New Frontiers in Artificial Intelligence: Joint JSAI 2001 Workshop Post-Proceedings (Lecture Notes in Computer Science, 2253) 3540430709, 9783540430704

Citation preview

Lecture Notes in Artificial Intelligence Subseries of Lecture Notes in Computer Science Edited by J. G. Carbonell and J. Siekmann

Lecture Notes in Computer Science Edited by G. Goos, J. Hartmanis, and J. van Leeuwen

2253

3

Berlin Heidelberg New York Barcelona Hong Kong London Milan Paris Tokyo

Takao Terano Toyoaki Nishida Akira Namatame Syusaku Tsumoto Yukio Ohsawa Takashi Washio (Eds.)

New Frontiers in Artificial Intelligence Joint JSAI 2001 Workshop Post-Proceedings

13

Series Editors Jaime G. Carbonell, Carnegie Mellon University, Pittsburgh, PA, USA J¨org Siekmann, University of Saarland, Saarbr¨ucken, Germany Volume Editors Takao Terano Yukio Ohsawa The University of Tsukuba, Graduate School of Business Science E-mail: {terano/osawa}@gssm.otsuka.tsukuba.ac.jp Toyoaki Nishida The University of Tokyo, School of Engineering E-mail: [email protected] Akira Namatame National Defense Academy, Dept. of Computer Science E-mail: [email protected] Syusaku Tsumoto Shimane Medical University, School of Medicines E-mail: [email protected] Takashi Washio Osaka University, The Institute of Scientific and Industrial Research E-mail: [email protected] Cataloging-in-Publication Data applied for Die Deutsche Bibliothek - CIP-Einheitsaufnahme New frontiers in artificial intelligence : joint JSAI 2001 workshop post proceedings / Takao Terano ... (ed.). - Berlin ; Heidelberg ; New York ; Barcelona ; Hong Kong ; London ; Milan ; Paris ; Tokyo : Springer, 2001 (Lecture notes in computer science ; Vol. 2253 : Lecture notes in artificial intelligence) ISBN 3-540-43070-9 CR Subject Classification (1998): I.2, H.2.8, H.3, F.1, H.4, H.5.2, I.5, J.1, J.3, K.4.3-4 ISSN 0302-9743 ISBN 3-540-43070-9 Springer-Verlag Berlin Heidelberg New York This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer-Verlag. Violations are liable for prosecution under the German Copyright Law. Springer-Verlag Berlin Heidelberg New York a member of BertelsmannSpringer Science+Business Media GmbH http://www.springer.de © Springer-Verlag Berlin Heidelberg 2001 Printed in Germany Typesetting: Camera-ready by author, data conversion by PTP-Berlin, Stefan Sossna Printed on acid-free paper SPIN: 10845999 06/3142 543210

Preface

The Japanese Society for Artificial Intelligence (JSAI) was established in July 1986. Since then, we have held conferences every year. Although JSAI is the second largest community in the world focusing on the area of Artificial Intelligence and we have over 3,000 members, the importance of the research presented and discussions held at the annual conferences has not been fully recognized in the Artificial Intelligence communities elsewhere in the world, partly because most presentations are made in the Japanese language. Therefore, the program committee of the Fifteenth Annual Conference of JSAI decided to open the door to the world and hold international workshops during the conference on May 20th and 25th, 2001 in Matsue City, Japan. The workshop proposals were gathered from the members of JSAI. We accepted the following up-to-date and exciting topics: 1) Social Intelligence Design chaired by Prof. Toyoaki Nishida, University of Tokyo, 2) Agent-Based Approaches in Economic and Social Complex Systems chaired by Prof. Akira Namatame, National Academy of Defense, 3) Rough Set Theory and Granular Computing chaired by Prof. Shusaku Tsumoto, Shimane Medical University, 4) Chance Discovery chaired by Prof. Yukio Osawa, and 5) Challenge in Knowledge Discovery and Data Mining chaired by Prof. Takashi Washio, Osaka University. These workshops were highly welcome and successful. A total of 116 people in Japan and 30 researchers from abroad participated in them. This volume of the proceedings contains selected papers presented at the workshops. The contents of the volume are divided into five parts, each of which corresponds to the topics of the workshops. Each paper was strictly reviewed by the committee members of the workshops. They also cover recent divergent areas of artificial intelligence. We believe that the volume is highly useful for both researchers and practitioners who have interests in recent advances in artificial intelligence.

October 2001

Takao Terano

JSAI Workshops as International Trends

Looking at the current economic, political, and ecological situations, we become aware of the dynamic environment surrounding all human activities. Hand in hand, the expansion of the World Wide Web is activating the whole globe as an information system including humans, computers, and networks. The workshop topics associated with JSAI 2001 were designed to hit such world wide trends. Social Information Designs are needed to aid the mutual progress of human society and various kinds of information flows. The Agent-Based Simulations consider social behavior from the aspect of economics, with the up-to-date viewpoint of complexity. Rough Set Theories may achieve a breakthrough with regard to dealing with uncertain real world events on the basis of established theories. Chance Discovery is a new direction proposed by Japanese researches, for helping people and agents be aware of novel information, significant for their own decisions in dynamic environments. KDD-Challengers are responding to requirements for new knowledge to be obtained from new data in new social situations. I am sure the selected papers from these first international workshops associated with JSAI will win the attention of people from several different areas of research, not only artificial intelligence but also social sciences and other areas looking into the future of human life. A piece of good news for those readers is that JSAI is becoming increasingly international, after many years as a semi-domestic Japanese AI community. With the foundation of five workshop themes this year, the new generation of AI researchers is finding new problems and new solutions in the creative atmosphere. On behalf of all the workshop organizers, I wish to draw readers’ attention to forthcoming international JSAI events. Before beginning the contents, let us express our gratitude to the great support given by the co-editors who organized each workshop, all authors and audiences, JSAI committee members, Shimane prefecture and Matsue city, and Jun’ichiro Mori of the University of Tokyo whose operations greatly aided this publication. October 2001

Yukio Ohsawa

Table of Contents

Part I. Social Intelligence Design 1. Social Intelligence Design – An Overview Toyoaki Nishida . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.2 Horizon of Social Intelligence Design . . . . . . . . . . . . . . . . . . . . . . 4 1.2.1 Methods of Establishing the Social Context . . . . . . . . . . 6 1.2.2 Embodied Conversational Agents and Social Intelligence 6 1.2.3 Collaboration Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.2.4 Public Discourse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 1.2.5 Theoretical Aspects of Social Intelligence Design . . . . . 8 1.2.6 Evaluations of Social Intelligence . . . . . . . . . . . . . . . . . . . 9 1.3 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2. FaintPop: In Touch with the Social Relationships Takeshi Ohguro, Kazuhiro Kuwabara, Tatsuo Owada, and Yoshinari Shirai . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.1 2.2 2.3 2.4

Social Intelligence Design for Communications . . . . . . . . . . . . . In Touch with the Social Relationships . . . . . . . . . . . . . . . . . . . . Initial Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion and Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . .

11 13 16 17

3. From Virtual Environment to Virtual Community A. Nijholt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Towards Multi-user Virtual Worlds . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 Interacting Embodied Personalities . . . . . . . . . . . . . . . . . 3.2.2 Embodied Personalities in Virtual Worlds . . . . . . . . . . . 3.3 Building a Theater Environment . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Interacting about Performances and Environment . . . . . . . . . . 3.5 Towards a Theater Community . . . . . . . . . . . . . . . . . . . . . . . . . . .

19 19 20 21 23 24 25

X

Table of Contents

4. Collaborative Innovation Tools John C. Thomas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 4.1 Importance of Collaboration: Practical and Scientific . . . . . . . . 27 4.2 New Technological Possibilities . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 4.3 Work of the Knowledge Socialization Group . . . . . . . . . . . . . . . 31 5. Bricks & Bits & Interaction R. Fruchter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 5.1 5.2 5.3 5.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Visibility, Awareness, and Interaction in Videoconference Space Mobile Learners in E-learning Spaces . . . . . . . . . . . . . . . . . . . . . Emerging Changes Influenced by Bricks & Bits & Interaction

35 36 39 41

6. A Distributed Multi-agent System for the Self-Evaluation of Dialogs Alain Cardon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 6.1 6.2 6.3 6.4 6.5 6.6

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . System General Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Representation of the Semantic of the Communication Act . . . Semantic Traits and Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Aspectual Agent Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . The Emerging Meaning of the Communication: The Morphological Agent Organization . . . . . . . . . . . . . . . . . . . . 6.7 Interpretation of the Morphological Organization: The Evocation Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

43 44 45 46 46 48 49 50

7. Public Opinion Channel: A System for Augmenting Social Intelligence of a Community Tomohiro Fukuhara, Toyoaki Nishida, and Shunsuke Uemura . . . . . 51 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Communication Costs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3 POC Prototype System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.1 POC Server . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.2 POC Client: POCViewer . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.5.1 Automatic Broadcasting System . . . . . . . . . . . . . . . . . . . . 7.5.2 POC and Narrative Intelligence . . . . . . . . . . . . . . . . . . . . 7.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

51 52 53 53 54 57 57 57 58 58

Table of Contents

XI

8. Enabling Public Discourse Keiichi Nakata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2 Enabling Individuals to Collect and Exchange Information and Opinions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3 Raising Social Awareness through Position-Oriented Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.1 Positioning-Oriented Discussion Interface . . . . . . . . . . . . 8.4 Towards “Social Intelligence Design” . . . . . . . . . . . . . . . . . . . . . . 8.5 Concluding Remark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

59 60 62 63 64 65

9. Internet, Discourses, and Democracy R. Luehrs, T. Malsch, and K. Voss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 9.1 9.2 9.3 9.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Online Support for Democratic Processes . . . . . . . . . . . . . . . . . . A Novel Participation Methodology . . . . . . . . . . . . . . . . . . . . . . . System Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

67 67 69 72

10. How to Evaluate Social Intelligence Design Nobuhiko Fujihara . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 10.1 Computer Networked Community as Social Intelligence . . . . . 10.2 The Importance of Control Condition in Evaluating Social Intelligence Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3 How to Evaluate POC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.4 Future Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

75 76 77 81

Part II. Agent-Based Approaches in Economic and Social Complex Systems 11. Overview Akira Namatame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 12. Analyzing Norm Emergence in Communal Sharing via Agent-Based Simulation Setsuya Kurahashi and Takao Terano . . . . . . . . . . . . . . . . . . . . . . . . . 88 12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2 Related Work on Studies of Norms . . . . . . . . . . . . . . . . . . . . . . . 12.3 Artificial Society Model TRURL . . . . . . . . . . . . . . . . . . . . . . . . . . 12.3.1 Agent Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.3.2 Communication and Action Energy . . . . . . . . . . . . . . . . . 12.3.3 Inverse Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

88 89 90 90 91 91 92

XII

Table of Contents

12.4.1 An Amount of Information in Each Society . . . . . . . . . . 12.4.2 Emergence and Collapse of a Norm . . . . . . . . . . . . . . . . . 12.4.3 Emergence and Control of Free Riders . . . . . . . . . . . . . . 12.4.4 Information Gap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

92 93 94 95 96 97

13. Toward Cumulative Progress in Agent-Based Simulation Keiki Takadama and Katsunori Shimohara . . . . . . . . . . . . . . . . . . . . 99 13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13.2 Can We Assist Cumulative Progress? . . . . . . . . . . . . . . . . . . . . . 13.2.1 Problems in Agent-Based Approaches . . . . . . . . . . . . . . . 13.2.2 Points for Cumulative Progress . . . . . . . . . . . . . . . . . . . . . 13.2.3 Cumulative Progress in Current Projects . . . . . . . . . . . . 13.3 Exploring Key Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13.3.1 Interpretation by Implementation . . . . . . . . . . . . . . . . . . 13.3.2 Applications of IbI Approach . . . . . . . . . . . . . . . . . . . . . . 13.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13.4.1 Cumulative Progress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13.4.2 Potential of Our Approach . . . . . . . . . . . . . . . . . . . . . . . . 13.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

99 100 100 100 101 101 102 103 104 104 105 107

14. Complexity of Agents and Complexity of Markets Kiyoshi Izumi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.2 The Efficient Market Hypothesis Seen from Complexity . . . . . 14.3 Artificial Market Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.3.1 Expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.3.2 Order . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.3.3 Price Determination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.3.4 Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.4 Simulation Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.4.1 Merit of Complicating a Prediction Formula . . . . . . . . . 14.4.2 The Demerit in the Whole Market . . . . . . . . . . . . . . . . . . 14.4.3 Development of the Complexity of a Market . . . . . . . . . 14.5 New Efficient Market Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . 14.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

110 111 112 112 113 113 113 114 114 115 115 118 119

Table of Contents

XIII

15. U-Mart Project: Learning Economic Principles from the Bottom by Both Human and Software Agents Hiroshi Sato, Hiroyuki Matsui, Isao Ono, Hajime Kita, Takao Terano, Hiroshi Deguchi, and Yoshinori Shiozawa . . . . . . . . . . . . . . . . . . . . . . 121 15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.2 Outlines of U-Mart System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.3 Outline of Open Experiment, Pre U-Mart 2000 . . . . . . . . . . . . . 15.3.1 Open Experiment and Its Objectives . . . . . . . . . . . . . . . . 15.3.2 Experimental System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.3.3 Configuration of Experiment . . . . . . . . . . . . . . . . . . . . . . . 15.4 Participated Agents and Their Strategies . . . . . . . . . . . . . . . . . . 15.5 Experimental Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.5.1 First Round . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.5.2 Second Round . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.5.3 Variety of Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.5.4 Reason of Heavy Rises and Falls . . . . . . . . . . . . . . . . . . . 15.6 Experiments with Human Agents . . . . . . . . . . . . . . . . . . . . . . . . . 15.7 Conclusion and Acknowledgements . . . . . . . . . . . . . . . . . . . . . . .

121 122 123 123 123 123 123 126 126 127 127 128 129 130

16. A Multi-objective Genetic Algorithm Approach to Construction of Trading Agents for Artificial Market Study Rikiya Fukumoto and Hajime Kita . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 16.1 16.2 16.3 16.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The U-Mart System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Multi-objective Genetic Algorithms (MOGA) . . . . . . . . . . . . . . Construction of Trading Agents with a MOGA . . . . . . . . . . . . . 16.4.1 Structure of Trading Agents . . . . . . . . . . . . . . . . . . . . . . . 16.4.2 Implementation of MOGA . . . . . . . . . . . . . . . . . . . . . . . . . 16.5 Results of Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

132 133 133 134 134 137 139 140

17. Agent-Based Simulation for Economic and Environmental Studies Hideyuki Mizuta and Yoshiki Yamagata . . . . . . . . . . . . . . . . . . . . . . . 142 17.1 17.2 17.3 17.4 17.5 17.6

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Agent-Based Simulation Framework: ASIA . . . . . . . . . . . . . . . . Market Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dynamic Online Auctions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Greenhouse Gas Emissions Trading . . . . . . . . . . . . . . . . . . . . . . . Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

142 143 145 146 147 151

XIV

Table of Contents

18. Avatamsaka Game Experiment as a Nonlinear Polya Urn Process Yuji Aruka . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 18.1 Characteristics of Avatamsaka Game . . . . . . . . . . . . . . . . . . . . . . 18.1.1 Synchronization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.1.2 A Two Person Game Form . . . . . . . . . . . . . . . . . . . . . . . . 18.1.3 No Complementarities Except for Positive Spillovers to Be Found . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.2 Avatamsaka Game Experiment as a Nonlinear Polya Urn Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18.2.1 The Elementary Polya Process . . . . . . . . . . . . . . . . . . . . . 18.2.2 A Generalized Polya Urn Process . . . . . . . . . . . . . . . . . . . 18.2.3 A Nonlinear Polya Process . . . . . . . . . . . . . . . . . . . . . . . .

154 154 155 156 157 157 158 160

19. Effects of Punishment into Actions in Social Agents Keji Suzuki . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 19.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19.2 The Tragedy of the Common . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19.3 Coevolving Levy Plan and Payoff Prediction . . . . . . . . . . . . . . . 19.3.1 Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19.3.2 Relation between Levy Plan and Payoff Prediction . . . . 19.3.3 Reward of Agent and Incoming Levy of Meta-agent . . . 19.3.4 Evaluation of Game . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19.3.5 Coevolution of Plan and Predictions . . . . . . . . . . . . . . . . 19.4 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19.4.1 Game without Meta-agent . . . . . . . . . . . . . . . . . . . . . . . . . 19.4.2 Simulations with Meta-agents . . . . . . . . . . . . . . . . . . . . . . 19.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

162 163 164 164 165 166 167 167 169 169 169 172

20. Analysis of Norms Game with Mutual Choice Tomohisa Yamashita, Hidenori Kawamura, Masahito Yamamoto, and Azuma Ohuchi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 20.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20.2 Mutual Choice in Group Formation . . . . . . . . . . . . . . . . . . . . . . . 20.2.1 Norms Game with Mutual Choice . . . . . . . . . . . . . . . . . . 20.2.2 Metanorms Game with Mutual Choice . . . . . . . . . . . . . . 20.3 Simulation Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20.4 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20.4.1 Maintenance of Norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20.4.2 Establishment of Norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

174 175 175 177 177 178 178 180 183

Table of Contents

XV

21. Cooperative Co-evolution of Multi-agents Sung-Bae Cho . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 21.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.2 Evolutionary Approach to IPD Game . . . . . . . . . . . . . . . . . . . . . 21.3 Cooperative Co-evolution of Strategies . . . . . . . . . . . . . . . . . . . . 21.3.1 Forming Coalition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.3.2 Evolving Strategy Coalition . . . . . . . . . . . . . . . . . . . . . . . . 21.3.3 Gating Strategies in Coalition . . . . . . . . . . . . . . . . . . . . . . 21.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.4.1 Evolution of Strategy Coalition . . . . . . . . . . . . . . . . . . . . 21.4.2 Gating Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

185 186 187 187 188 188 190 190 191 192

22. Social Interaction as Knowledge Trading Games Kazuyo Sato and Akira Namatame . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 22.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22.2 Knowledge Transaction as Knowledge Trading Games . . . . . . . 22.3 Knowledge Trading as Symmetric and Asymmetric Coordination Games . . . . . . . . . . . . . . . . . . . . . . . . . 22.4 Aggregation of Heterogeneous Payoff Matrices . . . . . . . . . . . . . . 22.5 The Collective Behavior in Knowledge Transaction . . . . . . . . . 22.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

195 197 198 201 203 206

23. World Trade League as a Standard Problem for Multi-agent Economics – Concept and Background Koichi Kurumatani and Azuma Ohuchi . . . . . . . . . . . . . . . . . . . . . . . . 208 23.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23.2 Concept of World Trade League . . . . . . . . . . . . . . . . . . . . . . . . . . 23.3 Elements of World Trade League . . . . . . . . . . . . . . . . . . . . . . . . . 23.3.1 Behavior Options of Agents and Market Structure . . . . 23.3.2 Game Settings and Complexity . . . . . . . . . . . . . . . . . . . . 23.3.3 Evaluation Function of Players . . . . . . . . . . . . . . . . . . . . . 23.4 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23.4.1 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23.4.2 Communication Protocol X-SS . . . . . . . . . . . . . . . . . . . . . 23.5 Requirements for Standard Problem in Multi-agent Economics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23.6 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

208 209 210 210 211 212 212 212 213 214 215 216

XVI

Table of Contents

24. Virtual Economy Simulation and Gaming —An Agent Based Approach— Hiroshi Deguchi, Takao Terano, Koichi Kurumatani, Taro Yuzawa, Shigeji Hashimoto, Hiroyuki Matsui, Akio Sashima, and Toshiyuki Kaneda . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218 24.1 24.2 24.3 24.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Agent Based Simulation Model for Virtual Economy . . . . . . . . Result of Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

218 219 223 225

25. Boxed Economy Foundation Model: Model Framework for Agent-Based Economic Simulations Takashi Iba, Yohei Takabe, Yoshihide Chubachi, Junichiro Tanaka, Kenichi Kamihashi, Ryunosuke Tsuya, Satomi Kitano, Masaharu Hirokane, and Yoshiaki Matsuzawa . . . . . 227 25.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25.2 Model Framework for Agent-Based Economic Simulations . . . 25.3 Boxed Economy Foundation Model . . . . . . . . . . . . . . . . . . . . . . . 25.3.1 EconomicActor, SocialGroup, Individual . . . . . . . . . . . . 25.3.2 Goods, Information, Possession . . . . . . . . . . . . . . . . . . . . . 25.3.3 Behavior, BehaviorManagement, Memory, Needs . . . . . 25.3.4 Relation, Path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25.4 Applying Boxed Economy Foundation Model . . . . . . . . . . . . . . . 25.4.1 Modeling Behavior Rather than Agent . . . . . . . . . . . . . . 25.4.2 Flexibility on the Boundary of Agent . . . . . . . . . . . . . . . 25.4.3 Example: Sellers in Distribution Mechanism . . . . . . . . . 25.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

227 228 228 229 231 232 232 233 233 233 234 235

Part III. Rough Set Theory and Granular Computing 26. Workshop on Rough Set Theory and Granular Computing – Summary Shusaku Tsumoto, Shoji Hirano, and Masahiro Inuiguchi . . . . . . . . . 239 27. Bayes’ Theorem Revised – The Rough Set View Zdzislaw Pawlak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240 27.1 27.2 27.3 27.4 27.5 27.6 27.7

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bayes’ Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Information Systems and Approximation of Sets . . . . . . . . . . . . Rough Membership . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Information Systems and Decision Rules . . . . . . . . . . . . . . . . . . . Probabilistic Properties of Decision Tables . . . . . . . . . . . . . . . . . Decision Tables and Flow Graphs . . . . . . . . . . . . . . . . . . . . . . . . .

240 241 242 244 244 245 246

Table of Contents

XVII

27.8 Comparison of Bayesian and Rough Set Approach . . . . . . . . . . 247 27.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249 28. Toward Intelligent Systems: Calculi of Information Granules Andrzej Skowron . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 28.1 28.2 28.3 28.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . AR-Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rough Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Decomposition of Information Granules . . . . . . . . . . . . . . . . . . .

251 254 255 256

29. Soft Computing Pattern Recognition: Principles, Integrations, and Data Mining Sankar K. Pal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261 29.1 29.2 29.3 29.4 29.5 29.6 29.7

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Relevance of Fuzzy Set Theory in Pattern Recognition . . . . . . Relevance of Neural Network Approaches . . . . . . . . . . . . . . . . . . Genetic Algorithms for Pattern Recognition . . . . . . . . . . . . . . . . Integration and Hybrid Systems . . . . . . . . . . . . . . . . . . . . . . . . . . Evolutionary Rough Fuzzy MLP . . . . . . . . . . . . . . . . . . . . . . . . . Data Mining and Knowledge Discovery . . . . . . . . . . . . . . . . . . . .

261 262 264 265 266 267 268

30. Identifying Upper and Lower Possibility Distributions with Rough Set Concept P. Guo and Hideo Tanaka . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272 30.1 Concepts of Upper and Lower Possibility Distributions . . . . . . 30.2 Comparison of Dual Possibility Distributions with Dual Approximations in Rough Sets Theory . . . . . . . . . . . . . . . 30.3 Identification of Upper and Lower Possibility Distributions . . 30.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

272 273 274 277

31. On Fractals in Information Systems: The First Step Lech Polkowski . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278 31.1 31.2 31.3 31.4 31.5

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Fractal Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rough Sets and Topologies on Rough Sets . . . . . . . . . . . . . . . . . Fractals in Information Systems . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

278 278 279 280 282

32. Generalizations of Fuzzy Multisets for Including Infiniteness Sadaaki Miyamoto . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283 32.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283

XVIII Table of Contents

32.2 32.3 32.4 32.5

Multisets and Fuzzy Multisets . . . . . . . . . . . . . . . . . . . . . . . . . . . . Infinite Memberships . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A Set-Valued Multiset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

284 285 286 287

33. Fuzzy c-Means and Mixture Distribution Model for Clustering Based on L1 -Space Takatsugu Koga, Sadaaki Miyamoto, and Osamu Takata . . . . . . . . . 289 33.1 33.2 33.3 33.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Fuzzy c-Means Based on L1 -Space . . . . . . . . . . . . . . . . . . . . . . . . Mixture Distribution Based on L1 -Space . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

289 289 291 293

34. On Rough Sets under Generalized Equivalence Relations Masahiro Inuiguchi and Tetsuzo Tanino . . . . . . . . . . . . . . . . . . . . . . . . 295 34.1 34.2 34.3 34.4 34.5

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The Original Rough Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Two Different Problem Settings . . . . . . . . . . . . . . . . . . . . . . . . . . Approximation by Means of Elementary Sets . . . . . . . . . . . . . . . Distinction among Three Regions . . . . . . . . . . . . . . . . . . . . . . . . .

295 296 297 298 298

35. Two Procedures for Dependencies among Attributes in a Table with Non-deterministic Information: A Summary Hiroshi Sakai . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301 35.1 35.2 35.3 35.4 35.5 35.6 35.7

Preliminary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Definitions of NISs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A Way to Obtain All Possible Equivalence Relations . . . . . . . Procedure 1 for Dependencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . Procedure 2 for Dependencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . Execution Time of Every Method . . . . . . . . . . . . . . . . . . . . . . . . . Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

301 302 303 303 304 304 305

36. An Application of Extended Simulated Annealing Algorithm to Generate the Learning Data Set for Speech Recognition System Chi-Hwa Song and Won Don Lee . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306 36.1 36.2 36.3 36.4 36.5 36.6

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Domain Definition for LDS Extraction . . . . . . . . . . . . . . . . . . . . The Numerical Formula for LDS Extraction . . . . . . . . . . . . . . . The Algorithm for Extraction of LDS . . . . . . . . . . . . . . . . . . . . . Experimental and Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

306 306 307 308 309 310

Table of Contents

XIX

37. Generalization of Rough Sets with α-Coverings of the Universe Induced by Conditional Probability Relations Rolly Intan, Masao Mukaidono, and Y.Y. Yao . . . . . . . . . . . . . . . . . . 311 37.1 37.2 37.3 37.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conditional Probability Relations . . . . . . . . . . . . . . . . . . . . . . . . Generalized Rough Sets Approximation . . . . . . . . . . . . . . . . . . . Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

311 312 313 315

38. On Mining Ordering Rules Y.Y. Yao and Ying Sai . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316 38.1 38.2 38.3 38.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ordered Information Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mining Ordering Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

316 317 318 320

39. Non-additive Measures by Interval Probability Functions Hideo Tanaka, Kazutomi Sugihara, and Yutaka Maeda . . . . . . . . . . . 322 39.1 39.2 39.3 39.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Interval Probability Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . Combination and Conditional Rules for IPF . . . . . . . . . . . . . . . Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

322 323 325 326

40. Susceptibility to Consensus of Conflict Profiles Ngoc Thanh Nguyen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327 40.1 40.2 40.3 40.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conflict Profiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Susceptibility to Consensus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

327 327 329 331

41. Analysis of Image Sequences for the Unmanned Aerial Vehicle Hung Son Nguyen, Andrzej Skowron, and Marcin S. Szczuka . . . . . 333 41.1 41.2 41.3 41.4 41.5 41.6

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The Task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

333 334 334 334 335 337

XX

Table of Contents

42. The Variable Precision Rough Set Inductive Logic Programming Model and Web Usage Graphs V. Uma Maheswari, Arul Siromoney, and K.M. Mehata . . . . . . . . . . 339 42.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42.2 The VPRSILP Model and Web Usage Graphs . . . . . . . . . . . . . . 42.2.1 A Simple–Graph–VPRSILP–ESD System . . . . . . . . . . . . 42.2.2 Web Usage Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42.3 Experimental Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

339 339 340 340 341 343

43. Optimistic Priority Weights with an Interval Comparison Matrix Tomoe Entani, Hidetomo Ichihashi, and Hideo Tanaka . . . . . . . . . . . 344 43.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43.2 Interval AHP with Interval Comparison Matrix . . . . . . . . . . . . 43.3 Choice of Optimistic Weights and Efficiency by DEA . . . . . . . . 43.3.1 DEA with Normalized Data . . . . . . . . . . . . . . . . . . . . . . . 43.3.2 Optimistic Importance Grades in Interval Importance Grades . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43.4 Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

344 345 346 346 346 347 348

44. Rough Set Theory in Conflict Analysis ´ ˛zak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349 Rafal Deja and Dominik Sle 44.1 44.2 44.3 44.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conflict Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

349 350 352 352

45. Dealing with Imperfect Data by RS-ILP Chunnian Liu and Ning Zhong . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354 45.1 45.2 45.3 45.4 45.5

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Imperfect Data in ILP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . RS-ILP for Missing Classification . . . . . . . . . . . . . . . . . . . . . . . . . RS-ILP for Too Strong Bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

354 355 356 357 357

46. Extracting Patterns Using Information Granules: A Brief Introduction Andrzej Skowron, Jaroslaw Stepaniuk, and James F. Peters . . . . . . 359 46.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359 46.2 Granule Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359

Table of Contents

XXI

47. Classification Models Based on Approximate Bayesian Networks ´ ˛zak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364 Dominik Sle 47.1 47.2 47.3 47.4 47.5 47.6

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Frequencies in Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Approximate Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bayesian Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Approximate Bayesian Networks . . . . . . . . . . . . . . . . . . . . . . . . . Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

364 364 365 366 367 368

48. Identifying Adaptable Components – A Rough Sets Style Approach Yoshiyuki Shinkawa and Masao J. Matsumoto . . . . . . . . . . . . . . . . . . 370 48.1 48.2 48.3 48.4 48.5

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Defining Adaptation of Software Components . . . . . . . . . . . . . . Identifying One-to-One Component Adaptation . . . . . . . . . . . . Identifying One-to-Many Component Adaptation . . . . . . . . . . . Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

370 370 371 373 374

49. Rough Measures and Integrals: A Brief Introduction Zdzislaw Pawlak, James F. Peters, Andrzej Skowron, Z. Suraj, S. Ramanna, and M. Borkowski . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375 49.1 49.2 49.3 49.4 49.5 49.6

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Classical Additive Set Functions . . . . . . . . . . . . . . . . . . . . . . . . . . Basic Concepts of Rough Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rough Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Relevance of a Sensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

375 376 376 377 378 378

50. Association Rules in Semantically Rich Relations: Granular Computing Approach T.Y. Lin and Eric Louie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380 50.1 50.2 50.3 50.4 50.5 50.6

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Relational Models and Rough Granular Structures . . . . . . . . . . Databases with Additional Semantics . . . . . . . . . . . . . . . . . . . . . Mining Real World or Its Representations . . . . . . . . . . . . . . . . . Clustered Association Rules-Mining Semantically . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

380 380 381 382 383 383

51. A Note on Filtration and Granular Reasoning Tetsuya Murai, Michinori Nakata, and Yoshiharu Sato . . . . . . . . . . . 385 51.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385 51.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385

XXII

Table of Contents

51.3 Relative Filtration with Approximation . . . . . . . . . . . . . . . . . . . 386 51.4 Example of Granular Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . 388 51.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389 52. A Note on Conditional Logic and Association Rules Tetsuya Murai, Michinori Nakata, and Yoshiharu Sato . . . . . . . . . . . 390 52.1 52.2 52.3 52.4 52.5

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Association Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Previous Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Graded Conditional Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

390 391 391 392 394

53. Analysis of Self-Injurious Behavior by the LERS Data Mining System Rachel L. Freeman, Jerzy W. Grzymala-Busse, Laura A. Riffel, and Stephen R. Schroeder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395 53.1 53.2 53.3 53.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Data Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

395 396 397 398

54. A Clustering Method for Nominal and Numerical Data Based on Rough Set Theory Shoji Hirano, Shusaku Tsumoto, Tomohiro Okuzaki, and Yutaka Hata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400 54.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54.2 Clustering Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54.2.1 Initial Equivalence Relation . . . . . . . . . . . . . . . . . . . . . . . . 54.2.2 Modification of Equivalence Relations . . . . . . . . . . . . . . . 54.2.3 Evaluation of Validity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54.3 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

400 401 401 402 403 404 404

55. A Design of Architecture for Rough Set Processor Akinori Kanasugi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406 55.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55.2 Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55.2.1 Data Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55.2.2 Execution Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55.2.3 Discernibility Matrix Maker . . . . . . . . . . . . . . . . . . . . . . . 55.2.4 Core Selector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55.2.5 Covering Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55.2.6 Reconstruction Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

406 406 406 407 407 408 408 408

Table of Contents XXIII

55.2.7 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409 55.2.8 Performance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409 55.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 410 Part IV. Chance Discovery 56. The Scope of Chance Discovery Yukio Ohsawa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413 57. Chance Discovery Using Dialectical Argumentation Peter McBurney and Simon Parsons . . . . . . . . . . . . . . . . . . . . . . . . . . . 414 57.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57.2 Argumentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57.3 The Discovery Agora: Formal Structure . . . . . . . . . . . . . . . . . . . 57.3.1 Discovery Dialogues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57.3.2 Model of a Discovery Dialogue . . . . . . . . . . . . . . . . . . . . . 57.3.3 Dialogue Game Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

414 415 417 417 418 420 423

58. Methodological Considerations on Chance Discovery Helmut Prendinger and Mitsuru Ishizuka . . . . . . . . . . . . . . . . . . . . . . 425 58.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58.2 Nature vs. Open Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58.2.1 Prediction in the Natural Sciences . . . . . . . . . . . . . . . . . . 58.2.2 Prediction in Open Systems . . . . . . . . . . . . . . . . . . . . . . . 58.3 Chance Discovery in Open Systems . . . . . . . . . . . . . . . . . . . . . . . 58.3.1 Enterprise Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58.3.2 The Limits of Regulatory Mechanisms . . . . . . . . . . . . . . 58.3.3 Chance Discovery as Anticipation . . . . . . . . . . . . . . . . . . 58.4 Chance Discovery, Uncertainty, Freedom . . . . . . . . . . . . . . . . . . . 58.4.1 Freedom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58.4.2 Explaining versus Predicting . . . . . . . . . . . . . . . . . . . . . . . 58.5 Scientific Evaluation of Theories . . . . . . . . . . . . . . . . . . . . . . . . . . 58.6 Chance Discovery vs. KDD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58.7 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

425 426 426 427 427 427 427 428 429 429 430 430 431 432

59. Future Directions of Communities on the Web Naohiro Matsumura, Yukio Ohsawa, and Mitsuru Ishizuka . . . . . . . 435 59.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59.2 Related Researches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59.2.1 Discovery of Communities . . . . . . . . . . . . . . . . . . . . . . . . . 59.2.2 Discovery of Future Directions . . . . . . . . . . . . . . . . . . . . . 59.3 Future Directions of Communities . . . . . . . . . . . . . . . . . . . . . . . .

435 436 436 437 438

XXIV Table of Contents

59.3.1 How to Discover the Future Directions? . . . . . . . . . . . . . 59.3.2 The Detailed Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59.4 Experiments and Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59.4.1 Future Directions of Portal Sites . . . . . . . . . . . . . . . . . . . 59.4.2 Future Directions of Book Site . . . . . . . . . . . . . . . . . . . . . 59.4.3 Future Directions of Artificial Intelligence . . . . . . . . . . . 59.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

438 439 440 440 441 442 442

60. A Document as a Small World Yutaka Matsuo, Yukio Ohsawa, and Mitsuru Ishizuka . . . . . . . . . . . . 444 60.1 60.2 60.3 60.4 60.5 60.6

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Small World . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Term Co-occurrence Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Finding Important Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

444 444 445 446 447 448

61. Support System for Creative Activity by Information Acquirement through Internet Wataru Sunayama and Masahiko Yachida . . . . . . . . . . . . . . . . . . . . . 449 61.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61.2 Framework for Creative Activity . . . . . . . . . . . . . . . . . . . . . . . . . 61.2.1 User Discovers a Viewpoint of the Combination . . . . . . 61.2.2 Support System for Search Systems . . . . . . . . . . . . . . . . . 61.2.3 Data Mining from Web Pages . . . . . . . . . . . . . . . . . . . . . . 61.2.4 Interface for Knowledge Refinement . . . . . . . . . . . . . . . . . 61.3 Experimental System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

449 449 450 450 451 451 452 453

62. An Approach to Support Long-Term Creative Thinking and Its Feasibility Hirohito Shibata and Koichi Hori . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455 62.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62.2 System Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62.3 Long-Term User Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62.3.1 Behavior Analysis on Pop-Up . . . . . . . . . . . . . . . . . . . . . . 62.3.2 Effects and Open Problems . . . . . . . . . . . . . . . . . . . . . . . . 62.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

455 456 458 458 460 460

63. Chance Discovery by Creative Communicators Observed in Real Shopping Behavior Hiroko Shoji and Koichi Hori . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462 63.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462

Table of Contents

XXV

63.2 Collecting Protocols of Actual Purchase Activities . . . . . . . . . . 63.3 Analysis and Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63.3.1 Expected Reaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63.3.2 Unexpected Reaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63.3.3 Successful Chance Discovery with Unexpected Reaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

463 463 463 465 465 466

64. The Role of Counterexamples in Discovery Learning Environment: Awareness of the Chance for Learning Tomoya Horiguchi and Tsukasa Hirashima . . . . . . . . . . . . . . . . . . . . . 468 64.1 64.2 64.3 64.4 64.5

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chance Discovery in Learning Environment . . . . . . . . . . . . . . . . How to Design Effective Counterexamples . . . . . . . . . . . . . . . . . Designing ‘Visible’ Counterexamples . . . . . . . . . . . . . . . . . . . . . . Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

468 469 470 471 473

65. Integrating Data Mining Techniques and Design Information Management for Failure Prevention Yoshikiyo Kato, Takehisa Yairi, and Koichi Hori . . . . . . . . . . . . . . . . 475 65.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65.2 Fault Detection of Spacecraft by Mining Association Rules of Housekeeping Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65.3 Managing Information for Failure Prevention . . . . . . . . . . . . . . . 65.3.1 Using Design Information for Failure Prevention . . . . . . 65.3.2 Design Information Repository . . . . . . . . . . . . . . . . . . . . . 65.3.3 Handling Anomalies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65.4 Current Work and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . .

475 476 477 477 479 480 480

66. Action Proposal as Discovery of Context (An Application to Family Risk Management) Yukio Ohsawa and Yumiko Nara . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481 66.1 Introduction : Which Opinions Grow into Consensus ? . . . . . . 66.2 KeyGraph for Noticing Consensus Seeds from Questionnaire . 66.3 Family Perception of Risks and Opportunities . . . . . . . . . . . . . . 66.3.1 The Results of KeyGraph . . . . . . . . . . . . . . . . . . . . . . . . . 66.3.2 Which Opinions Grew into Consensus? . . . . . . . . . . . . . . 66.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

481 482 483 484 485 485

67. Retrieval of Similar Time-Series Patterns for Chance Discovery Takuichi Nishimura and Ryuichi Oka . . . . . . . . . . . . . . . . . . . . . . . . . . 486 67.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 486

XXVI Table of Contents

67.2 Reference Interval-Free Active Search . . . . . . . . . . . . . . . . . . . . 487 67.3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488 67.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489 68. Fuzzy Knowledge Based Systems and Chance Discovery Vicen¸c Torra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491 68.1 68.2 68.3 68.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Fuzzy Knowledge Based Systems . . . . . . . . . . . . . . . . . . . . . . . . . System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

491 492 493 494

Part V. Challenge in Knowledge Discovery and Datamining 69. JSAI KDD Challenge 2001: JKDD01 Program Chair: Takashi Washio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499 70. Knowledge Discovery Support from a Meningoencephalitis Dataset Using an Automatic Composition Tool for Inductive Applications Hiromitsu Hatazawa, Hidenao Abe, Mao Komori, Yoshiaki Tachibana, and Takahira Yamaguchi . . . . . . . . . . . . . . . . . . 500 70.1 70.2 70.3 70.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ontologies for Inductive Learning . . . . . . . . . . . . . . . . . . . . . . . . . Basic Design of CAMLET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A Case Study of Knowledge Discovery Support Using a Meningoencephalitis Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70.4.1 Learning Rules from the View of Precision . . . . . . . . . . . 70.4.2 Learning Rules from the View of Specificity . . . . . . . . . . 70.5 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . .

500 501 502 503 504 505 507

71. Extracting Meningitis Knowledge by Integration of Rule Induction and Association Mining T.B. Ho, S. Kawasaki, and D.D. Nguyen . . . . . . . . . . . . . . . . . . . . . . . 508 71.1 71.2 71.3 71.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . LUPC: Learning Unbalanced Positive Class . . . . . . . . . . . . . . . . Finding Rules from Meningitis Data . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

508 508 509 512

72. Basket Analysis on Meningitis Data Takayuki Ikeda, Takashi Washio, and Hiroshi Motoda . . . . . . . . . . . 516 72.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516 72.2 Method for Selection and Discretization . . . . . . . . . . . . . . . . . . . 517

Table of Contents XXVII

72.2.1 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72.2.2 Performance Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72.3 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72.4 Result and Expert’s Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . 72.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

517 518 520 521 523

73. Extended Genetic Programming Using Apriori Algorithm for Rule Discovery Ayahiko Niimi and Eiichiro Tazaki . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525 73.1 73.2 73.3 73.4

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Genetic Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Approach of Proposed Combined Learning . . . . . . . . . . . . . . . . Apply to Rule Discovery from Database . . . . . . . . . . . . . . . . . . 73.4.1 ADF-GP Only . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73.4.2 Proposed Technique (Association Rules + ADF-GP) . . 73.4.3 Discussion for the Results . . . . . . . . . . . . . . . . . . . . . . . . . 73.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

525 526 527 528 529 529 531 531

74. Medical Knowledge Discovery on the Meningoencephalitis Diagnosis Studied by the Cascade Model Takashi Okada . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533 74.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74.2 The Cascade Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74.3.1 Computation by DISCAS . . . . . . . . . . . . . . . . . . . . . . . . . 74.3.2 Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74.3.3 Detection of Bacteria or Virus . . . . . . . . . . . . . . . . . . . . . 74.3.4 Prognosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74.4 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

533 533 535 535 536 538 539 540

75. Meningitis Data Mining by Cooperatively Using GDT-RS and RSBR Ning Zhong, Ju-Zhen Dong, and Setsuo Ohsuga . . . . . . . . . . . . . . . . . 541 75.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75.2 Rule Discovery by GDT-RS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75.2.1 GDT and Rule Strength . . . . . . . . . . . . . . . . . . . . . . . . . . . 75.2.2 A Searching Algorithm for Optimal Set of Rules . . . . . . 75.3 Discretization Based on RSBR . . . . . . . . . . . . . . . . . . . . . . . . . . . 75.4 Application in Meningitis Data Mining . . . . . . . . . . . . . . . . . . . . 75.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

541 542 542 544 546 546 547

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 549 Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 551

1 . S o c ia l I n te llig e n c e D e s ig n – A n O v e r v ie w Toy oa k i N i s h i d a D ep a r tm en t of In for m a ti on a n d Com m u n i ca ti on E n g i n eer i n g G r a d u a te Sch ool of In for m a ti on Sci en ce a n d Tech n ol og y Th e Un i ver s i ty of Tok y o 7-3 -1 Hon g o, B u n k y o-k u , Tok y o 113 -8 656, Ja p a n n i s h i d a @ k c.t.u -tok y o.a c.j p

1 .1 I n t r o d u c t io n Th e a d ven t of th e In ter n et a n d i n for m a ti on tech n ol og y h a s b r ou g h t a b ou t s i g n i fi ca n t p r og r es s i n a u g m en ti n g th e wa y p eop l e ca n i n ter a ct wi th ea ch oth er i n a tota l l y n ew fa s h i on th a t wa s n ot p os s i b l e i n th e p a s t. E x a m p l es of n ew tech n ol og i es i n cl u d e con ver s a ti on a l a g en ts th a t m ed i a te p eop l e i n g etti n g to k n ow a n d com m u n i ca te wi th ea ch oth er , a col l a b or a ti ve vi r tu a l en vi r on m en t for l a r g e-s ca l e d i s cu s s i on s , p er s on a l i z ed i n for m a ti on tool s for h el p i n g cr os s -cu l tu r a l com m u n i ca ti on , i n ter a cti ve com m u n i ty m ed i a for a u g m en ti n g com m u n i ty a wa r en es s a n d m em or y , to n a m e j u s t a few. Som eti m es n ew tech n ol og i es i n d u ce th e em er g en ce of a n ew l a n g u a g e a n d l i fes ty l e. F or ex a m p l e, i n ter a cti ve m u l ti m ed i a web s i tes a r e a n ew m ed i u m a n d p r ob a b l y even a n ew l a n g u a g e, wi th i n ter es ti n g n ew con ven ti on s , a n d i n cr ea s i n g a d a p ta ti on to th e s u p p or t of com m u n i ti es . Ja p a n es e teen a g er s h a ve d evel op ed a n ew l a n g u a g e for u s e or i g i n a l l y wi th b eep er s a n d n ow wi th m ob i l e p h on es . Th es e a r e b oth n ew m a i n s tr ea m r ea l wor l d d evel op m en ts th a t s h ou l d b e s tu d i ed fu r th er , a n d cou l d p r ob a b l y g i ve s om e va l u a b l e i n s i g h ts . Th e th em e of Soci a l In tel l i g en ce D es i g n i s r ea l l y a n a n g l e on th e s u p p or t of g r ou p s i n p u r s u i t of th ei r g oa l s , wh eth er th a t i s m ed i ca l k n owl ed g e, s tock tr a d i n g , or teen a g e g os s i p . Soci a l In tel l i g en ce D es i g n g i ves s om e n ew l i fe to Ag en t Tech n ol og y a n d Ar ti fi ci a l In tel l i g en ce r es ea r ch i n g en er a l i n th a t h u m a n s a r e i n teg r a l p a r t of a b i g p i ctu r e b y s h i fti n g th e focu s , fr om b u i l d i n g a r ti fa cts wi th th e p r ob l em s ol vi n g or l ea r n i n g a b i l i ty , to d es i g n i n g a fr a m ewor k of i n ter a cti on th a t l ea d s to cr ea ti on of n ew k n owl ed g e a n d r el a ti on s h i p a m on g p a r ti ci p a n ts . P r om i s i n g a p p l i ca ti on d om a i n s of Soci a l In tel l i g en ce D es i g n i n cl u d e col l a b or a ti ve en vi r on m en t, e-l ea r n i n g , k n owl ed g e m a n a g em en t, com m u n i ty s u p p or t s y s tem s , s y m b i os i s of h u m a n s a n d a r ti fa cts , cr i s i s m a n a g em en t, a n d d i g i ta l d em ocr a cy . In wh a t fol l ows , I wi l l over vi ew m a j or i s s u es i n vol ved i n Soci a l In tel l i g en ce D es i g n a n d a ttem p t a t s tr u ctu r e th em i n a coh er en t s tor y .1

1

Th e fol l owi n g d es cr i p ti on i s i n d eb ted to th e d i s cu s s i on s a t JSAI-Sy n s op h y In ter n a ti on a l Wor k s h op on Soci a l In tel l i g en ce D es i g n , Ma ts u e, Ja p a n , Ma y 21-22, 2001.

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 -10, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

4

T. N i s h i d a

1 .2 H o r iz o n o f S o c ia l I n t e llig e n c e D e s ig n Soci a l In tel l i g en ce D es i g n i s a d i s ci p l i n e a i m ed a t u n d er s ta n d i n g a n d s u p p or ti n g s oci a l i n tel l i g en ce. Con ven ti on a l l y , s oci a l i n tel l i g en ce h a s b een d i s cu s s ed i n th e con tex t of a n i n d i vi d u a l ’ s a b i l i ty , e.g ., a n a b i l i ty to b e a b l e to m a n a g e r el a ti on s h i p wi th oth er a g en ts a n d a ct wi s el y i n a s i tu a ti on g over n ed b y a n i m p l i ci t or ex p l i ci t s et of s h a r ed r u l es , b a s ed on a n a b i l i ty of m on i tor i n g a n d u n d er s ta n d i n g of oth er a g en ts ’ m en ta l s ta te. It i s d i s ti n g u i s h ed fr om oth er k i n d s of i n tel l i g en ce s u ch a s p r ob l em s ol vi n g i n tel l i g en ce (a b i l i ty to s ol ve l og i ca l l y com p l ex p r ob l em s ) or em oti on a l i n tel l i g en ce (a b i l i ty to m on i tor on e’ s own a n d oth er s ’ em oti on s a n d to u s e th e i n for m a ti on to g u i d e on e’ s th i n k i n g a n d a cti on s ). Al ter n a ti vel y , s oci a l i n tel l i g en ce m i g h t b e a ttr i b u ted to a col l ecti on of a g en ts a n d d efi n ed a s a n a b i l i ty to m a n a g e com p l ex i ty a n d l ea r n fr om ex p er i en ces a s a fu n cti on of th e d es i g n of s oci a l s tr u ctu r e. Th i s vi ew em p h a s i z es th e r ol e of s oci a l r u l es or cu l tu r e th a t con s tr a i n th e wa y i n d i vi d u a l a g en ts b eh a ve. We m i g h t a ttr i b u te a g ood s oci a l b eh a vi or to a g ood s oci a l s tr u ctu r e a n d con s i d er th a t a g ood s oci a l s tr u ctu r e a ffor d s th e m em b er s of th e com m u n i ty to l ea r n fr om ea ch oth er . In Soci a l In tel l i g en ce D es i g n , we i n ter m i n g l e th es e two vi ews a n d l ook a t b oth s i d es of s oci a l i n tel l i g en ce. Th e " s oci a l i n tel l i g en ce a s a n i n d i vi d u a l ’ s a b i l i ty " vi ew i s r el a ted to d es i g n i n g a p er s on a l a s s i s ta n ce or s oci a l l y i n tel l i g en t a g en ts . On th e oth er h a n d , th e " s oci a l i n tel l i g en ce a s a col l ecti ve a b i l i ty " vi ew i s con cer n ed wi th th e d es i g n of g r ou p /com m u n i ty s u p p or t s y s tem s . Soci a l In tel l i g en ce D es i g n i s tr u l y a n i n ter d i s ci p l i n a r y fi el d . Th e en g i n eer i n g a s p ects of Soci a l In tel l i g en ce D es i g n i n vol ve d es i g n a n d i m p l em en ta ti on of s y s tem s th a t r a n g e fr om g r ou p /tea m or i en ted col l a b or a ti on s u p p or t s y s tem s th a t fa ci l i ta te i n ti m a te, g oa l -or i en ted i n ter a cti on a m on g p a r ti ci p a n ts , to com m u n i ty s u p p or t s y s tem s th a t s u p p or t l a r g e-s ca l e on l i n e-d i s cu s s i on . Th e s ci en ti fi c a s p ects of Soci a l In tel l i g en ce D es i g n a r e con cer n ed wi th cog n i ti ve a n d s oci a l p s y ch ol og i ca l u n d er s ta n d i n g of s oci a l i n tel l i g en ce. In a d d i ti on , econ om y , s oci ol og y , eth i cs a n d m a n y oth er d i s ci p l i n es con s ti tu te th e fou n d a ti on of Soci a l In tel l i g en ce D es i g n . E n g i n eer i n g a p p r oa ch es s h ou l d b e ti g h tl y cou p l ed wi th s oci ol og i ca l a n d cog n i ti ve a p p r oa ch es to p r ed i ct a n d a s s es s th e effects of s oci a l i n tel l i g en ce a u g m en ta ti on s y s tem s on th e h u m a n s oci ety . On th e oth er h a n d , n ovel i n s i g h ts m a y b e ob ta i n ed i n s oci ol og y , cog n i ti ve p s y ch ol og y a n d oth er h u m a n i ty s tu d i es b y i n ves ti g a ti n g a n ew vi r tu a l i z ed s oci ety wh er e h u m a n s a n d a r ti fa cts coh a b i t. Ty p i ca l a p p l i ca ti on s of Soci a l In tel l i g en ce D es i g n a r e g r ou p /tea m s u p p or t s y s tem s a n d com m u n i ty s u p p or t s y s tem s . Com m u n i ty s u p p or t s y s tem s p r ovi d e r a th er l on g -r a n g e, b ottom -u p com m u n i ca ti ve fu n cti on s i n th e b a ck g r ou n d of d a i l y l i fe. Ma j or i s s u es a r e: (i ) ex ch a n g i n g a wa r en es s wi th oth er m em b er s , (i i ) ex p l or i n g h u m a n a n d k n owl ed g e n etwor k s , (i i i ) b u i l d i n g com m u n i ty k n owl ed g e, (i v) or g a n i z i n g p u b l i c even ts , (v) for m i n g a g r ou p /tea m for col l a b or a ti ve wor k , (vi ) h el p i n g n eg oti a te wi th oth er s , a n d (vi i ) s u p p or ti n g p u b l i c d i s cu s s i on s a n d d eci s i on m a k i n g a b ou t th e com m u n i ty . In con tr a s t, g r ou p /tea m s u p p or t s y s tem s focu s on fa ci l i ta ti n g m or e i n ti m a te col l a b or a ti on a m on g m em b er s . Th u s , g r ou p /tea m s u p p or t s y s tem s em p h a s i z e m or e ta s k -d r i ven , s h or t-r a n g e col l a b or a ti on , a l th ou g h a wa r en es s i s eq u a l l y em p h a s i z ed .

1. Soci a l In tel l i g en ce D es i g n – An Over vi ew

5

T a b l e 1 . Hor i z on of Soci a l In tel l i g en ce D es i g n • − − − • − − − − − − • − − − • − − − • − − − − • − −

m eth od s of es ta b l i s h i n g th e s oci a l con tex t a wa r en es s of con n ected n es s [ 16] ci r cu l a ti n g p er s on a l vi ews [ 6] s h a r i n g s tor i es [ 20] em b od i ed con ver s a ti on a l a g en ts a n d s oci a l i n tel l i g en ce k n owl ed g e ex ch a n g e b y vi r tu a l i z ed eg os [ 8 ] con ver s a ti on a l a g en ts for m ed i a ti n g d i s cu s s i on s [ 18 ] a vi r tu a l wor l d h a b i ted b y a u ton om ou s con ver s a ti on a l a g en ts [ 15] s oci a l l ea r n i n g wi th a con ver s a ti on a l i n ter fa ce [ 9] con ver s a ti on s a s a p r i n ci p l e of d es i g n i n g com p l ex s y s tem s [ 7] a r ti fa cts ca p a b l e of m a k i n g em b od i ed com m u n i ca ti on [ 19] col l a b or a ti on d es i g n i n teg r a ti n g th e p h y s i ca l s p a ce, el ectr on i c con ten t, a n d i n ter a cti on [ 3 ] u s i n g m u l ti a g en t s y s tem to h el p p eop l e i n a com p l ex s i tu a ti on [ 2] eva l u a ti n g com m u n i ca ti on i n fr a s tr u ctu r e i n ter m s of col l a b or a ti on s u p p or t [ 11] p u b l i c d i s cou r s e vi s u a l i z a ti on [ 14 ] s oci a l a wa r en es s s u p p or t [ 14 ] i n teg r a ti n g Su r vey s , D el p h i s a n d Med i a ti on for d em ocr a ti c p a r ti ci p a ti on [ 10] th eor eti ca l a s p ects of s oci a l i n tel l i g en ce d es i g n u n d er s ta n d i n g g r ou p d y n a m i cs of k n owl ed g e cr ea ti on [ 1] u n d er s ta n d i n g con s en s u s for m a ti on p r oces s [ 13 ] th eor y of com m on g r ou n d i n l a n g u a g e u s e [ 17] a tta ch m en t-b a s ed l ea r n i n g for s oci a l l ea r n i n g [ 12] eva l u a ti on of s oci a l i n tel l i g en ce n etwor k a n a l y s i s [ 5] h y b r i d m eth od [ 4 ]

Th e s cop e of Soci a l In tel l i g en ce D es i g n a s a d i s ci p l i n e of u n d er s ta n d i n g a n d s u p p or ti n g s oci a l i n tel l i g en ce i s s u m m a r i z ed i n Ta b l e 1. On th e on e h a n d , Soci a l In tel l i g en ce D es i g n i s con cer n ed wi th d es i g n a n d i m p l em en ta ti on of n ovel com m u n i ca ti on m ea n s for m ed i a ti n g i n ter a cti on a m on g p eop l e a n d a g en ts . Th e s cop e r a n g es fr om p r el i m i n a r y a n d p r ep a r a tor y i n ter a cti on s a m on g p eop l e s u ch a s k n owi n g wh o’ s wh o, to m or e i n ti m a te i n ter a cti on s u ch a s col l a b or a ti on . Su p p or ti n g a g r ou p for m a ti on , col l a b or a ti on , n eg oti a ti on , p u b l i c d i s cu s s i on or s oci a l l ea r n i n g i s con s i d er ed to b e a n i m p or ta n t a p p l i ca ti on of Soci a l In tel l i g en ce D es i g n . Th eor eti ca l a s p ects , a s wel l a s p r a g m a ti c a s p ects , s h ou l d b e ta k en i n to a ccou n t i n d es i g n i n g , d ep l oy i n g , a n d eva l u a ti n g s oci a l i n tel l i g en ce s u p p or t tool s . In th e r es t of th i s s ecti on , I wi l l over vi ew m a j or i s s u es i n Soci a l In tel l i g en ce D es i g n .

6

T. N i s h i d a

1 .2 .1 M e t h o d s o f E s t a b lis h in g t h e S o c ia l C o n t e x t Th e com m on g r ou n d n eed to b e es ta b l i s h ed i n or d er for s oci a l i n tel l i g en ce to em er g e fr om i n ter a cti on a m on g a g en ts , es p eci a l l y wh en th e a g en ts a r e l oca ted i n a g eog r a p h i ca l l y d i s ta n t p l a ces . A s u b fi el d of Soci a l In tel l i g en ce D es i g n i s d evoted to th e d es i g n of a n ew com m u n i ca ti on m ed i u m for a com m u n i ty or a g r ou p . Th e r ol e of com m u n i ca ti on m ed i u m i s n ot on l y to m eet p r i m a r y com m u n i ca ti on g oa l s , i .e., tr a n s m i tti n g a n i n ten d ed con ten t, b u t a l s o p r ovi d i n g con tex tu a l i n for m a ti on th a t m a y h el p i n ter p r et th e con ten t. It i s often th e ca s e i n ou r d a i l y l i fe th a t con ver s a ti on i s n ot for a ch i evi n g h i g h er -l evel g oa l s s u ch a s i n for m a ti on s eek i n g , b u t m er el y for s oci a l i n ter a cti on s u ch a s m a i n ta i n i n g h u m a n r el a ti on . Su ch s oci a l i n ter a cti on i s i m p or ta n t to con s ti tu te a s oci a l con tex t s u ch a s tr u s t. On e a p p r oa ch i s to s u p p or t s oci a l a wa r en es s . Oh g u r o p r op os es to s u p p or t th e a wa r en es s of con n ected n es s wi th F a i n tP op , wh i ch i s a n on ver b a l com m u n i ca ti on d evi ce s i m i l a r to a p h oto fr a m e [ 16] i n wh i ch s m a l l p h otos or i con s of th e u s er ’ s col l ea g u es a r e d i s p l a y ed . F a i n tP op a l l ows th e u s er to com m u n i ca te h er /h i s feel i n g towa r d s h er /h i s col l ea g u es b y u s i n g th e th r ee ty p es of tou ch i n g (a ta p to com m u n i ca te a n eu tr a l feel i n g , a p et a p os i ti ve feel i n g , a n d a h i t a n eg a ti ve feel i n g ). In con tr a s t, on e m a y d es i g n a ver b a l com m u n i ca ti on m ed i u m to ex ch a n g e m or e ex p l i ci t i n for m a ti on . Th e P u b l i c Op i n i on Ch a n n el (P OC) [ 6] i s a com m u n i ty -wi d e i n ter a cti ve b r oa d ca s ti n g s y s tem . A P OC con ti n u ou s l y col l ects m es s a g es fr om p eop l e i n a com m u n i ty a n d b r oa d ca s ts s u m m a r i z ed m es s a g es to th e com m u n i ty . P OC i s n ot i n ten d ed to b e a s y s tem th a t b r oa d ca s ts p u b l i c op i n i on s p er s e. In s tea d , i t i s i n ten d ed to b r oa d ca s t p eop l e’ s p er s on a l vi ews a r i s i n g i n a d a i l y l i fe, e.g ., q u es ti on s , s tor i es , fi n d i n g s , j ok es , or p r op os a l s . Th es e m es s a g es a r e con s i d er ed to for m a s oci a l con tex t th a t ca n s er ve a s a b a s i s of p u b l i c op i n i on for m a ti on . IB M’ s Wor l d Ja m [ 20] i s a l a r g e-s ca l e cor p or a te-wi d e d i s cu s s i on wh er ei n a l l IB Mer s wor l d wi d e a r e i n vi ted to p a r ti ci p a te i n . Th e s y s tem p r ovi d es a n i n ter fa ce th a t a l l ows ea ch p a r ti ci p a n t to q u i ck l y s ee th e con cu r r en t vi ew of wh o el s e i s p r es en t a n d wh i ch top i cs a r e b ei n g d i s cu s s ed . Th om a s s u g g es ts th a t k ey s to i n n ova te a r e wi th d es i g n i n g i n ter fa ce th a t ca n (i ) fa ci l i ta te en g a g em en t, (i i ) a l l ow th e u s er to b r i n g to b ea r n eces s a r y s k i l l s , ta l en ts , a n d k n owl ed g e s ou r ces on th e p r ob l em , a n d (i i i ) u s e a p p r op r i a te r ep r es en ta ti on s of th e s i tu a ti on . He a l s o p oi n ts ou t th e i m p or ta n ce of s tor i es a n d or g a n i z a ti on a l i s s u es . Stor i es a l l ow th e u s er to a s s oci a te th e con ten t wi th p r evi ou s ex p er i en ce. Th e or g a n i z a ti on a l s tr u ctu r e con s i s ti n g of s u ch p eop l e a s m od er a tor s a n d fa ci l i ta tor s p l a y s a cr i ti ca l r ol e i n th e Wor l d Ja m l a r g e-s ca l e d i s cu s s i on ex p er i m en t. 1 .2 .2 E m b o d ie d C o n v e r s a t io n a l A g e n t s a n d S o c ia l I n t e llig e n c e Con ver s a ti on p l a y s va r i eti es of r ol es i n h u m a n s oci eti es . It n ot on l y a l l ows p eop l e to ex ch a n g e i n for m a ti on i n a ca s u a l fa s h i on , b u t i t a l s o h el p s th em cr ea te n ew i d ea s or m a n a g e h u m a n r el a ti on s . E m b od i ed con ver s a ti on a l a g en ts ca n b e u s ed to a u g m en t s oci a l i n tel l i g en ce b y m ed i a ti n g con ver s a ti on s a m on g p eop l e. Ku b ota a n d N i s h i d a u s e a ta l k i n g vi r tu a l i z ed -eg os m eta p h or i n E g oCh a t [ 8 ] to en a b l e a s op h i s ti ca ted a s y n ch r on ou s

1. Soci a l In tel l i g en ce D es i g n – An Over vi ew

7

com m u n i ca ti on a m on g com m u n i ty m em b er s . A vi r tu a l i z ed eg o m a i n l y p l a y s two fu n cti on s . F i r s t, i t s tor es a n d m a i n ta i n s th e u s er ’ s p er s on a l m em or y . Secon d , i t p r es en ts th e con ten t of th e p er s on a l m em or y on b eh a l f of th e u s er a t a p p r op r i a te s i tu a ti on s . A vi r tu a l i z ed eg o s er ves a s a p or ta l to th e m em or y a n d k n owl ed g e of a p er s on . It a ccu m u l a tes i n for m a ti on a b ou t a p er s on a n d a l l ows h er /h i s col l ea g u e to a cces s th e i n for m a ti on b y fol l owi n g a n or d i n a r y s p ok en -l a n g u a g e con ver s a ti on m od e, n ot b y g oi n g u p a n d d own a com p l ex d i r ector y i n s ea r ch for p os s i b l y ex i s ten t i n for m a ti on , or b y d el i b er a tel y i s s u i n g com m a n d s for i n for m a ti on r etr i eva l . In a d d i ti on , vi r tu a l i z ed eg o m a y em b od y ta ci t a n d n on -ver b a l k n owl ed g e a b ou t th e p er s on s o th a t m or e s u b tl e m es s a g es s u ch a s a tti tu d e ca n b e com m u n i ca ted . Ta k a h a s h i a n d Ta k ed a u s e a va ta r -l i k e con ver s a ti on a l a g en ts i n a s i m i l a r vei n [ 18 ] . Th e u s er ca n u s e h er /h i s a g en t to g i ve com m en ts on a web p a g e. It ex ten d s col l a b or a ti ve a n n ota ti on i n s u ch a wa y th a t th e u s er s ca n en cod e s u b tl e feel i n g s i n em oti on a l ex p r es s i on s of a g en ts . In a m or e s op h i s ti ca ted a p p l i ca ti on s , b u i l d i n g a r i ch con ver s a ti on a l en vi r on m en t b ecom es m or e i m p or ta n t. N i j h ol t a r g u es b u i l d i n g a th ea ter en vi r on m en t th a t p r ovi d es th e u s er wi th a n i n for m a ti on -r i ch vi r tu a l en vi r on m en t m i m i ck i n g r ea l th ea ter b u i l d i n g s i n a r ea l town wh er e a u ton om ou s a g en ts wi th va r y i n g a b i l i ti es coh a b i t [ 15] . Th e th ea ter en vi r on m en t a l l ows th e u s er to b e i m m er s ed i n th e vi r tu a l wor l d a n d fol l ow con ti n u ou s ver b a l /n on ver b a l i n ter a cti on s wi th a g en ts . He i s i n tr od u ci n g th e i n ter n a l m od el of a u ton om ou s a g en ts i n ter m s of b el i efs , d es i r es , p l a n s , a n d em oti on s , to r ea l i z e a th ea ter com m u n i ty . Con ver s a ti on a l ch a r a cter s ca n a l s o b e em p l oy ed i n th e l ea r n i n g en vi r on m en t. Ka k tu s i s a com p u ter g a m e en vi r on m en t th a t i s d es i g n ed s o th a t th e teen a g er s tu d en t ca n i n ter a ct wi th s em i -a u ton om ou s em oti on a l l y i n tel l i g en t ch a r a cter s to l ea r n s oci o-em oti on a l r el a ti on s [ 9] . Th e n oti on s of con ver s a ti on s a n d s oci a l i n tel l i g en ce a r e u s efu l i n d es i g n i n g com p l ex s y s tem s . G og u en s u g g es ts ex p er i m en ti n g wi th a n a p p r op r i a te b l en d of i n ter a cti on m eta p h or s i n b u i l d i n g i n ter fa ces to th eor em p r over s [ 7] . In th e r ea l wor l d a p p l i ca ti on s , m or e i s s u es s u ch a s em b od i m en t s h ou l d b e ta k en i n to a ccou n t. Ter a d a a n d N i s h i d a d i s cu s s d es i g n i n g a n a r ti fa ct ca p a b l e of m a k i n g em b od i ed com m u n i ca ti on wi th p eop l e a n d oth er a g en ts [ 19] . An i n ter es ti n g i s s u e i s h ow on e ca n a l l ow a g en ts wi th d i ffer en t em b od i m en t to com m u n i ca te wi th ea ch oth er . 1 .2 .3 C o lla b o r a t io n D e s ig n Col l a b or a ti on d es i g n i s con cer n ed wi th g oa l -or i en ted , m or e i n ti m a te i n ter a cti on . In a d d i ti on to b a s i c com m u n i ca ti on fa ci l i ti es , th e n a tu r e of i n ter a cti on i n col l a b or a ti ve a cti vi ti es s h ou l d b e s tu d i ed i n d eta i l . P r i n ci p l es a n d g u i d el i n es a r e n eces s a r y to d es i g n col l a b or a ti on s u p p or t s y s tem s . F r u ch ter p oi n ts ou t th a t i t i s b en efi ci a l to con s i d er i n ter m s of th r ee p er s p ecti ves , n a m el y , p h y s i ca l s p a ces (" b r i ck s " ), el ectr on i c con ten t (" b i ts " ) a n d th e wa y p eop l e com m u n i ca te wi th ea ch oth er (" i n ter a cti on " ) [ 3 ] . Sh e s u g g es ts th a t b y p r op er l y u n d er s ta n d i n g th e r el a ti on s h i p b etween b r i ck s , b i ts , a n d i n ter a cti on , on e ca n d es i g n s p a ces th a t b etter a ffor d com m u n i ca ti ve even ts , d evel op col l a b or a ti on tech -

8

T. N i s h i d a

n ol og i es th a t ca n b es t s u p p or t th e j oi n t a cti vi ti es of p eop l e, a n d en g a g e p eop l e i n r i ch com m u n i ca ti ve ex p er i en ces th a t en a b l e th em to i m m er s e i n th ei r a cti vi ty . Som eti m es , e.g ., i n th e ca s e of em er g en ce, i t i s d es i r a b l e for a col l ecti on of p eop l e to b e g u i d ed b y s oci a l l y i n tel l i g en t a g en ts i n or d er to a voi d a p a n i c. Ca r d on p r op os es to em p l oy a h i er a r ch y of m u l ti p l e a g en ts con s i s ti n g of wh a t h e ca l l s a s p ectu a l a g en ts a n d m or p h ol og i ca l a g en ts [ 2] . In or d er to cop e wi th th e ou tcom e of u n ex p ected s tr u ctu r e, h e em p h a s i z es th e i m p or ta n ce of a m ech a n i s m th a t a l l ows m ea n i n g to b e d y n a m i ca l l y g en er a ted i n com m u n i ca ti on . Th e com m u n i ca ti on i n fr a s tr u ctu r e m a y i n fl u en ce th e wa y of d i s ta n t col l a b or a ti on . F or ex a m p l e, r ep l a ci n g HD TV (Hi g h D efi n i ti on TV) b y n or m a l vi d eo m a y m a k e a q u a l i ta ti ve d i ffer en ce i n col l a b or a ti on s ty l e. Ma r k a n d D eF l or i o s u g g es t th a t s i n ce th e HD TV p r ovi d es h i g h -r es ol u ti on i m a g e, p eop l e d o n ot u s e ex a g g er a ted g es tu r es or m ovem en ts to con vey ex p r es s i on th r ou g h th e HD TV i m a g e, wh i ch wa s r ep or ted to h a p p en i n n or m a l vi d eocon fer en ces [ 11] . 1 .2 .4 P u b lic D is c o u r s e A g r ou p /com m u n i ty /s oci ety a s a wh ol e h a s to m a k e d eci s i on fr om ti m e to ti m e. E ffecti ve u s e of i n for m a ti on a n d com m u n i ca ti on tech n ol og i es a r e s ou g h t to s u p p or t p u b l i c d i s cu s s i on a n d d eci s i on -m a k i n g . N a k a ta a r g u es th a t cr i ti ca l i s s u es i n d es i g n i n g a d i s cu s s i on s u p p or t s y s tem a r e (i ) ea s e of i n for m a ti on a cces s a n d p r oa cti ve i n for m a ti on g a th er i n g , (i i ) u s er -fr i en d l y a cces s to a s ci en ti fi c a n a l y s i s tool k i t, (i i i ) eva l u a ti on of d el i b er a ti ve s ta tes , a n d (i v) g u i d i n g d i s cu s s i on th r ou g h d i s cu s s i on a n d con s en s u s g en er a ti on m od el s [ 14 ] . He a l s o p oi n ts ou t th e i m p or ta n ce of s u p p or ti n g i n d i vi d u a l s s o th a t th ey ca n col l ect a n d ex ch a n g e i n for m a ti on a n d op i n i on s . In for m a ti on a n d com m u n i ca ti on tech n ol og y m i g h t b r i n g a n ovel p a r ti ci p a ti on a n d d i s cu s s i on s ch em e i n to d em ocr a cy . L u eh r s et a l [ 10] a ttem p t a t com b i n i n g s u r vey tech n i q u es , d el p h i a p p r oa ch es , a n d m ed i a ti on m eth od i n to a n ew m eth od ol og y for on -l i n e d em ocr a ti c p a r ti ci p a ti on a n d i n ter a cti ve con fl i ct r es ol u ti on . Th ei r s y s tem i n teg r a tes m a s s op i n i on p ol l s , cy cl i ca l d eci s i on -m a k i n g p r oces s ex p l oi ti n g ex p er t k n owl ed g e, a n d a n op en p r oces s of p a r ti ci p a ti ve con fl i ct r es ol u ti on , a d a p ted fr om Su r vey s , D el p h i , a n d Med i a ti on , r es p ecti vel y . 1 .2 .5 T h e o r e t ic a l A s p e c t s o f S o c ia l I n t e llig e n c e D e s ig n Th eor i es p l a y s ever a l r ol es i n Soci a l In tel l i g en ce D es i g n . In a d d i ti on to th ei r p r i n ci p a l r ol e of p r ovi d i n g a fr a m ewor k for u n d er s ta n d i n g p h en om en on , th eor i es tel l u s m or e d i r ect i m p l i ca ti on s s u ch a s g u i d el i n es of d es i g n i n g com m u n i ty /g r ou p s u p p or t s y s tem s or a n i n ven tor y of k n own p i tfa l l s th a t s h ou l d b e ta k en i n to a ccou n t i n s y s tem d es i g n . In s oci a l p s y ch ol og y , n otor i ou s ex a m p l es s u ch a s g r ou p th i n k (i .e., a p h en om en on th a t col l ecti ve cr ea ti vi ty d oes n ot ex ceed i n d i vi d u a l cr ea ti vi ty ) or th e h os ti l i ty to ou t-g r ou p s (i .e., a g r ou p m em b er h a s h os ti l i ty to ou t-g r ou p s ea s i l y ) a r e k n own to h i n d er effecti ve k n owl ed g e cr ea ti on i n a n etwor k ed com m u n i ty . Az ech i cl a s s i -

1. Soci a l In tel l i g en ce D es i g n – An Over vi ew

9

fi es th e con ten t of a m es s a g e i n to d r y a n d wet i n for m a ti on [ 1] . D r y i n for m a ti on p r i m a r i l y con ta i n s l og i ca l l i n g u i s ti c i n for m a ti on a n d con s ti tu tes th e cor e of a m es s a g e. In con tr a s t, wet i n for m a ti on i s m a i n l y n on l i n g u i s ti c, m eta -i n for m a ti on i n ci d en ta l to th e con ten ts of th e m es s a g e. Az ech i a r g u es th a t com m u n i ty -wi d e d i s cu s s i on for a ch i evi n g s om e p r a cti ca l g oa l s h ou l d b e m a d e on l y wi th d r y i n for m a ti on , oth er wi s e r a ti on a l d i s cu s s i on wi l l b e h i n d er ed d u e to th e p a th ol og y of a g r ou p . Ma ts u m u r a a d d r es s es th e con s en s u s for m a ti on i n n etwor k ed com m u n i ti es [ 13 ] . B a s ed on s oci a l p s y ch ol og i ca l ex p er i m en ts , h e h a s fou n d th a t (i ) m i n or i ty m em b er s ten d to over es ti m a te th e n u m b er of oth er m em b er s wh o s h a r e th e s a m e a tti tu d e, (i i ) m i n or i ty m em b er s ten d to u n d er es ti m a te th e a tti tu d e of oth er m em b er s , (i i i ) m i n or i ty m em b er s wh o u n d er es ti m a te th e p r op or ti on of th e m i n or i ty ’ s op i n i on ten d to l os e a n i n ten ti on to a ct. Su ch i n a ccu r a cy i n cog n i ti on of op i n i on d i s tr i b u ti on i s ca l l ed th e fa l s e con s en s u s effect. Th es e ob s er va ti on s s h ou l d b e ta k en i n to a ccou n t i n d es i g n i n g d i s cu s s i on s u p p or t s y s tem s s o th a t u s efu l d i s cu s s i on s ca n b e ex p ected b y r efl ecti n g m i n or i ty op i n i on s . Th eor i es of l a n g u a g e u s e i n i n ter a cti on a r e r el eva n t to es ta b l i s h i n g th e com m on g r ou n d i n col l a b or a ti on . R os en b er g s u g g es ts th a t k ey i s s u es a r e i n for m a ti on i n teg r a ti on i n to a com m on g r ou n d , th e r el a ti on b etween l i n g u i s ti c ch a n n el s a n d s h a r ed k n owl ed g e, a n d th e m ech a n i s m of r eta i n i n g s h a r ed k n owl ed g e i n th e com m on g r ou n d of d i ffer en t k i n d s of p a r ti ci p a n t [ 17] . In th e con tex t of s oci a l l ea r n i n g , Ma r l ow a n d P er etti ex p l or e a tta ch m en t-b a s ed l ea r n i n g com p r i s i n g r es p on s e i m p r i n ti n g a n d m i m i cr y . Th ey h a ve b u i l t a l ea r n i n g en vi r on m en t to tes t th e h y p oth es i s [ 12] . 1 .2 .6 E v a lu a t io n s o f S o c ia l I n t e llig e n c e Soci a l In tel l i g en ce D es i g n i s cer ta i n l y a n em p i r i ca l s tu d y . We h a ve to r ep ea t th e d es i g n -i m p l em en t-eva l u a ti on cy cl e u n ti l we r ea ch b etter s y s tem s . N etwor k An a l y s i s i s a p ower fu l m ea n s of eva l u a ti n g or com p a r i n g em p i r i ca l d a ta . It p r ovi d es u s wi th a m ea n s for ca l cu l a ti n g va r i ou s a s p ects of a g i ven n etwor k i n ter m s of cen tr a l i ty , d en s i ty or coh es i on . B y com p a r i n g th os e fea tu r es fr om on e n etwor k a g a i n s t th os e fr om a n oth er , we ca n d es cr i b e th e s i m i l a r i ty a n d d i ffer en ce i n q u a n ti ta ti ve ter m s . F u j i ta h a s con d u cted a fi el d tr i a l a n d em p l oy ed n etwor k a n a l y s i s to s h ow th e effecti ven es s of th ei r com m u n i ty s u p p or t s y s tem [ 5] . F u j i h a r a h a s a l s o a p p l i ed n etwor k a n a l y s i s to a l og col l ected fr om ex p er i m en ts wi th a P OC p r ototy p e [ 6] for s ever a l m on th s to s ee i f P OC a ctu a l l y fa ci l i ta tes com m u n i ty k n owl ed g e cr ea ti on [ 4 ] . He a l s o p oi n ts ou t th a t n etwor k a n a l y s i s a l on e i s n ot en ou g h to eva l u a te com m u n i ty s u p p or t s y s tem s , a n d h en ce i t s h ou l d b e com b i n ed wi th s ever a l oth er m eth od s s u ch a s th e u s er ’ s s u b j ecti ve a n a l y s i s or l og a n a l y s i s .

10

T. N i s h i d a

1 .3 C o n c lu d in g R e m a r k s Soci a l In tel l i g en ce D es i g n i s a d i s ci p l i n e a i m ed a t u n d er s ta n d i n g a n d s u p p or s oci a l i n tel l i g en ce. In th i s p a p er , I h a ve over vi ewed m a j or i s s u es i n vol ved i n ci a l In tel l i g en ce D es i g n a n d a ttem p ted a t s tr u ctu r e th em i n a coh er en t s tor y . con tem p or a r y vi ew of Soci a l In tel l i g en ce D es i g n con s i s ts of m eth od s of es l i s h i n g th e s oci a l con tex t, em b od i ed con ver s a ti on a l a g en ts , col l a b or a ti on d es p u b l i c d i s cou r s e, th eor eti ca l a s p ects of s oci a l i n tel l i g en ce d es i g n , a n d eva l u a of s oci a l i n tel l i g en ce.

ti n g SoTh e ta b ig n , ti on

R e fe r e n c e s 1. 2. 3 . 4 . 5. 6. 7. 8 . 9. 10. 11. 12. 13 . 14 . 15. 16. 17. 18 . 19. 20.

Az ech i , S.: Com m u n i ty b a s ed s oci a l i n tel l i g en ce, SID -2001, 2001. Ca r d on , A.: A d i s tr i b u ted m u l ti -a g en t s y s tem for th e s el f-eva l u a ti on of d i a l og s , i n th i s vol u m e. F r u ch ter , R .: B r i ck s & b i ts & i n ter a cti on , i n th i s vol u m e. F u j i h a r a , N .: How to eva l u a te s oci a l i n tel l i g en ce d es i g n , i n th i s vol u m e. F u j i ta , K. et a l : P os s i b i l i ty of n etwor k com m u n i ti es : n etwor k a n a l y s i s of a com m u n i ty or g a n i z er ex p er i m en t, i n th i s vol u m e. F u k u h a r a , T. et a l : P u b l i c Op i n i on Ch a n n el : a s y s tem for a u g m en ti n g s oci a l i n tel l i g en ce of a com m u n i ty , i n th i s vol u m e. G og u en , J.: Ar e a g en ts a n a n s wer or a q u es ti on ? , p u b l i s h -on l y p a p er , SID -2001, 2001. Ku b ota , H. a n d N i s h i d a , T.: Kn owl ed g e cr ea ti n g con ver s a ti on a l a g en ts , SID -2001, 2001. L a a k s ol a h ti , J., P er s s on , P ., a n d P a l o, C.: Ka k tu s : a s oci o-em oti on a l l y r i ch i n ter a cti ve n a r r a ti ve, p u b l i s h -on l y p a p er , SID -2001, 2001. L u eh r s , R ., Ma l s ch , T., a n d Vos s , K.: In ter n et, D i s cou r s es a n d D em ocr a cy , i n th i s vol u m e. Ma r k , G . a n d D eF l or i o, P .: HD TV: a ch a l l en g e to tr a d i ti on a l vi d eo con fer en ci n g ? , p u b l i s h -on l y p a p er s , SID -2001, 2001. Ma r l ow, C. a n d P er etti , J.: Mod el i n g s oci a l i n tel l i g en ce th r ou g h a tta ch m en t-b a s ed l ea r n i n g , p u b l i s h -on l y p a p er s , SID -2001, 2001. Ma ts u m u r a , K.: Con s en s u s for m a ti on p r oces i n n etwor k com m u n i ty , SID -2001, 2001. N a k a ta , K.: E n a b l i n g p u b l i c d i s cou r s e, i n th i s vol u m e. N i j h ol t, A.: F r om vi r tu a l en vi r on m en t to vi r tu a l com m u n i ty , i n th i s vol u m e. Oh g u r o, T. et a l : F a i n tP op : In tou ch wi th th e s oci a l r el a ti on s h i p s , i n th i s vol u m e. R os en b er g , D .: Com m u n i ca ti ve a s p ects of s oci a l i n tel l i g en ce, p u b l i s h -on l y p a p er , SID 2001, 2001. Ta k a h a s h i , T. a n d Ta k ed a , T.: N a r r a ti ve ed i ti n g of web con tex ts on on l i n e com m u n i ty s y s tem wi th a va ta r -l i k e a r ti fa cts , SID -2001, 2001. Ter a d a , K. a n d N i s h i d a , T.: E m b od i ed com m u n i ca ti on b etween h u m a n s a n d a r ti fa cts , SID -2001, 2001. Th om a s , J.: Col l a b or a ti ve i n n ova ti on tool s , i n th i s vol u m e. SID -2001: JSAI-Sy n s op h y In ter n a ti on a l Wor k s h op on Soci a l In tel l i g en ce D es i g n , Ma ts u e, Ja p a n , Ma y 21-22, 2001.

2. FaintPop: In Touch with the Social Relationships Takeshi Ohguro1 , Kazuhiro Kuwabara2 , Tatsuo Owada2 , and Yoshinari Shirai2 1 2

NTT-ME Corporation, Marketing HQ V, Tokyo 100-8132, Japan. email: [email protected] NTT Communication Science Laboratories, NTT Corporation, Kyoto 6190237, Japan

We propose a tool called FaintPop. It is intended to be an alternative media that is suitable for a very light-weight, acknowledge-only, mode of communication. Furthermore, it intuitively provides, through memories of communication, a general overview of the communication activities. The tool is designed for a community, with which the sense of connectedness can be shared among members. Results from an initial experiment are reported briefly.

2.1 Social Intelligence Design for Communications Although the IT (Information Technology) bubble is said to have burst, the Internet and IT remain essential and are experiencing continued significant advances. There are several evidences that support the trend, just to mention an example: Mobile phone services are rushing toward the 3G era in which more ubiquitous and broadband communications will be fully utilized. The trend shows that our lifestyle, as well as our society, is surely being impacted by the Internet and IT. It is hard nowadays to imagine to work, live, or communicate without the network. Now the important question is determining what design will best augment social intelligence for the network age. More specifically, we focus on the communication environment for emerging networked societies, since communication is the very basis of the societies. In challenging the question, we first look at the problem that is currently appearing and would increase in the future. The problem, Communication Overflow [2.15], consists of two related subproblems. One is that our opportunities for communications are much greater than ever before. This trend is sometimes so overwhelming that our communications become segmented into pieces, that we lose the general view on our own communication activities. The other is that we do not have enough variants of network communication media to support the various communication modes common in our daily lives. For example, current network communication media seem too heavy for simply saying “Hi,” which is a frequently-used communication mode in physical environments. Using non-suitable media requires cognitive load. The notion of “Communication Overflow” is closely related to the problem of “Information Overflow.” However, our focus is not information itself. In T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 11− 18 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

12

T. Ohguro et al.

other words, our primary focus is not on the “developer’s side,” which mainly addresses the tools and abilities offered by technologies. Instead, we focus on the experience of users [2.18]. In other words, our primary focus is on the “user’s side,” that mainly concerns how tools are used, in what situations, and by whom. Therefore we use the term “Communication Overflow,” to clarify that the problem is with the possible overflow on users’ opportunities and awareness on his/her own communication activities. To answer the above problem, we proposed a new communication environment [2.14]. Awareness of Connectedness is the key notion for understanding the environment. Here we focus on the awareness of (the communication activities of) oneself. Moreover, to transmit and share the sense of connectedness (awareness of “connected” status with others) are also primary concern. This is contrasted to the term “awareness” used in the area of groupware, in which awareness information of the other participants involved in the current communication is the central issue, where the information is to supplement the contents of communication (e.g., [2.7]). Two candidate tools for the environment are introduced. One is called the Indicator, which is intended to provide feedback of the user’s communication activities [2.4]. It provides a general overview of user’s communication activities, which is easily lost in the current segmentation of communications. The other is called Gleams of People, which is a simple, intuitive interactive media that exchanges the presence and statuses of users [2.16]. It is designed to be an alternative communication media which is very light-weight and suitable for the acknowledge-only mode of communication. As the first tools for the new communication environment, both tools were designed for personal use, since the individuals’ awareness of communication and connectedness is fundamental to social communications. The aspects of communities and societies are not addressed directly by the tools, though they can be derived implicitly as participants of communications in individuals’ communication activities. However since we mostly belong and act in communities, tools that address these aspects will also be needed. Therefore, in this paper, we introduce the third tool for the communication environment. The tool, called FaintPop, subsumes the functions of the two tools mentioned above, but is designed to be a media for a community. More specifically, it provides an alternative communication media that is very light-weight and suitable for the acknowledge-only mode of communication, through which the sense of connectedness will be shared across the community. Moreover, it provides the general overview of communication activities in the community. The tools works in a suggestive way [2.15]. That is, it does not provide logical analysis such as comment chains or statistics; Instead, a general view is provided that offers a more intuitive but vague picture of what’s is going on in the community. In this way, the tool will retain the social relationships among the community members.

2. FaintPop: In Touch with the Social Relationships

13

2.2 In Touch with the Social Relationships We incorporated a scenario-based technique [2.3] in designing the tool. The scenario: This snapshot shows my old friends. Most of us live far apart, but our friendship might remain. Sometimes I feel like contacting them, but find it hard to do so. Isn’t odd to make a phone call or to write a letter without any “important” business? What I want is mere a slight touch that we still are the friends; Just a faint sense of connectedness would suffice. I wonder if I can do this using just this snapshot. FaintPop implements this scenario; It is a media for sharing the sense of connectedness in a community. Messages exchanged using this media are a sort of things that not so important to talk to, but worth expressing. That is, the communication established by this media is not about important business matters, but about feelings which are very important in social relationships. The communication does not involve written or spoken language, the more intuitive technique of touching is employed. Moreover, memories of communication are summarized and represented graphically in an intuitive way. It gives the users a general view of what’s going on in the community. Figure 2.1 shows two FaintPop prototypes. It is a hardware device shaped to resemble a photo frame. Each member of the community has his/her own device, and all pictures are the same initially. All of them have networked. Instead of using real photographs, small pictures (or icons) of faces of all members of the community (possibly extracted from the original photo) are displayed. Members can communicate each other by touching the images/icons of friends. Written or spoken language is not supported, because in case of contents-oriented mode of communication where such languages are involved, conventional media such as e-mail and phone are more suitable. Instead, FaintPop is oriented to the very light-weight, acknowledge-only mode of communications, in which to notify some content is not the main objective but to share the sense of connectedness is the main purpose.

Fig. 2.1. FaintPop prototype

14

T. Ohguro et al.

Three touching types are provided. A tap to represent neutral feeling, a pet to represent more positive feeling, and a hit to represent rather negative feeling. Due to the limitation of the device, these three types are currently implemented as a click, a long click, and a double click, respectively. Most users easily learned to input the right type by touching the screen with his/her finger. Our current design choice is to offer just these three types. Weather these three are enough remains to be confirmed. Some studies indicate that six or more basic emotions exist [2.1]. However some of these emotions (e.g., fear and anger) are not appropriate for the light-weight, acknowledge-only mode of communication. Furthermore, providing too many types would confuse users, complicate the interface, and conflict with the objective of the media. Touching a picture of a friend means that one of the three feelings are passed to that friend. The touch is encoded and distributed (via the network connections) to all the members in the community, so that all members can share what is going on in the community. The sending of a touch is displayed in all member’s screen as an animation effect: A small ball travels from the sender to the recipient, with different colors and speeds according to the three types of touching. The picture of a friend who received a message with positive (negative) feeling blinks larger (resp. smaller) for a while. For neutral feeling, the picture oscillates for a while. Touching his/her own picture means to broadcast a message to all community members —In other word, the user calls out the community in that feeling. Figure 2.2 shows the animation effect that a user broadcasts a message with negative feeling. The tool is modeled after a photo frame so that it can be placed and embedded naturally in daily lives. Therefore, the interface should not annoy the user such as flashing the whole screen. However such non-disturbing design has a drawback that the user possibly miss the communication that being held. One solution is to use a faint sound to indicate that some activity related to the user is occurring (for example, a message from another user is arrived). Different sounds are used according to the three types of messages.

Fig. 2.2. Screen image of FaintPop. A broadcasting message, traces of communications and recent activities of users are shown.

2. FaintPop: In Touch with the Social Relationships

15

Another technique is to provide memories of communications. In the background of pictures, traces of animation effects, which corresponds to the communications held, are left. Moreover, the informations that which members have been actively communicated recently is indicated by the changing color of the bulls-eye surrounding the picture of the friends; It is represented as a pie chart that indicates which types of messages are sent by the user recently (Figure 2.2). These memories gradually disappear with time. This provides a general view of communications in the community. The feature well suits the nature of the tool; One of the typical use case is that the user would glance at the “photo frame” occasionally and notice that something had happened among the friends. Therefore the tool has the aspect of asynchronous communication media, in addition to the aspect of synchronous communication by touches. Hence, memory retention periods range from hours to a day, longer than those of most (synchronous) communication media. The touches that users make are not only visualized as animation effects and memories of communication but affect the default locations of each picture of friend. FaintPop holds the parameters of closeness, which are naturally asymmetric, between the friends. When the user touches a friend positively (negatively), the acquaintance parameter from the user to the friend is increased (resp. decreased). Then the picture of the friend is moved closer to (resp. apart from) that of the user. Therefore, the locations of pictures displayed on a user’s “photo frame” represent the closeness from the user to the friends. A single touch triggers just a slight change. Again, this is the long term effect as so is similar to the memories of communication. The user can know the closeness between friends (or the closeness from friends to the user) by dragging the picture of friends (or self). When the picture of friend A is dragged to that of friend B, the picture of friend B responds. If the acquaintance parameter from B to A is high (low), the picture of B moves close to (resp. apart from) the picture of A. Note that the acquaintance parameters are asymmetric: Dragging the picture of B to that of A may cause different move. This effect ends when the user stops the dragging, and all the pictures are returned to their default locations. The dragging itself do not generate a message nor is shared among the friends, but the information that the user performed dragging is distributed. It is shown in the activity summary (pie chart) surrounding the picture of the user. Privacy and one-to-one communication issues are important but not addressed directly by this media. It is because our main focus is to provide an alternative communication media that will allow the sense of connectedness to be shared in the community. For one-to-one communication, we have introduced another media called Gleams of People [2.16]. It might be desirable to integrate these media in the future.

16

T. Ohguro et al.

2.3 Initial Experiment To verify whether the basic objectives of the tool were accomplished, we conducted a preliminary experimental study. To match our scenario (section 2.2), 6 subjects of similar ages who knew each other, that reside in different office locations and belong to different work teams, were selected from our laboratory members. Before the experiment, the subjects were instructed the basic usage of the tool. However, the objective of the tool, as well as when and in what purpose they were supposed to use the tool, was not explained. A traffic log of the tool was collected during the experiment. After the one-week experiment, subjects were asked to answer a questionnaire, mainly on at what occasions they sent messages, and with what they expected to communicate. It was well accepted as an alternative communication media for a community. Communication using FaintPop was frequent than e-mail and phone calls: An average of 13.4 messages per subject per day. Moreover, it was reported that subjects would like to use a media like FaintPop with close, intimate friends, while they wouldn’t with non-close persons or bosses. Although the objective of the tool and our scenario were not instructed, subjects understood the nature and objective of the tool through the experiment. FaintPop was used as a very light-weight media for an acknowledge-only mode of communication. Subjects sent single message mainly to express casual greetings and simple replies (acknowledgment) to the message received. Broadcast messages were used mainly to express friendly greetings when their status change (e.g., “see you tomorrow”). Figure 2.3 shows the daily usage of FaintPop. In 10:00 period, the largest number of broadcast messages were sent: Subjects issued friendly greetings, saying good morning to the community. Subjects actively dragged the picture of friends in 15:00 period: They were between tasks, and their moods changed (or they were trying to change). In 17:00 period, subjects actively sent single messages, trying to change his/her mood by expressing casual greetings. The questionnaire indicated that “Around 17:00, I felt sympathy with friends that they also were taking a pause between tasks, because many friends actively used the tool.” The general overview of the communication activities was accepted positively. Memories of communications, both the pie charts of recent activities and the traces of communications, were accepted positively. However detailed opinions varied. For example, some subjects reported that too many traces lasted too long, others reported that traces disappeared too quickly. Therefore, there is room to refine the representation. It can retain the social relationships among the members of the community. The questionnaire replies indicated a slight improvement in the sense of closeness among the subjects, but it was not evident. However one subject reported: “I often used FaintPop when I heard a sound from it. I felt the sence of connectedness through the sounds, then confirmed the situation by watching the screen. Now the experiment is over and I miss the sounds.”

2. FaintPop: In Touch with the Social Relationships dragging

single message

17

broadcast message

50 45 40

# of activities

35 30 25 20 15 10 5 0 8

9

10

11

12

13

14

15

16

17

18

19

20

hour

Fig. 2.3. Total number of activities during the oneweek experiment

2.4 Conclusion and Related Works A tool called FaintPop is introduced, to demonstrate one answer to the problem of communication overflow. It is a media designed for a community, with which the sense of connectedness can be shared among members. It is intended to be an alternative media that is suitable for very light-weight, acknowledge-only mode of communication. Furthermore, it provides, through memories of communication, a general overview of the communication activities in the community. Results from initial experiment using the media are reported. It is suggested that the basic objectives of the tool are achieved. More long-term experiment will reveal the details on whether its objectives are accomplished and how people accept and use (or refuse) the media. For communities and groups, several visualization tools for communications have been studied [2.5, 2.10, 2.11]. However, they are sometimes oriented towards the logical, analytic aspect of the activities, or the communication media and the visualization are separated. On the other hand, FaintPop is intended to be a communication media that also offers intuitive visualization of (memories of) communication. There are several researches that try to support communities [2.8, 2.9, 2.13]. These studies are closely related to ours, however, most focus on the contents-oriented mode of communication. Several studies that use devices modeled after a photo frame are found. Kodak and StoryBox Network (www.storybox.com) started a service named Smart Picture Frame. While the sense of connectedness seems to be in its view, it merely shares pictures, not being a communication media nor using touches. In [2.12] the concept of digital picture frame is introduced. It tries to provide the visualization of everyday life activities of the person in the picture by using icons on the frame. It intends to foster relationships between distributed families, and so is closely related to Familyware [2.6]. Though the objective is close to ours, their main concern is sensing and visualizing the status of a member. A light-weight communication media that uses photo frames and feathers is proposed [2.17]. inTouch is also a light-weight media using touches [2.2]. However, these works are basically for one-to-one communication, and memories of communication are not well supported.

18

T. Ohguro et al.

Acknowledgments. The authors would like to express thanks to Yoshikazu Inui (Boctok), Takeshi Kawashima and Ryoji Murata (Kawashima-Lab.) for their collaboration in implementing the tool. The authors would like to thank Yuta Tsuboi (NAIST) for his cooperation in conducting the experiment, and Kunihiko Fujita (NTT.com) for coordinating the experiment. The authors wish to thank the colleagues who happily participated in the experiment.

References 2.1 Baron, R.A., Byrne, D. (1996): Social Psychology –8th ed. Allyn and Bacon 2.2 Brave, S., Ishii, H., Dahley, A. (1998): Tangible Interfaces for Remote Collaboration and Communication. Proc. CSCW’98, 169–178, ACM 2.3 Carroll, J.M. (2000): Making Use: Scenario-Based Design Human-Computer Interactions. MIT Press 2.4 Delgermaa, M., Ohguro, T. (2000): Intuitional Presentation for E-mail Usage Statistics —Toward the Improvement of the Awareness for Communication Activities. Proc. 60th Natl. Conv. IPSJ, 4, 167–168, IPS Japan (Japanese) 2.5 Donath, J.S. (1995): Visual Who: Animating the Affinities and Activities of an Electronic Community. Proc. Multimedia’95, 99–107, ACM 2.6 Go, K., Carroll, J., Imamiya, A. (2000): Familyware: Communicating with Someone You Love. Home Informatics and Telematics: Information, Technology and Society, 125–140, Kluwer Academic Publishers 2.7 Greenberg, S., Kuzuoka, H. (2000): Using Digital but Physical Surrogates to Mediate Awareness, Communication and Privacy in Media Spaces. Personal Technologies, 4 (1), Elsevier 2.8 Hattori, F., et al. (1999): Socialware: Multiagent Systems for Supporting Network Communities. Commun. ACM, 42 (3), 55–61 2.9 Ishida, T. (ed.) (1998): Community Computing —Collaboration over Global Information Networks. John Wiley & Sons 2.10 Kamei, K., et al. (2001): Community Organizer: Supporting the Formation of Network Communities through Spatial Representation. Proc. SAINT 2001, 207–214, IEEE 2.11 Matsubara, S., Ohguro, T. (1999): CommunityBoard 2: Mediating between Speakers and an Audience in Computer Network Discussions. Proc. Agents’99, 370–371, ACM 2.12 Mynatt, E.D., et al. (2001): Digital Family Portraits: Supporting Peace of Mind for Extended Family Members. Proc. CHI’01, 333–340, ACM 2.13 Nishida, T., et al. (1999): Public Opinion Channel for Communities in the Information Age. New Generation Computing, 17 (4), 417–427 2.14 Ohguro, T. (2000): Enhancing the Awareness for Connectedness via Intuitional Presentation of Communication. Proc. KES 2000, 341–344, IEEE 2.15 Ohguro, T. (2001): Towards Agents which are Suggestive of “Awareness of Connectedness.” Trans. IEICE, E84-D (8), (to appear), IEICE 2.16 Ohguro, T., Yoshida, S., Kuwabara, K. (1999): Gleams of People: Monitoring the Presence of People with Multi-agent Architecture. LNAI 1733 (Proc. PRIMA’99), 170–182, Springer-Verlag 2.17 Strong, R., Gaver, B. (1996): Feather, Scent, and Shaker: Supporting Simple Intimacy. Proc. CSCW’96 (short papers), 444, ACM 2.18 Sunaga, T., Nagai, Y. (2000): Information Design —Giving Comprehensive Being to Information. IPSJ Magazine, 41 (11), 1258–1263 (Japanese)

3 . F r o m

V ir tu a l E n v ir o n m e n t to V ir tu a l C o m m u n ity

A. N i j h ol t Un i ver s i ty of Twen te, D ep a r tm en t of Com p u ter Sci en ce, P O B ox 217, 7500 AE E n s ch ed e, N eth er l a n d s

3 .1

I n tr o d u c tio n

We d i s cu s s a vi r tu a l r ea l i ty th ea ter en vi r on m en t a n d i ts tr a n s i ti on to a vi r tu a l com m u n i ty b y a d d i n g d om a i n a g en ts a n d b y a l l owi n g m u l ti p l e u s er s to vi s i t th i s en vi r on m en t. Th e en vi r on m en t h a s b een b u i l t u s i n g VR ML (Vi r tu a l R ea l i ty Mod el i n g L a n g u a g e). We d i s cu s s h ow ou r i d ea s a b ou t th i s en vi r on m en t ch a n g ed i n ti m e b y a d d i n g m or e fa ci l i ti es to i t a n d b y p a y i n g m or e a tten ti on to p oten ti a l u s er s . R a th er th a n a g oa l -d i r ected i n for m a ti on a n d tr a n s a cti on s y s tem , th e en vi r on m en t i s evol vi n g i n to a vi r tu a l com m u n i ty wh er e d i ffer en ces b etween vi s i tor s a n d a r ti fi ci a l a g en ts ca n b ecom e b l u r r ed . B efor e g oi n g i n to a d es cr i p ti on of ou r own en vi r on m en t a n d i ts d evel op m en t we s u r vey th e r es ea r ch a r ea s th a t n ow a l l ow th e b u i l d i n g of 3 D em b od i ed a n d a n i m a ted a g en ts th a t s h ow i n tel l i g en ce a n d p er s on a l i ty a n d th a t ca n i n h a b i t ou r en vi r on m en t.

3 .2

T o w a r d s M u lti-u s e r V ir tu a l W o r ld s

Th e fi r s t n etwor k ed vi r tu a l wor l d s wer e tex t-b a s ed . Th ey b eca m e k n own a s MUD s (Mu l ti -Us er D om a i n s ) a n d th ey a l l owed com m u n i ca ti on b etween u s er s a n d a cces s to a s h a r ed d a ta b a s e wi th tex t d es cr i p ti on s of u s er s a n d ob j ects . In th es e en vi r on m en ts th e p er s on a l i ty of a u s er s h ows i n th e con ten ts a n d th e s ty l e of th e tex t u tter a n ces th e u s er p r od u ces , h i s tu r n ta k i n g b eh a vi or a n d m or e g en er a l l y th e m ood s (a s th ey s h ow) a n d a tti tu d es towa r d s th e com m u n i ty th a t ca n d evel op i n s u ch en vi r on m en ts . G r a p h i ca l m u l ti -u s er en vi r on m en ts wer e i n tr od u ced i n th e 198 0s . In a ty p i ca l s etti n g we h a ve a b a ck g r ou n d i m a g e s h owi n g th e en tr a n ces to s ever a l l oca ti on s or r oom s i n th e en vi r on m en t or we a r e i n on e of th es e 2D l oca ti on s a n d we ca n ch oos e on e of th e oth er vi s i tor s (or a l l of th em ) to ta l k to. Ty p i ca l l y , vi s i tor s ca n p r es en t th em s el ves b y ch oos i n g a n a va ta r (a 2D ob j ect) a n d i ts p r ed efi n ed a n i m a ti on s . Th es e a n i m a ti on s a r e s i m p l e (a wa vi n g g es tu r e, a j u m p of j oy , . . .). Mos t i n ter a cti on s a r e tex t-b a s ed , b y u s i n g ch a t wi n d ows a n d tex t b a l l oon s th a t a p p ea r a b ove th e h ea d of a va ta r s th a t ta k e p a r t i n th e d i s cu s s i on . Wi th th e a d ven t of VR ML , vi r tu a l wor l d s cou l d b e d es i g n ed for Wor l d wi d e Web . R a th er th a n for ch a tti n g , th e wor l d s wer e m ea n t to b e ex p l or ed , to ex p l a i n or to a l l ow th e s i m u l a ti on of a p a r ti cu l a r a cti vi ty i n wh i ch th e vi s i tor wa s i n vol ved . Vi r tu a l r ea l i ty a p p l i ca ti on s wer e a l r ea d y th er e a n d r a th er th a n con s i d er d i s tr i b u ted vi r tu a l r ea l i ty a s a tech n ol og y to d es i g n com m u n i ti es i t wa s ex p l or ed for a l l k i n d s T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 19-26, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

20

A. N i j h ol t

of a p p l i ca ti on s . Vi r tu a l wor l d s i n ten d ed to m eet oth er p eop l e en ter ed th e a r en a . In th es e wor l d s m u l ti p l e vi s i tor s ca n s h a r e th e s cen es . In th e m or e a d va n ced wor l d s u s er s ca n ch a n g e p a r ts of th e wor l d a n d ca n h a ve s op h i s ti ca ted vi s u a l r ep r es en ta ti on s th a t ca n i n ter a ct wi th ou t b ei n g r es tr i cted to p r ed efi n ed g es tu r es . An a va ta r ca n b e m a d e to r es em b l e th e h u m a n u s er b y p h otog r a p h i c m ea n s . Th e wor l d s th a t we con s i d er m a y h a ve col l i s i on a n d g r a vi ty fea tu r es th a t b ecom e vi s i b l e i n th e m ovem en ts of a va ta r s . Th er e ca n b e r ea l -ti m e voi ce com m u n i ca ti on a n d i n a d d i ti on th er e ca n b e l i p -s y n c fa ci a l g es tu r es . D es p i te a d d i n g s u ch fea tu r es , th er e r em a i n s a n en or m ou s g a p wh en we com p a r e th e ca p a b i l i ti es of th e a va ta r s a n d ta l k i n g h ea d s wi th th os e of th e h u m a n s th ey r ep r es en t. On e wa y to cl os e th i s g a p i s to g i ve th e h u m a n u s er th e a b i l i ty to con tr ol th e a va ta r i n a m u ch m or e d eta i l ed wa y . On e p os s i b i l i ty i s to h a ve th em ex p l i ci tl y con tr ol l ed on l i n e b y th e u s er a n d ca p tu r ed fr om ver b a l a n d n on -ver b a l i n p u t or fr om b od y m ovem en ts . Al s o, i n a d d i ti on to th e a va ta r s th a t r ep r es en t h u m a n s we ca n a d d d om a i n a va ta r s to th e en vi r on m en t to i n cr ea s e th e s en s e of r ea l i ty . Th ey s h ou l d b e a n i m a ted , b u t p r efer a b l y th er e s h ou l d b e p os s i b i l i ti es to g i ve th em p er s on a l i ty a n d ca p a b i l i ti es to a ct on th ei r own or on b eh a l f of a u s er of th e a va ta r or own er of th e en vi r on m en t. Th a t i s , th ey n eed a p p r op r i a te i n ter n a l m od el i n g to a l l ow a u ton om ou s b eh a vi or . 3 .2 .1

I n te r a c tin g E m b o d ie d P e r s o n a litie s

Ag en t tech n ol og y i s a r es ea r ch fi el d th a t em er g ed i n th e 1990’ s a n d th a t ca n b e con s i d er ed a s a fi el d i n wh i ch a ctor s a r e d evel op ed , a l th ou g h n ot n eces s a r i l y i n th e con tex t of h u m a n -com p u ter i n ter a cti on or vi r tu a l com m u n i ti es . Wi th ou t g oi n g i n to d eta i l s a n d es p eci a l l y con tr over s i a l d eta i l s , we wa n t to m en ti on p r op er ti es of s oftwa r e m od u l es th a t a r e g en er a l l y a s s u m ed to b e p r es en t b efor e b ei n g ‘ a l l owed ’ to ta l k a b ou t th em a s a g en ts : a u ton om y , r ea cti ve a n d p r oa cti ve b eh a vi or a n d th e a b i l i ty to i n ter a ct wi th oth er a g en ts (or h u m a n s ). F or a n a g en t to a ct a p p r op r i a tel y i n a d om a i n i t h a s b een u s efu l to d i s ti n g u i s h b el i efs (wh a t th e a g en t r eg a r d s to b e tr u e, th i s m a y ch a n g e i n ti m e), d es i r es (th e g oa l s th e a g en t h a s com m i tted h i m s el f to) a n d th e i n ten ti on s (s h or t-ter m p l a n s th a t i t tr i es to ex ecu te). B el i eva b i l i ty i s a n oti on th a t h a s b een em p h a s i z ed b y Jos ep h B a tes , a g a i n i n th e ea r l y 1990’ s . An a g en t i s ca l l ed b el i eva b l e, i f s om e ver s i on of a p er s on a l i ty s h ows i n th e i n ter a cti on wi th a h u m a n . Two m a i n th eor i es on p er s on a l i ty wh i ch ca n b e u s ed to d es i g n b el i eva b l e a g en ts a r e tr a i t th eor y , wh er e p er s on a l i ty i s a s et of p s y ch ol og i ca l tr a i ts th a t ch a r a cter i z es a p er s on ’ s b eh a vi or a n d s oci a l l ea r n i n g th eor y , wh er e a p p r a i s a l of th e s i tu a ti on a n d th e i n d i vi d u a l ’ s h i s tor y a r e ta k en i n to a ccou n t. Ma i n r eq u i r em en ts for b el i eva b i l i ty a r e (L oy a l l [ 8 ] ): p er s on a l i ty , em oti on , s el fm oti va ti on , ch a n g e, s oci a l r el a ti on s h i p s a n d con s i s ten cy of ex p r es s i on . Wh en we z oom i n on th e r ol e of em oti on s , i t s h ou l d b e m en ti on ed th a t th er e a r e m a n y s u b tl eti es i n vol ved wh en con vey i n g th em . Ca r toon ch a r a cter s a r e a l l owed to ex a g g er a te, g i vi n g m or e cu es to th e ob s er ver . E m oti on a l cu es s h ou l d n ’ t b e i n con fl i ct wi th con tex tu a l cu es . E m oti on a l cu es s h ou l d b e con s i s ten t d u r i n g i n ter a cti on ; n ever th el es s th ey m a y ch a n g e wh en i n ter a cti on h a s ta k en p l a ce wi th th e s a m e u s er d u r i n g a l on g er p er i od , i n ti m e. Com p u ta ti on a l m od el s fr om wh i ch em oti on a l b eh a vi or ca n b e g en er a ted ex i s t, b u t a r e n ot b a s ed on wel l -d evel op ed th eor y . Th er e-

3 . F r om

Vi r tu a l E n vi r on m en t to Vi r tu a l Com m u n i ty

21

for e, r a th er th a n h a vi n g em er g en t em oti on a l b eh a vi or b a s ed on a n a g en t’ s cog n i ti ve a p p r a i s a l m od el , we s ee a p p l i ca ti on s i n p r ototy p e (l ea r n i n g ) en vi r on m en ts wi th p r ep r og r a m m ed em oti on a l d i s p l a y . N ow th a t we h a ve d i s cu s s ed r ea s on a b l e, s oci a l , i n tel l i g en t, b el i eva b l e a n d , i n d eed , wh a tever k i n d of cog n i ti ve b eh a vi or , i t i s ti m e to con s i d er th e r ol e of em b od i m en t. E m b od i m en t a l l ows m or e a g en t m u l ti m od a l i ty , th er efor e m a k i n g i n ter a cti on m or e n a tu r a l a n d r ob u s t. Sever a l a u th or s h a ve i n ves ti g a ted n on ver b a l b eh a vi or a m on g h u m a n s a n d th e r ol e a n d u s e of n on ver b a l b eh a vi or to s u p p or t h u m a n -com p u ter i n ter a cti on . See e.g . (Ca s s el l [ 1] ) for a col l ecti on of ch a p ter s on p r op er ti es a n d i m p a ct of em b od i ed con ver s a ti on a l a g en ts (wi th a n em p h a s i s on coh er en t fa ci a l ex p r es s i on s , g es tu r es , i n ton a ti on , p os tu r e a n d g a z e i n com m u n i ca ti on ) a n d for th e r ol e of em b od i m en t (a n d s m a l l ta l k ) on fos ter i n g s el f-d i s cl os u r e a n d tr u s t b u i l d i n g . Wh i l e th e p r evi ou s i n ves ti g a ti on s we m en ti on ed ca n b e u n d er s tood to em p h a s i z e th e cog n i ti ve vi ewp oi n t of em b od i m en t, we ca n a s m u ch em p h a s i z e th e p os s i b i l i ty of a n em b od i ed a g en t to wa l k a r ou n d , to p oi n t a t ob j ects i n a vi s u a l i z ed d om a i n , to m a n i p u l a te ob j ects or to ch a n g e a vi s u a l i z ed (vi r tu a l ) en vi r on m en t. In th es e ca s es th e em b od i m en t ca n p r ovi d e a p oi n t of th e focu s for i n ter a cti on . F r om a tech n i ca l p oi n t of vi ew, ex tr em el y m u ch h a s to b e d on e on h u m a n l i k e (fr om a p h y s i ca l a n d cog n i ti ve p oi n t of vi ew) a g en t b eh a vi or . F r om a d om a i n p oi n t of vi ew i t h a s to b e d eci d ed wh en a n d wh y s u ch b eh a vi or i s u s efu l . Ou r n ex t s tep i s fr om em b od i m en t to vi r tu a l h u m a n s . A l i s t of r es ea r ch top i cs i n vol ved i n cl u d es n a tu r a l l ook i n g m ovem en t a n d d efor m a ti on of vi s i b l e b od y s u r fa ce, a n i m a ti on of s k el eton , h a n d s a n d fa ce, h a i r , s k i n a n d cl oth es r ep r es en ta ti on , n a tu r a l l ook i n g wa l k i n g a n d g r a s p i n g a n i m a ti on a n d , ver y i m p or ta n tl y i n th e vi ew of th e p r evi ou s top i cs , b eh a vi or a l a n i m a ti on wh i ch s tr i ves a t g i vi n g ch a r a cter a n d p er s on a l i ty to th e a n i m a ti on . Th i s l i s t of vi ewp oi n ts ca n b e com p l em en ted wi th vi ewp oi n ts fr om cog n i ti ve a n d p er cep tor y s ci en ces . Vi r tu a l h u m a n s h a ve to a ct i n vi r tu a l en vi r on m en ts wh er e a vi s u a l , a n a u d i tor y a n d a h a p ti c/k i n a es th eti c en vi r on m en t i n ter s ect. 3 .2 .2

E m b o d ie d P e r s o n a litie s in V ir tu a l W

o r ld s

Ag en ts a r e fi n d i n g th ei r wa y i n vi r tu a l en vi r on m en ts . Th e fi r s t a p p l i ca ti on s of em b od i ed a g en ts ca n b e fou n d i n tr a i n i n g , s i m u l a ti on , ed u ca ti on a n d en ter ta i n m en t. Th es e en vi r on m en ts m a y i n cl u d e a s i n g l e a g en t wi th wh i ch th e u s er ca n i n ter a ct, b u t th e u s er i ts el f, or p a r t of th e u s er , ca n b e i n cl u d ed i n th e en vi r on m en t. In tea m tr a i n i n g we ca n h a ve s ever a l a g en ts i n th e en vi r on m en t or s ever a l u s er s a r e r ep r es en ted i n th e en vi r on m en t. R es ea r ch i n to cr owd m od el i n g a l s o s tu d i es th e b eh a vi or of g r ou p s of p eop l e i n vi r tu a l en vi r on m en ts . However , a p a r t fr om th es e a p p l i ca ti on s we a l s o s ee d evel op m en ts wh er e 2D a n d 3 D ex ten s i on s of ch a t wor l d s a n d d i g i ta l ci ti es b ecom e i n h a b i ted b y em b od i ed a g en ts , b oth a s r ep r es en ta ti on s of vi s i tor s a n d a s a u ton om ou s d om a i n a g en ts . In th e n ea r fu tu r e we ca n ex p ect th a t com p a n i es , fa m i l i es or g r ou p s of p eop l e th a t s h a r e i n ter es ts h a ve th e op p or tu n i ty to d es i g n a n d u s e s u ch en vi r on m en ts . B el ow we m en ti on a few p r oj ects i n wh i ch th es e fu tu r e d evel op m en ts b ecom e vi s i b l e.

22

A. N i j h ol t

Sever a l i m p r es s i ve r es ea r ch s y s tem s em p l oy i n g a n i m a ted p ed a g og i ca l a g en ts h a ve b een b u i l t a n d a r e i n a p r oces s of fu r th er d evel op m en t. E m b od i ed p ed a g og i ca l a g en ts ca n s h ow h ow to m a n i p u l a te ob j ects , th ey ca n d em on s tr a te ta s k s a n d th ey ca n em p l oy g es tu r e to focu s a tten ti on . As s u ch th ey ca n g i ve m or e cu s tom i z ed a d vi ce i n a n i n for m a ti on -r i ch en vi r on m en t. L es ter et a l . [ 7] u s e th e ter m d ei cti c b el i eva b i l i ty for a g en ts th a t a r e s i tu a ted i n a wor l d th a t th ey co-i n h a b i t wi th s tu d en ts a n d i n wh i ch th ey u s e th ei r k n owl ed g e of th e wor l d , th ei r r el a ti ve l oca ti on a n d th ei r p r evi ou s a cti on s to cr ea te n a tu r a l d ei cti c g es tu r es , m oti on s , a n d u tter a n ces . On e ex a m p l e of a n en vi r on m en t th a t em p l oy s em b od i ed a g en ts i s th e Soa r Tr a i n i n g E x p er t for Vi r tu a l E n vi r on m en ts (STE VE , s ee Joh n s on et a l . [ 5] ). Th i s i s a n i m m er s i ve 3 -D l ea r n i n g en vi r on m en t wi th a vi r tu a l a g en t ca l l ed Steve. Steve d em on s tr a tes h ow to p er for m a p h y s i ca l , p r oced u r a l ta s k . It i s a ty p i ca l ex a m p l e of a n en vi r on m en t wh er e a s tu d en t ca n g et h a n d s -on ex p er i en ce. D u e to th e s tu d en t’ s h ea d -m ou n ted d i s p l a y , Steve' s p er cep ti on m od u l e k n ows a b ou t th e s tu d en t’ s p os i ti on i n th e vi r tu a l wor l d , a b ou t th e s tu d en t’ s l i n e of s i g h t a n d wh i ch ob j ects a r e i n th e s tu d en t' s fi el d of vi ew. Steve h a s b een d es i g n ed to s u p p or t tea m tr a i n i n g . A s econ d ex a m p l e we wa n t to m en ti on i s a B od y Ch a t (Vi l h j a l m s s on [ 14 ] ), a r es ea r ch en vi r on m en t on con ver s a ti on a l em b od i ed a g en ts . Th a t i s , th er e i s n ot r ea l l y a ta s k to b e p er for m ed or l ea r n ed . P eop l e ex ch a n g e i n for m a ti on a n d ch a t. In th i s en vi r on m en t s ever a l u s er s ca n h a ve a con ver s a ti on u s i n g th e k ey b oa r d wh i l e th ei r ca r toon -l i k e 3 D a n i m a ted a va ta r s d i s p l a y cor r es p on d i n g s a l u ta ti on s a n d tu r n ta k i n g b eh a vi or . Th ey l ook a wa y d u r i n g p l a n n i n g a n u tter a n ce, th ey b a ck -ch a n n el feed b a ck a n d fa ci a l ex p r es s i on a n d l ook to th e n ex t s p ea k er wh en en d i n g . Wa ta n a b e [ 15] r ep or ts a b ou t s i m i l a r r es ea r ch . An oth er s y s tem b y Vi l h j a l m s s on , ca l l ed Si tu a ted Ch a t i s i n d evel op m en t. Th i s s y s tem a l s o a n i m a tes a va ta r s i n a n on l i n e g r a p h i ca l ch a t en vi r on m en t. However , s i n ce i t k n ows a b ou t th e s h a r ed vi s u a l en vi r on m en t th e g en er a ti on of a va ta r m ovem en ts ca n i n cl u d e r efer r i n g g es tu r es wh en m a k i n g i m p l i ci t or ex p l i ci t r efer en ces to th e en vi r on m en t d u r i n g th e con ver s a ti on . As a th i r d r a n g e of ex a m p l es we l ook a t s y s tem s th a t h a ve b ecom e k n own a s i n ter a cti ve th ea ter , wh er e p l a y er s con n ected b y a n etwor k ca n ta k e p a r t i n a p er for m a n ce a s a ctor s . Th er e i s a h os t s er ver for th e p r od u cer a n d th er e a r e cl i en t com p u ter s for th e p er for m er s . Th e l a tter a r e r ep r es en ted a s a va ta r s i n th e vi r tu a l en vi r on m en t a n d wi th m oti on ca p tu r e s y s tem s (ca m er a s or s en s or s ) a va ta r m ovem en ts r efl ect p l a y er a cti on s . G es tu r es , tou ch a n d fa ci a l ex p r es s i on s of th e p l a y er s ca n b e tr a ck ed a n d g i ven to th e a n i m a ti on a l g or i th m s . Th e vi r tu a l s ta g e m a y h a ve a ctor s th a t a r e p r ovi d ed b y th e th ea ter a n d th a t s h ow a u ton om ou s b eh a vi or a ccor d i n g to s om e a cti on p a tter n s . Th ey h a ve a r ol e, b u t th e wa y th ey p er for m th i s r ol e i s a l s o d eter m i n ed i n i n ter a cti on s wi th th e h u m a n p l a y er s a n d th ei r a l ter eg o a va ta r s . See Ta k a h a s h i et a l . [ 12] a n d Tos a et a l . [ 13 ] for ex a m p l es of i n ter a cti ve th ea ter .

3 . F r om

3 .3

Vi r tu a l E n vi r on m en t to Vi r tu a l Com m u n i ty

23

B u ild in g a T h e a te r E n v ir o n m e n t

Th e m a i n th ea ter b u i l d i n g i n ou r u n i ver s i ty town i s ca l l ed ‘ Het Mu z i ek Cen tr u m ’ . It i n cl u d es th e u s u a l r oom s : p er for m a n ce h a l l s , d r es s i n g r oom s for a r ti s ts , r ecr ea ti on a l l oca ti on s (for th e a u d i en ce a n d p er for m er s ), wa r d r ob es , etceter a . It a l s o i n cl u d es a m u s i c a ca d em y . Th er e a r e a l s o s om e oth er th ea ter b u i l d i n g s i n th e town . At th i s m om en t s om e of th e b u i l d i n g s , th ei r s u r r ou n d i n g s a n d th e s tr eets l ea d i n g fr om on e l oca ti on to th e oth er a r e b ei n g m od el ed i n VR ML a n d Ja va 3 D . Th e vi r tu a l th ea ter wa s b u i l t a ccor d i n g to th e d es i g n d r a wi n g s of th e a r ch i tects of th e r ea l b u i l d i n g . Or i g i n a l l y th e en vi r on m en t wa s b u i l t a r ou n d a n a l r ea d y ex i s ti n g n a tu r a l l a n g u a g e d i a l og u e s y s tem th a t p r ovi d es i n for m a ti on a b ou t th ea ter p er for m a n ces a n d th a t a l l ows r es er va ti on s to b e m a d e. In th e vi r tu a l en vi r on m en t th e d i a l og u e s y s tem h a s b een a s s i g n ed to a vi s u a l i z ed em b od i ed a g en t. On ce we h a d th i s a g en t a n d ex ten d ed th e en vi r on m en t, th er e g r ew th e n eed to a d d oth er a g en ts th a t wer e a b l e to h el p th e vi s i tor . Th i s r a i s ed ou r i n ter es t i n h a vi n g th es e a g en ts com m u n i ca te wi th ea ch oth er a s wel l a n d to en d ow th em wi th s om e for m of a u ton om ou s b eh a vi or . R a th er th a n towa r d s a g oa l -d i r ected i n for m a ti on a n d tr a n s a cti on s y s tem com p a r a b l e to a voi ce-on l y tel ep h on e i n for m a ti on s y s tem , th e en vi r on m en t i s n ow evol vi n g i n to a vi r tu a l com m u n i ty wh er e d i ffer en ces b etween vi s i tor s a n d a r ti fi ci a l a g en ts b ecom e b l u r r ed a n d wh er e r es ea r ch top i cs s h ow a wi d e va r i ety i n cl u d i n g a s s i g n i n g p er s on a l i ti es a n d em oti on s to a r ti fi ci a l a g en ts , u s a b i l i ty s tu d i es i n vol vi n g a n a vi g a ti on a l a s s i s ta n t, for m a l s p eci fi ca ti on of (i n ter a cti on s i n ) vi r tu a l en vi r on m en ts a n d r ei n for cem en t l ea r n i n g for a g en ts i n th i s m u l ti m od a l en vi r on m en t to i n cr ea s e th ei r a u ton om y . Wh en we en ter ou r Vi r tu a l Mu z i ek Cen tr u m , we s ee th e i n for m a ti on a g en t ca l l ed Ka r i n , wa i ti n g to tel l u s a b ou t p er for m a n ces , a r ti s ts a n d a va i l a b l e ti ck ets . Vi s i tor s ca n ex p l or e th i s vi r tu a l en vi r on m en t, wa l k i n g fr om on e l oca ti on to a n oth er , l ook i n g a t p os ter s , cl i ck i n g on ob j ects a n d s o on . Ka r i n ca n b e a s k ed n a tu r a l l a n g u a g e q u es ti on s a b ou t p er for m a n ces i n th e th ea ter . Sh e h a s a cces s to a d a ta b a s e con ta i n i n g a l l th e p er for m a n ces i n th e va r i ou s th ea ter s d u r i n g th e cu r r en t s ea s on . Ka r i n h a s a 3 -D fa ce th a t a l l ows s i m p l e fa ci a l ex p r es s i on s a n d s i m p l e l i p m ovem en ts th a t a r e s y n ch r on i z ed wi th a tex t-to-s p eech s y s tem m ou th i n g th e s y s tem ’ s u tter a n ces to th e u s er (s ee N i j h ol t & Hu l s ti j n [ 9] for d eta i l s ). Oth er a g en ts h a ve b een i n tr od u ced i n th i s en vi r on m en t. F or ex a m p l e, a n a vi g a ti on a g en t, th a t k n ows a b ou t th e g eog r a p h y of th e b u i l d i n g a n d th a t ca n b e a d d r es s ed u s i n g ty p ed i n n a tu r a l l a n g u a g e u tter a n ces . Th e vi s i tor ca n a s k th e a g en t a b ou t ex i s ti n g l oca ti on s i n th e th ea ter . Wh en th e r eq u es t i s u n d er s tood , a r ou te i s com p u ted a n d th e vi ewp oi n t i n th e wor l d i s g u i d ed a l on g th i s r ou te to th e d es ti n a ti on . Th e n a vi g a ti on a g en t h a s n ot b een vi s u a l i z ed a s a 3 D em b od i ed a g en t. A Ja va b a s ed a g en t fr a m ewor k h a s b een i n tr od u ced to p r ovi d e th e p r otocol for com m u n i ca ti on b etween a g en ts a n d th e i n tr od u cti on of oth er a g en ts . F or ex a m p l e, wh y n ot a l l ow th e vi s i tor to ta l k to th e m a p of th e s ea ts i n th e m a i n con cer t h a l l or to a p os ter d i s p l a y i n g a n i n ter es ti n g p er for m a n ce? In fa ct, we ca n h a ve a m u l ti tu d e of p oten ti a l a n d u s efu l a g en ts i n ou r en vi r on m en t, wh er e s om e j u s t tr i g g er a n a n i m a ti on , oth er s ca n wa l k a r ou n d a n d oth er s h a ve b u i l t-i n i n tel l i g en ce th a t a l l ows th em to ex ecu te cer ta i n a cti on s b a s ed on i n ter a cti on s wi th vi s i tor s . Som e of th e 3 D a va ta r s th a t l i ve i n ou r en vi r on m en t h a ve n ot y et b een i n cor p or a ted i n th e fr a m e-

24

A. N i j h ol t

wor k i n a wa y th a t vi s i tor s ca n com m u n i ca te wi th th em (a b a r oq u e d a n cer , a p i a n o p l a y er ). We h a ve b een ex p er i m en ti n g wi th em b ed d i n g ou r en vi r on m en t i n a m u l ti u s er s h el l (R ei tm a y r et a l . [ 11] ) th a t a l l ows to en ter ta i n m u l ti p l e vi s i tor s th a t ca n m a k e th em s el ves vi s i b l e to ea ch oth er a s a va ta r s (VR ML ob j ects ). Th es e a va ta r s m ove a l on g wi th th e vi s i tor , b u t th ey ca n a l s o b e a s s i g n ed a n i m a ti on s , i n tel l i g en ce a n d i n ter a cti on a b i l i ti es . Hen ce, we ca n h a ve d i ffer en t h u m a n -l i k e a g en ts . Som e of th em a r e a u ton om ou s em b od i ed a g en ts s ta n d i n g or m ovi n g a r ou n d i n th e vi r tu a l wor l d a n d a l l owi n g i n ter a cti on wi th vi s i tor s of th e en vi r on m en t. Oth er s r ep r es en t h u m a n vi s i tor s of th e en vi r on m en t. We wa n t a n y vi s i tor to b e a b l e to com m u n i ca te wi th a u ton om ou s a g en ts a n d vi s i tor s , wh eth er vi s u a l i z ed or n ot. Th a t m ea n s we ca n h a ve i n ter a cti on s b etween a g en ts , b etween vi s i tor s , a n d b etween vi s i tor s a n d a g en ts . Th i s i s a r a th er a m b i ti ou s g oa l wh i ch ca n n ot b e r ea l i z ed y et com p l etel y .

3 .4

I n te r a c tin g a b o u t P e r fo r m a n c e s a n d E n v ir o n m e n t

How d oes i n ter a cti on b etween d om a i n a g en ts a n d vi s i tor s ta k e p l a ce? We d eci d ed to i n tr od u ce a m od el of n a tu r a l l a n g u a g e i n ter a cti on b etween Ka r i n a n d u s er th a t i s r a th er p r i m i ti ve fr om a l i n g u i s ti c p oi n t of vi ew, b u t s u ffi ci en tl y i n tel l i g en t fr om a p r a cti ca l a n d p r a g m a ti c p oi n t of vi ew. Th i s n a tu r a l l a n g u a g e u n d er s ta n d i n g s y s tem m ed i a tes b etween th e u s er a n d a d a ta b a s e con ta i n i n g i n for m a ti on a b ou t p er for m a n ces , a r ti s ts a n d p r i ces . Al th ou g h th e ‘ l i n g u i s ti c i n tel l i g en ce’ i s r a th er p oor , th e ou tcom e of a l i n g u i s ti c a n a l y s i s ca n b e p a s s ed on to p r a g m a ti c m od u l es th a t p r od u ce r el eva n t s y s tem r es p on s es i n th e m a j or i ty of ca s es . Th e s y s tem p r om p ts m a k e u s er s a d a p t th ei r b eh a vi or to th e s y s tem . Ka r i n p r es en ts h er i n for m a ti on u s i n g tex tto-s p eech s y n th es i s a n d l i p m ovem en ts . Wh en th er e a r e too m a n y p er for m a n ces to r ea d ou t, s h e p r es en ts a ta b l e a n d d r a ws th e u s er ’ s a tten ti on to th i s ta b l e u s i n g ey e m ovem en t a n d a n a tu r a l l a n g u a g e u tter a n ce. Th e d i a l og u e s y s tem ca n i n ter p r et a n d g en er a te r efer en ces to i tem s i n th i s ta b l e. It m a y b e cl ea r h ow to a d d r es s Ka r i n . However , vi s i tor s m a y wa n t to a d d r es s oth er d om a i n a g en ts a n d a g en ts th a t r ep r es en t u s er s . As m en ti on ed , th i s i s wor k i n p r og r es s . We a r e fol l owi n g s ever a l a p p r oa ch es to s ol ve th i s p r ob l em . Th ey a r e r el a ted a n d ca n b e i n teg r a ted s i n ce a l l of th em a r e a g en t-or i en ted a n d b a s ed on a com m on fr a m ewor k of com m u n i ca ti n g a g en ts . In a d d i ti on , we h a ve b u i l t th i s fr a m ewor k i n s u ch a wa y th a t d i ffer en t a g en ts wi th d i ffer en t a b i l i ti es ca n b ecom e p a r t of i t: a s i m p l e a n i m a ted p i a n o p l a y er , a b a r oq u e d a n cer th a t ‘ u n d er s ta n d s ’ th e m u s i c s h e i s d a n ci n g on , Ka r i n wh o k n ows a b ou t th ea ter p er for m a n ces , a n d a n a vi g a ti on a g en t th a t k n ows a b ou t th e g eog r a p h y of th e b u i l d i n g . D evel op i n g n a vi g a ti on a g en ts l ea d s to a n u m b er of q u es ti on s . How ca n we b u i l d n a vi g a ti on i n tel l i g en ce i n to a n a g en t? Wh a t d oes n a vi g a ti on i n tel l i g en ce m ea n ? How ca n we con n ect th i s i n tel l i g en ce to l a n g u a g e a n d vi s i on i n tel l i g en ce? Vi s i tor s of ou r en vi r on m en t a r e l a n g u a g e u s er s a n d , m or eover , th ey k n ow a n d i n ter p r et wh a t th ey s ee. Th er e i s a con ti n u ou s i n ter a cti on b etween ver b a l a n d n on ver b a l i n for m a ti on wh en i n ter p r eti n g a s i tu a ti on i n ou r vi r tu a l en vi r on m en t. Th i s i n ter a cti on a n d th e r ep r es en ta ti on a n d i n ter p r eta ti on of s ou r ces a n d th en th e g en er a ti on of m u l ti m ed i a fr om th em a r e a m on g th e m a i n top i cs of ou r r es ea r ch .

3 . F r om

Vi r tu a l E n vi r on m en t to Vi r tu a l Com m u n i ty

25

We ver y m u ch fol l ow D a r k en & Si l b er t [ 3 ] i n ou r a p p r oa ch to n a vi g a ti on . To a s s i s t th e vi s i tor i n n a vi g a ti n g th r ou g h ou r vi r tu a l th ea ter , we h a ve a d d ed b oth a m a p a n d a n i n tel l i g en t n a vi g a ti on a g en t. Th e vi s i tor ca n a s k q u es ti on s , g i ve com m a n d s a n d p r ovi d e i n for m a ti on wh en p r om p ted b y th e a g en t. Th i s i s d on e b y ty p i n g n a tu r a l l a n g u a g e u tter a n ces or b y m ovi n g th e m ou s e p oi n ter over th e m a p to l oca ti on s a n d ob j ects th e u s er i s i n ter es ted i n . On th e m a p th e u s er ca n fi n d th e p er for m a n ce h a l l s , th e l ou n g es a n d b a r s , s el l i n g p oi n ts , i n for m a ti on d es k s a n d oth er i n ter es ti n g l oca ti on s a n d ob j ects . Th e cu r r en t p os i ti on of th e vi s i tor i n th e vi r tu a l en vi r on m en t i s m a r k ed on th e m a p . Wh i l e m ovi n g i n VR th e vi s i tor ca n ch eck h i s or h er p os i ti on on th i s m a p . Wh en u s i n g th e m ou s e to p oi n t a t a p os i ti on on th e m a p , r efer en ces ca n b e m a d e b y b oth u s er (i n n a tu r a l l a n g u a g e) a n d s y s tem to th e ob j ect or l oca ti on p oi n ted a t. We h a ve a n n ota ted a s m a l l cor p u s of ex a m p l e u tter a n ces th a t a p p ea r i n n a vi g a ti on d i a l og u es . An ex a m p l e of a q u es ti on i s : “ Wh a t i s th i s ? ” wh i l e p oi n ti n g a t a n ob j ect on th e m a p , or “ Is th er e a n en tr a n ce for wh eel ch a i r s ? ” . E x a m p l es of com m a n d s a r e “ B r i n g m e th er e.” or “ B r i n g m e to th e i n for m a ti on d es k .” E x a m p l es of s h or t p h r a s es a r e “ N o, th a t on e.” or “ Ka r i n .” F r om th e a n n ota ted cor p u s a g r a m m a r wa s i n d u ced a n d ou r u n i fi ca ti on -ty p e p a r s er for D u tch ca n b e u s ed to p a r s e th es e u tter a n ces i n to fea tu r e s tr u ctu r es . Th r ee a g en ts com m u n i ca te to fi l l i n m i s s i n g i n for m a ti on i n th e fea tu r e s tr u ctu r e a n d to d eter m i n e th e a cti on th a t h a s to b e u n d er ta k en (a n s wer i n g th e q u es ti on , p r om p ti n g for cl a r i fi ca ti on or m i s s i n g i n for m a ti on , d i s p l a y i n g a r ou te on th e m a p or g u i d i n g th e u s er i n VR to a cer ta i n p os i ti on ). Th e n a vi g a ti on a g en t, th e d i a l og u e m a n a g er a n d th e Cos m o Ag en t d o th i s i n co-op er a ti on . N ot y et i m p l em en ted i s th e p os s i b i l i ty th a t n ot on l y th e p os i ti on b u t a l s o wh a t i s i n th e ey es i g h t of th e vi s i tor i s k n own . Th i s wi l l a l l ow i n ter p r eta ti on of r efer en ces to ob j ects th a t a r e vi s i b l e to a vi s i tor .

3 .5

T o w a r d s a T h e a te r C o m m u n ity

Th e l en g th of th i s p a p er d oes n ot a l l ow a com p r eh en s i ve s u r vey of a l l th e p r ob l em s we h a ve to d ea l wi th wh en we wa n t a n a g en t-or i en ted d es i g n of ou r en vi r on m en t a n d h a ve i t i n h a b i ted b y a g en ts th a t ca n b e em b od i ed , h a ve i n tel l i g en ce a n d p er s on a l i ty a n d ca n com m u n i ca te wi th ea ch oth er a n d wi th a g en ts th a t r ep r es en t vi s i tor s . To d es i g n a n d m a i n ta i n a n en vi r on m en t l i k e th a t we n eed s om e u n i for m i ty fr om wh i ch we ca n d i ver g e i n s ever a l d i r ecti on s : a g en t i n tel l i g en ce, a g en t i n ter a cti on ca p a b i l i ti es , a g en t vi s u a l i z a ti on a n d a g en t a n i m a ti on (cf. N i j h ol t & Hon d or p [ 10] ). Sta n d a r d s a r e n eed ed to a l l ow fr a m ewor k s for com m u n i ca ti on , i n ter n a l m od el l i n g , a n d a n i m a ti on of em b od i ed a g en ts . Th es e s ta n d a r d s s h ou l d a l s o a d d r es s i s s u es con cer n ed wi th m u l ti -u s er a n d m u l ti -d evel op er en vi r on m en ts . In E g g es et a l . [ 4 ] we i n tr od u ce a n a p p r oa ch to th e i n ter n a l m od el l i n g of a g en ts we th i n k we ca n u s e i n ou r m u l ti -a g en t a n d m u l ti -u s er en vi r on m en t. Ou r a p p r oa ch d i s cu s s ed th er e, i s l i m i ted , b u t n ever th el es s a l l ows m od el i n g of ‘ i n tel l i g en ce’ i n ter m s of b el i efs , d es i r es a n d p l a n s , a n d p os s i b l e ex ten s i on s to th e m od el i n g of em oti on s a n d a n a g en t’ s k n owl ed g e a b ou t m ovem en ts , p os tu r es a n d n on -ver b a l com m u n i ca ti on . Ou r cu r r en t em oti on r es ea r ch i s r ep or ted i n Kes ter en et a l . [ 6] a n d B u i et a l . [ 2] .

26

A. N i j h ol t

R e fe r e n c e s 3 .1 3 .2 3 .3 3 .4

3 .5

3 .6

3 .7 3 .8 3 .9

3 .10

3 .11 3 .12 3 .13 3 .14 3 .15

J. Ca s s el l , J. Su l l i va n , S. P r evos t & E . Ch u r ch i l l (ed s .). E m b o d i e d C o n v e r s a t i o n a l A g e n t s . MIT P r es s , Ca m b r i d g e, 2000. B u i Th e D u y , D . Hey l en , M. P oel & A. N i j h ol t. G en er a ti on of fa ci a l ex p r es s i on fr om em oti on u s i n g a fu z z y r u l e b a s ed s y s tem . Su b m i tted for p u b l i ca ti on , Ju l y 2001. R .P . D a r k en & J.L . Si l b er t. Wa y fi n d i n g s tr a teg i es a n d b eh a vi or s i n vi r tu a l wor l d s . C H I ’ 9 6 , 14 2-14 9. A. E g g es , A. N i j h ol t & R . op d en Ak k er . D i a l og s wi th B D P Ag en ts i n Vi r tu a l E n vi r on m en ts . In : P r oceed i n g s 2n d IJCAI wor k s h op on K n o w l e d g e a n d R e a s o n i n g i n P r a c t i c a l D i a l o g u e S y s t e m s . Sea ttl e, Wa s h i n g ton , Au g u s t 2001, to a p p ea r W. L . Joh n s on , J. W. R i ck el , a n d J. C. L es ter . An i m a ted P ed a g og i ca l Ag en ts : F a ceto-F a ce In ter a cti on i n In ter a cti ve L ea r n i n g E n vi r on m en ts . T h e I n t e r n a t i o n a l J o u r n a l o f A r t i f i c i a l I n t e l l i g e n c e i n E d u c a t i o n (2000) 11, 4 7-78 . A.-J. va n Kes ter en , R . op d en Ak k er , M. P oel & A. N i j h ol t. Si m u l a ti on of em oti on s of a g en ts i n vi r tu a l en vi r on m en ts u s i n g n eu r a l n etwor k s . In : L e a r n i n g t o B e h a v e : I n t e r n a l i s i n g K n o w l e d g e . P r oceed i n g s Twen te Wor k s h op s on L a n g u a g e Tech n ol og y 18 (TWL T 18 ), j oi n t Cel e-Twen te wor k s h op , N ovem b er 2000, 13 7-14 7. J.C. L es ter , S. G . Town s , C.B . Ca l l a wa y , J.L . Voer m a n & P .J. F i tz g er a l d . D ei cti c a n d em oti ve com m u n i ca ti on i n a n i m a ted p ed a g og i ca l a g en ts . In : Ca s s el l et a l ., 2000. A.B . L oy a l l . B e l i e v a b l e A g e n t s : B u i l d i n g i n t e r a c t i v e P e r s o n a l i t i e s . P H.D . Th es i s , CMU-CS-97-123 , Ma y 1997, Ca r n eg i e Mel l on Un i ver s i ty . A. N i j h ol t & J. Hu l s ti j n . Mu l ti m od a l In ter a cti on s wi th Ag en ts i n Vi r tu a l Wor l d s . In : F u tu r e D ir e c tio n s fo r In te llig e n t In fo r m a tio n S y s te m s a n d In fo r m a tio n S c ie n c e , N . Ka s a b ov (ed .), P h y s i ca -Ver l a g : Stu d i es i n F u z z i n es s a n d Soft Com p u ti n g , 2000. A. N i j h ol t & H. Hon d or p . Towa r d s com m u n i ca ti n g a g en ts a n d a va ta r s i n vi r tu a l wor l d s . In : P r oc. E U R O G R A P H I C S 2 0 0 0 , A. d e Sou s a & J.C. Tor r es (ed s .), Au g u s t 2000, In ter l a k en , 91-95. G . R ei tm a y r et a l . D eep Ma tr i x : An op en tech n ol og y b a s ed vi r tu a l en vi r on m en t s y s tem . T h e V i s u a l C o m p u t e r J o u r n a l 15: 3 95-4 12, 1999. K. Ta k a h a s h i , J. Ku r u m i s a wa & T. Yots u k u r a . N etwor k ed th ea ter . P r oc. F i r s t IE E E P a c i f i c - R i m C o n f . O n M u l t i m e d i a , D ecem b er 2000, Un i ver s i ty of Sy d n ey , Au s tr a l i a . N . Tos a & R . N a k a ts u . E m oti on r ecog n i ti on -b a s ed i n ter a cti ve th ea ter – R om eo & Ju l i et i n Ha d es . P r oc. E U R O G R A P H I C S ’ 9 9 , M. Al b er ti et a l . (ed s .), 179-18 2. H.H. Vi l h j a l m s s on . Au ton om ou s com m u n i ca ti ve b eh a vi or s i n a va ta r s . Ma s ter ’ s Th es i s , MIT Med i a L a b or a tor y , 1997. T. Wa ta n a b e, M. Ok u b o & Y. Is h i i . An em b od i ed vi r tu a l fa ce-to-fa ce com m u n i ca ti on s y s tem wi th vi r tu a l a ctor a n d vi r tu a l wa ve for h u m a n i n ter a cti on s h a r i n g . In : P r oc. 4 th Wor l d Mu l ti con fer en ce on S y s t e m i c s , C y b e r n e t i c s a n d I n f o r m a t i c s ( S C I ' 0 0 ) , Vol . III, Or l a n d o, USA, 14 6-151, 2000.

4 . C o lla b o r a tiv e I n n o v a tio n T o o ls Joh n C. Th om a s IB M T. J. Wa ts on R es ea r ch P O B ox 704 , Yor k town Hei g h ts N ew Yor k 10598 USA j cth om a s @ u s .i b m .com

4 .1 I m p o r t a n c e o f C o lla b o r a t io n : P r a c t ic a l a n d S c ie n t if ic We l i ve i n a n i n cr ea s i n g l y i n ter con n ected wor l d . In r efl ecti on of th i s tr en d , th e fi el d of h u m a n -com p u ter i n ter a cti on h a s s h i fted focu s fr om i n d i vi d u a l s to tea m s a n d l a r g e or g a n i z a ti on s [ 3 5] . F r om a s ci en ti fi c p er s p ecti ve, we l ea r n m os t a b ou t th e ob j ect of s tu d y d u r i n g tr a n s i ti on s . Th u s , a l ea r n i n g tes t i s g en er a l l y m or e d i a g n os ti c of b r a i n fu n cti on th a n a tes t of s tor ed k n owl ed g e; a g l u cos e tol er a n ce tes t tel l s u s m or e th a n a r es ti n g b l ood s u g a r l evel ; a s tr es s tes t r evea l s m or e a b ou t th e h ea r t th a n d oes r es ti n g h ea r t r a te. Si m i l a r l y , th i s cen tu r y ’ s r a p i d tr a n s i ti on s s h ou l d a l l ow u s to l ea r n a g r ea t d ea l a b ou t col l ecti ve h u m a n b eh a vi or . At th e s a m e ti m e, we fa ce en or m ou s p l a n eta r y p r ob l em s i n cl u d i n g g l ob a l fou l i n g of th e ecos p h er e, i n eq u i ty i n econ om i c op p or tu n i ty , i n cr ea s ed ch a n ces for ca ta s tr op h i c d i s ea s e, a n d i n ter n a ti on a l ter r or i s m . Th es e p r ob l em s a r os e wi th cu r r en t a p p r oa ch es a n d l i m i ta ti on s to col l a b or a ti on a n d wi l l on l y b e s ol ved vi a b r ea k th r ou g h s i n col l a b or a ti on . F r om a m or e m u n d a n e vi ewp oi n t, s i m i l a r ch a l l en g es ex i s t tod a y for l a r g e, i n ter n a ti on a l or g a n i z a ti on s . F or i n s ta n ce, th e wor l d i s ch a n g i n g m or e q u i ck l y b u t cr ea ti ve d es i g n a b i l i ty h a s n ot i n cr ea s ed . As a r es u l t, th er e i s a wi d en i n g g a p b etween th e d eg r ee of fl ex i b i l i ty a n d cr ea ti vi ty n eed ed to a d a p t a n d th e ca p a ci ty of i n d i vi d u a l s a n d or g a n i z a ti on s to d o s o [ 12] . D es i g n p r ob l em s a r e often ex tr em el y h i g h l ever a g e for or g a n i z a ti on s . F or i n s ta n ce, er r or s i n d es i g n , wh eth er i n s oftwa r e, d r u g s , b u s i n es s p r oces s es , or a u tom ob i l es a r e ex tr em el y cos tl y . Con ver s el y , effecti ve a n d i n n ova ti ve d es i g n s ca n b e ex tr em el y l u cr a ti ve; a r e a h a l l m a r k s of l on g -l i ved com p a n i es [ 7, 10] . E ven a m od es t i n cr ea s e i n th e a b i l i ty of or g a n i z a ti on s to cr ea te m or e effecti ve d es i g n s cou l d g r ea tl y i n cr ea s e p r ofi ts i n ex i s ti n g m a r k ets a n d cr ea te wh ol e n ew m a r k ets . In cr ea s i n g d es i g n effecti ven es s wi l l r eq u i r e col l a b or a ti on b r ea k th r ou g h s . Hu m a n b ei n g s evol ved n a tu r a l l a n g u a g e a s a m eth od of col l a b or a ti on a m on g s m a l l g r ou p s of p eop l e wh o g en er a l l y s h a r ed con tex t, g oa l s , ex p er i en ce a n d cu l tu r e. Un d er th os e ci r cu m s ta n ces , s eq u en ti a l h u m a n s p eech s er ved fa i r l y wel l , e.g ., th e tel l i n g of s tor i es for s h a r i n g ex p er i en ces [ 3 4 ] . However , u n a i d ed s p eech i s n ot wel l -s u i ted to l a r g e-s ca l e col l a b or a ti on s ; p a r ti cu l a r l y n ot wh en th e p eop l e i n vol ved h a ve va s tl y d i ffer en t a s s u m p ti on s , cu l tu r a l b a ck g r ou n d s , g oa l s , con tex ts , ex p er i en ces a n d n a ti ve l a n g u a g es . We h a ve n ot y et i n ven ted a n en ti r el y effecti ve T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 27-3 4 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

28

J.C. Th om a s

r ep l a cem en t of n a tu r a l l a n g u a g e for l a r g e, d i ver s e g r ou p s th ou g h s tor y tel l i n g ca n b e u s efu l i n b r i d g i n g g a p s a m on g g r ou p s wh en i n cor p or a ted i n to th e a p p r op r i a te p r oces s [ 3 , 4 , 3 7] . Ca n we fu r th er ex ten d s u ch tech n i q u es to fa ci l i ta te com m u n i ca ti on a m on g l a r g er , m or e d i ver s e g r ou p s ? Or , s h ou l d we l i m i t s u ch i n ter a cti on s to " d r y " i n ter a cti on s [ 2] ? On e of th e s p eci a l ch a l l en g es offer ed b y col l a b or a ti on tod a y i s th a t often i t i n vol ves r em ote p a r ti ci p a n ts ; s om eti m es , wor l d wi d e[ 25] . In m a n y con ver s a ti on s a n d p a p er s , a n i m p l i ci t a s s u m p ti on i s th a t r em ote col l a b or a ti on i s l i m i ted b y b a n d wi d th a l on e a n d th a t th e cu r r en t s u p er i or i ty of fa ce to fa ce over r em ote col l a b or a ti on wi l l d i s a p p ea r on ce b a n d wi d th b ecom es l a r g e en ou g h . Su ch a n a n a l y s i s over l ook s two a d d i ti on a l a n d p oten ti a l l y q u i te i m p or ta n t a s p ects of fa ce to fa ce col l a b or a ti on . F i r s t, fa ce to fa ce col l a b or a ti on a l l ows p eop l e to s ee a n d ex p er i en ce th e p h y s i ca l a n d s oci a l con tex t of th ei r col l a b or a tor s . P er h a p s th ey s ee th e b u i l d i n g wh er e oth er s wor k ; tr y th e s a m e food ; fi n d ou t wh eth er th ey wor k i n a q u i et or n oi s y en vi r on m en t; wh a t th e m ood s a r e of th os e th a t p a s s b y i n th e h a l l wa y s . Secon d , s h a r i n g a n a ctu a l p h y s i ca l s p a ce a l l ows th e p os s i b i l i ty of m u ch d eep er i n ter a cti on a n d th a t p os s i b i l i ty m a y wel l a ffect tr u s t even i f th e p os s i b i l i ty n ever m a ter i a l i z es . Con s i d er two r a th er ex tr em e ex a m p l es . F i r s t, two p eop l e s h a r i n g a p h y s i ca l s p a ce m a y b e s u b j ect to a n a tu r a l d i s a s ter s u ch a s a n ea r th q u a k e a n d on e m a y s a ve th e l i fe of th e oth er . Al th ou g h ob vi ou s l y a ver y l ow p r ob a b i l i ty even t, th e m er e p os s i b i l i ty m a y wel l p u t p eop l e’ s p er cep tu a l a n d em oti on a l a p p a r a tu s i n to a h ei g h ten ed s ta te of a r ou s a l . Secon d , i f two p eop l e s h a r e a com m on p h y s i ca l s p a ce, on e cou l d p h y s i ca l l y i n j u r e th e oth er . Si n ce A’ s tr u s t of B i s en h a n ced b y s i tu a ti on s wh er ei n A cou l d h u r t B b u t i n fa ct, d oes n ot, th e ty p i ca l fa ce to fa ce i n ter a cti on m a y en h a n ce tr u s t i n j u s t th i s wa y . It i s n ot on l y th e m ed i u m a n d con tex t of com m u n i ca ti on th a t i m p a ct col l a b or a ti on , b u t a l s o th e con ten t. In p a r ti cu l a r , we a r g u e th a t ex p r es s i ve com m u n i ca ti on m a y offer a n op p or tu n i ty for col l a b or a tor s to g a i n m or e com p r eh en s i ve m od el s of ea ch oth er th a n i n s tr u m en ta l com m u n i ca ti on a l on e. In s tr u m en ta l com m u n i ca ti on i s com m u n i ca ti on th a t i s r eq u i r ed to a ccom p l i s h th e cu r r en t ta s k . E x p r es s i ve com m u n i ca ti on i s com m u n i ca ti on th a t tel l s a b ou t th e com m u n i ca tor a s wel l a s th e s u b j ect; i t i s com m u n i ca ted m or e b eca u s e th e com m u n i ca tor wa n ts to th a n b eca u s e th ey n eed to. Z h en g , B os , Ol s on , a n d Ol s on [ 3 8 ] s h owed th a t col l a b or a ti on a n d tr u s t ca n b e, i n effect, " j u m p -s ta r ted " wi th s oci a l ch i tch a t. Stor i es ca n a l s o h el p p eop l e d evel op m or e tr u s t th a n th e ex ch a n g e of i n for m a ti on p er s e. A s tor y i s n ot s i m p l y a n ob j ecti ve r ecou n ti n g of even ts ; i t a l wa y s i m p l i es a n u m b er of r evea l i n g ch oi ces . Th e s tor y tel l er ch oos es wh i ch even ts to ta l k a b ou t; wh er e to s ta r t; ton e; vi ewp oi n t; wh i ch d eta i l s to d es cr i b e a n d s o on . Th r ou g h s u ch ch oi ces , th e s tor y tel l er i n evi ta b l y r evea l s th em s el ves a s wel l a s th e s u b j ect. So l on g a s col l a b or a ti on p r oceed s a l on g p r ed i cta b l e l i n es , m od el s b u i l t fr om ex p r es s i ve com m u n i ca ti on m a y b e u n n eces s a r y . B u t, i f s ta n d a r d p r oced u r es b r ea k d own , th en col l a b or a tor s wh o h a ve d evel op ed m or e com p l ex m od el s of ea ch oth er wi l l b e a b l e to r ea ct m or e effecti vel y a n d effi ci en tl y a s a tea m . Of cou r s e, th er e i s a l s o a d a n g er h er e. As p er h a p s h i n ted a t b y Az ech i [ 2] , s tor i es m i g h t a l s o r evea l

4 . Col l a b or a ti ve In n ova ti on Tool s

29

ch a r a cter i s ti cs of th e s tor y tel l er th a t oth er col l a b or a tor s m i g h t fi n d q u i te n eg a ti ve wh i l e p u r el y i n s tr u m en ta l com m u n i ca ti on s a r e u n l i k el y to d o s o. A ch a l l en g e for k n owl ed g e s oci a l i z a ti on i s to d eter m i n e th e con d i ti on s u n d er wh i ch i t i s b etter to k eep com m u n i ca ti on s “ d r y ” or “ i n s tr u m en ta l ” a n d wh en i t i s d es i r a b l e to i n cl u d e m or e ex p r es s i ve or “ wet” m od es of com m u n i ca ti on . If th e l a tter i s n eces s a r y , we a l s o n eed to d evel op m eth od s of p r og r es s i ve d i s cl os u r e th a t m i n i m i z e fr i cti on a n d m a x i m i z e em p a th y .

4 .2 N e w T e c h n o lo g ic a l P o s s ib ilit ie s R ecen t a d va n ces i n com p u ti n g p ower , i n ter fa ce tech n ol og i es , b a n d wi d th , s tor a g e, a n d s oci a l en g i n eer i n g p r ovi d e m a n y p os s i b i l i ti es for n ovel s ol u ti on s to l a r g e-s ca l e col l a b or a ti on m a y b e d es i g n ed , tes ted , a n d i m p r oved . In th e " r ea l wor l d " effecti ve on -l i n e col l a b or a ti on s y s tem s b oth a t a d i s ta n ce [ 16] a n d fa ce-to-fa ce [ 17] , a r e a l r ea d y b ei n g fa ci l i ta ted b y tech n ol og y . We b el i eve fu r th er a d va n ces ca n b e m a d e b y i n cor p or a ti n g cr ea ti vi ty a i d s , s u g g es ti on s for p r oces s es [ 3 3 ] , a n d b y p r ovi d i n g tool s for a l ter n a ti ve r ep r es en ta ti on s [ 3 1] . F a i l u r e to i n n ova te i s n ot r a n d om , b u t ca n b e a s cr i b ed to on e of s ever a l m a i n d i ffi cu l ti es : 1. In d i vi d u a l s or g r ou p s d o n ot en g a g e i n effecti ve a n d effi ci en t p r oces s es of i n n ova ti ve d es i g n . 2. Th e n eces s a r y s k i l l s , ta l en ts , a n d k n owl ed g e s ou r ces a r e n ot b r ou g h t to b ea r on th e p r ob l em . 3 . Ap p r op r i a te r ep r es en ta ti on s of th e s i tu a ti on a r e n ot u s ed . L a b or a tor y [ 6, 15, 29] a s wel l a s fi el d r es ea r ch [ 24 , 3 6] h a s es ta b l i s h ed th a t th e m a j or p r oces s d i ffi cu l ti es a r e m a i n l y d u e to a l i m i ted n u m b er of p r even ta b l e er r or s . An a p p r op r i a te over a l l s tr u ctu r e m a y fa ci l i ta te g r ou p s th r ou g h s tep s of i n n ova ti on a n d h el p g u i d e th es e s ep a r a te s tep s ; d i s ti n ct g u i d el i n es a r e a p p r op r i a te wi th i n ea ch of th es e s tep s [ 28 , 3 3 ] . A com m on p r ob l em i s th a t p eop l e ty p i ca l l y fa i l to s p en d s u ffi ci en t ti m e i n th e ea r l y s ta g es of d es i g n ; vi z ., p r ob l em fi n d i n g a n d p r ob l em for m u l a ti on [ 27] . A com m on fa i l u r e d u r i n g a s p eci fi c s ta g e of i n n ova ti ve d es i g n i s th a t p eop l e often b r i n g cr i ti ca l j u d g m en t i n to p l a y too ea r l y i n th e i d ea g en er a ti on p h a s e of p r ob l em s ol vi n g . As a n oth er ex a m p l e, u n l i k e N ewel l a n d Si m on ' s [ 22] n or m a ti ve m od el of i d ea l p r ob l em s ol vi n g , i n fa ct, p eop l e' s b eh a vi or i s p a th -d ep en d en t a n d th ey a r e often u n wi l l i n g to ta k e wh a t a p p ea r s to b e a s tep th a t u n d oes a p r evi ou s a cti on even i f th a t s tep i s a ctu a l l y n eces s a r y for a s ol u ti on [ 29] . R eg a r d i n g th e s econ d i s s u e (b r i n g i n g to b ea r n eces s a r y s k i l l s , ta l en ts a n d k n owl ed g e s ou r ces ), wh i l e s oftwa r e tool s ca n n ot fu l l y s u b s ti tu te for h u m a n ex p er ts , evi d en ce s u g g es ts th a t i n d i vi d u a l s h a ve a l a r g e a m ou n t of r el eva n t i m p l i ci t k n owl ed g e wh i ch th ey often wi l l n ot b r i n g to b ea r on a p r ob l em a n d th a t g i vi n g a p p r op r i a te s tr a teg i es [ 29] , or k n owl ed g e s ou r ces [ 3 0] ca n h el p . R eg a r d i n g th e th i r d i s s u e of a p p r op r i a te r ep r es en ta ti on , con tr ol l ed l a b or a tor y ex p er i m en ts h a ve s h own th a t s u b j ects d i d s i g n i fi ca n tl y b etter , for ex a m p l e, i n a tem p or a l d es i g n ta s k wh en th ey u s ed a s p a ti a l r ep r es en ta ti on ; y et, ver y few s u b j ects s p on ta n eou s l y a d op ted s u ch a r ep r es en ta ti on [ 6] . Th e i m p a ct of g ood r ep r es en ta ti on s , h owever , i s n ot con fi n ed to l a b or a tor y d em on s tr a ti on s . Sp eech r es ea r ch

3 0

a d sp p h p r im

J.C. Th om a s

va n cem en ts a ccel er a ted g r ea tl y wh en wa vefor m s wer e l a r g el y r ep l a ced wi th eech s p ectr og r a m s a n d F ey n m a n d i a g r a m s a l l owed b r ea k th r ou g h s i n a tom i c y s i cs . B y p r ovi d i n g p eop l e wi th a va r i ety of p oten ti a l r ep r es en ta ti on s a n d s om e oces s es to en cou r a g e th e ex p l or a ti on of va r i ou s a l ter n a ti ves , we cou l d p r ob a b l y p r ove p er for m a n ce s i g n i fi ca n tl y . Ad va n ces i n s p eech r ecog n i ti on , com b i n ed wi th n a tu r a l l a n g u a g e p r oces s i n g a n d d a ta m i n i n g r a i s e th e p os s i b i l i ty of l a r g e-s ca l e r ea l ti m e col l a b or a ti on s . Sp eech r ecog n i ti on ca n tu r n r a w s p eech i n to tex t. Sta ti s ti ca l tech n i q u es ca n a u tom a te th e for m a ti on of " a ffi n i ty g r ou p s " th a t s h a r e va r i ou s i n ter es ts , va l u es , or g oa l s [ 23 ] . Sp eech r ecog n i ti on , i n th i s con tex t, n eed n ot p r od u ce p er fect tr a n s cr i p ts of wh a t i s s a i d b u t on l y tr a n s cr i b e en ou g h con ten t to en a b l e n a tu r a l l a n g u a g e p r oces s i n g s oftwa r e to cl u s ter s eg m en ts of tex t. Ad d i ti on a l b en efi ts s tem fr om a s p eech to tex t to cl u s ter i n g s y s tem . In th e p a s t, con ver s a ti on s wer e tr a n s i en t. Th er e wa s n o " ob j ecti ve" evi d en ce of th ei r con ten t or s tr u ctu r e. It often h a p p en s , e.g ., i n a g r ou p m eeti n g th a t th e fi r s t p er s on to r a i s e a n ew i d ea i s n ot r ecog n i z ed a s h a vi n g d on e s o. In s tea d , th e s econ d or th i r d p er s on to m en ti on th e i d ea i f often cr ed i ted wi th i t, q u i te p os s i b l y b eca u s e th e fi r s t m en ti on i s u n a s s i m i l a b l e b y th e cu r r en t m en ta l m od el of th e l i s ten er s b u t ca u s es a ch a n g e i n m en ta l m od el s s o th a t a s u b s eq u en t m en ti on i s com p r eh en s i b l e. Th e m or e g en er a l p oi n t i s th a t com p u ter i z ed r ecor d s of g r ou p m eeti n g s a n d l a r g er s ca l e col l a b or a ti on s a l l ow th e p os s i b i l i ty of feed i n g b a ck to th e p a r ti ci p a n ts va r i ou s vi s u a l i z a ti on s of b eh a vi or , m a k i n g th e com p u ter a n a cti ve p a r ti ci p a n t i n g r ou p com m u n i ca ti on [ 3 2] . In con j u n cti on wi th effecti ven es s m etr i cs , s u ch feed b a ck m ech a n i s m s m a y a l l ow g r ou p s to i m p r ove effecti ven es s . At IB M, we r ecen tl y en g a g ed i n a cor p or a te-wi d e ex p er i m en t ca l l ed " Wor l d Ja m " wh er ei n a l l IB Mer s wor l d wi d e wer e i n vi ted to a th r ee-d a y el ectr on i c m eeti n g to d i s cu s s ten i s s u es of i n ter es t to IB Mer s i n cl u d i n g em p l oy ee r eten ti on , wor k l i fe b a l a n ce, a n d wor k i n g r em otel y . Over 52, 600 em p l oy ees p a r ti ci p a ted a n d p os ted over 6000 s u g g es ti on s a n d com m en ts . E a ch top i c h a d a m od er a tor a n d fa ci l i ta tor s . E a ch m od er a tor , i n tu r n , h a d b een a s k ed to a s s em b l e a top i c-k n owl ed g ea b l e " B oa r d of Ad vi s or s " to p r ovi d e r efer en ces , web s i tes , a n d oth er r el eva n t m a ter i a l s a h ea d of ti m e a s wel l a s p a r ti ci p a ti on d u r i n g th e on -l i n e con fer en ce. In a d d i ti on , th e s et of m od er a tor s a n d fa ci l i ta tor s com m u n i ca ted wi th ea ch oth er th r ou g h a s y s tem ca l l ed " B a b b l e" wh i ch wa s d es i g n ed , d evel op ed , a n d d ep l oy ed a t IB M R es ea r ch . Th e B a b b l e s y s tem b l en d s s y n ch r on ou s a n d a s y n ch r on ou s tex t com m u n i ca ti on . In d i vi d u a l s i n th e s y s tem a r e r ep r es en ted a s col or ed d ots . Th e p os i ti on of a d ot wi th i n a s i m p l e vi s u a l i z a ti on ca l l ed a " s oci a l p r ox y " a l l ows ea ch p a r ti ci p a n t to q u i ck l y s ee wh o el s e i s p r es en t a n d wh i ch top i cs a r e b ei n g d i s cu s s ed . Wh en a u s er of th e s y s tem ty p es a n en tr y or s cr ol l s th r ou g h r ecor d ed d i s cu s s i on , th ei r d ot m oves to th e cen ter of th e s oci a l p r ox y for th a t top i c. Sever a l " B a b b l es " a r e n ow a cti ve wi th i n IB M i n cl u d i n g on e for " Com m u n i ty B u i l d er s " ; th a t i s , p eop l e i n va r i ou s or g a n i z a ti on s th r ou g h ou t IB M i n ter es ted i n th e p r oces s , tool s , a n d m eth od s for com m u n i ty b u i l d i n g ; " KM B l u e" wh i ch i n cl u d es a s i m i l a r cr os s -or g a n i z a ti on a l g r ou p i n ter es ted i n k n owl ed g e m a n a g em en t a n d " D es i g n er s " wh i ch b r i n g s tog eth er p eop l e wh os e p r i m a r y p r ofes s i on a l i d en ti fi ca ti on i s a s a d es i g n er . In th e ca s e or Wor l d Ja m , B a b b l e en a b l ed th e m od er a tor s a n d fa ci l i ta tor s to tr a d e b es t p r a cti ces a n d en g a g e i n j oi n t p r ob l em

4 . Col l a b or a ti ve In n ova ti on Tool s

3 1

s ol vi n g i n a ti m el y m a n n er . Ad d i ti on a l i n for m a ti on a b ou t th e fea tu r es , fu n cti on s , d es i g n r a ti on a l e for a n d em p i r i ca l s tu d i es of B a b b l e i s a va i l a b l e i n [ 13 , 14 ] . In ea r l i er wor k , we s h owed th a t th e i n tr od u cti on of p r ob l em s ol vi n g a i d s to b r ea k s et i n cr ea s ed p er for m a n ce a n d cr ea ti vi ty [ 3 0] a n d th a t i n s tr u cti on s to ta k e on m u l ti p l e vi ewp oi n ts i n cr ea s ed p r ob l em s fou n d i n h eu r i s ti c eva l u a ti on of a s oftwa r e d es i g n [ 11] . Th e u s e of m u l ti p l e vi ewp oi n ts h a s b een q u i te con s ci ou s l y u s ed b y th e Ir oq u oi s (a n d oth er cu l tu r es ) for th ou s a n d s of y ea r s [ 3 6] . Oth er wr i ter s on cr ea ti vi ty h a ve s u g g es ted s i m i l a r m eth od s [ 9, 28 ] .

4 .3 W

o r k o f th e K n o w le d g e S o c ia liz a tio n G r o u p

Th e wor k of ou r own g r ou p ob vi ou s l y r el a tes to a ti n y a r ea of th e va s t s p a ce ou tl i n ed a b ove. Ou r wor k com p r i s es s ever a l i n ter l a ced th r ea d s . In on e th r ea d , we a r e con cep tu a l i z i n g , d es i g n i n g , a n d b u i l d i n g tool s to s u p p or t th e cr ea ti on , ca p tu r e, or g a n i z a ti on , u n d er s ta n d i n g , a n d u ti l i z a ti on of s tor i es a s a m eth od for g r ou p s to b u i l d a n d s h a r e k n owl ed g e. In th e " Va l u e Mi n er " , e.g ., n a tu r a l l a n g u a g e p r oces s i n g m eth od s a r e u s ed to fi n d va l u es a s ex p r es s ed i n tex t. Th i s cou l d b e a p p l i ed to con ver s a ti on s , d ocu m en ts , a n d web -s i tes a s wel l a s s tor i es . Th e Va l u e Mi n er fi n d s va l u e-r el a ted wor d s a n d p h r a s es a n d tr i es to ca teg or i z e th es e. A r el a ted , " P oi n t Of Vi ew" tool s h ows th e va l u e s i m i l a r i ti es a n d d i ffer en ces of p a r ti ci p a n ts . We a r e a l s o wor k i n g on s tor y vi s u a l i z a ti on s a i m ed a t h el p i n g i n d i vi d u a l s a n d g r ou p s cr ea te, u n d er s ta n d , a n d fi n d s tor i es r el eva n t to a s i tu a ti on a t h a n d . F or ex a m p l e, i n on e l i n e of d evel op m en t, we a r e s h owi n g ti m el i n es of p l ot p oi n ts a n d ch a r a cter d evel op m en t. In a n oth er l i n e of r ep r es en ta ti on r es ea r ch , we s h ow a top l evel vi ew of th e k i n d s of a ttr i b u tes th a t a r e u s ed to d es cr i b e ch a r a cter s . B y cl i ck i n g on a top l evel vi ew, th e u s er m a y z oom on to th e va l u e a s s oci a ted wi th th a t a ttr i b u te a n d u l ti m a tel y to th e u n d er l y i n g tex t. In a d d i ti on to vi s u a l i z a ti on s , th er e a r e g u i d el i n es a n d m ea s u r es b a s ed on k n own h eu r i s ti cs of s tor y wr i ti n g th a t ca n b e i n cor p or a ted i n to g r ou p wa r e [ 18 , 21] . In or d er to p r ovi d e a com m on u n d er p i n n i n g for th e va r i ou s s tor y r el a ted tool s th a t we h a ve d evel op ed , we h a ve p r op os ed a fi r s t p a s s a t a " Stor y ML " ; th a t i s , a m a r k u p l a n g u a g e s p eci fi ca l l y g ea r ed towa r d s tor i es . In th i s r ep r es en ta ti on , th er e a r e th r ee d i ffer en t b u t r el a ted " vi ews " of s tor y : Stor y F or m (wh a t i s i n th e s tor y ); Stor y F u n cti on (wh a t a r e th e p u r p os es of th e s tor y ); a n d Stor y Tr a ce (wh a t i s th e h i s tor y of th e s tor y ). In tu r n , th e Stor y F or m ca n b e b r ok en d own i n to d i m en s i on s of E n vi r on m en t, Ch a r a cter , P l ot, a n d N a r r a ti ve. Th e i d ea of th e Stor y ML i s th a t i t i s ex p a n d a b l e a ccor d i n g to p u r p os e. F or s om e p u r p os es , th e u s er (e.g ., a s tu d en t s tu d y i n g m y s ter y p l ots ) m a y b e s a ti s fi ed wi th m i n i m a l d eta i l con cer n i n g F u n cti on a n d Tr a ce b u t n eed to ex p a n d cer ta i n a s p ects of th e Stor y F or m i n g r ea t d eta i l . In a n oth er con tex t, a d i ffer en t u s er (e.g ., a h i s tor i a n com p a r i n g cer ta i n th em es a cr os s ti m e a n d cu l tu r es ) m i g h t h a ve a ver y h i g h l evel vi ew of Stor y F or m a n d Stor y F u n cti on b u t wa n t to p r ovi d e a d eta i l ed d es cr i p ti on of Stor y Tr a ce. At th i s p oi n t, th e m eta -d a ta i n Stor y ML m u s t b e s u p p l i ed b y a k n owl ed g ea b l e h u m a n b ei n g . On ce a b a s e of p oten ti a l l y u s efu l s tor i es b ecom es l a r g e i n a n y on e col l ecti on or d om a i n , i t ca n b ecom e a ch a l l en g e to fi n d th e " r i g h t" s tor y or s tor i es . If on e i s

3 2

J.C. Th om a s

l ook i n g for s tor i es wi th p a r ti cu l a r ob j ects , p eop l e, or p l a ces i n th em , " k ey wor d i n con tex t" s ea r ch es a r e g en er a l l y s u ffi ci en t. B u t, i f on e i s l ook i n g for s tor i es a b ou t a cti vi ti es , a m or e s u b tl e a p p r oa ch i s r eq u i r ed . In r es p on s e to th i s ch a l l en g e, we h a ve d evel op ed a s cr i p t-b a s ed s tor y b r ows er . Th e " s cr i p t" i s a d efa u l t s et of p a r a m eter s a b ou t a n a cti vi ty ; i t m a y s p eci fy r ol es , g oa l s , ob j ects , a n d a s eq u en ce of even ts . In th e s tor y b r ows er , a u s er m a y ch oos e a n a cti vi ty a n d fi n d s tor i es r el a ted to th a t a cti vi ty or r el a ted a cti vi ti es th r ou g h a com b i n a ti on of s ea r ch i n g a n d b r ows i n g . Al th ou g h th i s a cti vi ty -b a s ed s ea r ch wor k s a t a h i g h er l evel of s em a n ti cs th a n ty p i ca l s ea r ch es , i n m a n y ca s es , a p er s on i s s ea r ch i n g for a s tor y th a t i l l u s tr a tes a p a r ti cu l a r k i n d of ver y a b s tr a ct p oi n t a n d even th e p a r ti cu l a r a cti vi ty i s n ot th a t i m p or ta n t. F or i n s ta n ce, th e s tor y of Od y s s eu s h i d i n g h i s wa r r i or s i n Th e Tr oj a n Hor s e m a y b e a p p l i ca b l e i n a wi d e va r i ety of d om a i n s s u ch a s d i s ea s e con tr ol or com p u ter s ecu r i ty . In s u ch ca s es , to fi n d s tor i es th a t a r e p oten ti a l l y a p p l i ca b l e, we r ea l l y n eed a s y s tem b a s ed on a b s tr a ct p l a n n i n g a n d p r ob l em s ol vi n g s tr a teg i es . In ou r l a b , An d r ew G or d on [ 20] h a s d evel op ed s u ch a n on tol og y for a b s tr a ct p l a n n i n g a n d p r ob l em s ol vi n g b y i n ter vi ewi n g ex p er ts a n d r ea d i n g s tr a teg y b ook s i n a wi d e va r i ety of d om a i n s a n d th en for m u l a ti n g th es e s tr a teg i es i n a b s tr a ct ter m s . In th e n ex t s tep , th es e ter m s ca n b e u s ed to ca teg or i z e s tor i es a ccor d i n g to th e s tr a teg i es th a t a r e u ti l i z ed . Th i s wi l l en a b l e i n d i vi d u a l p r ob l em s ol ver s , ed u ca tor s , a n d tea m s to fi n d s tor i es th a t a r e p oten ti a l l y a p p l i ca b l e to i m p r ovi n g s p eci fi c s i tu a ti on s or s ol vi n g p a r ti cu l a r p r ob l em s . We a r e a l s o en g a g ed i n a ttem p ti n g to ex ten d th e a r ch i tect Ch r i s top h er Al ex a n d er ’ s [ 1] con cep t of a P a tter n L a n g u a g e to s tor i es . A P a tter n L a n g u a g e con s i s ts of a l a tti ce of i n ter r el a ted p a tter n s . E a ch p a tter n h a s a Ti tl e, a d es cr i p ti on of a con tex t i n wh i ch a p r ob l em i s l i k el y to occu r , a d es cr i p ti on of op p os i n g for ces , a n d th e b a s i c ou tl i n e of a s ol u ti on . A p a tter n a l s o often con ta i n s a d i a g r a m i l l u s tr a ti n g th e b a s i c s ol u ti on , a n d m a y con ta i n r efer en ces or oth er evi d en ce a b ou t i ts effi ca cy . E a ch p a tter n a l s o i n cl u d es l i n k s to h i g h er l evel a n d l ower l evel p a tter n s . Th e n oti on s of p a tter n s a n d A P a tter n L a n g u a g e h a ve b een a p p l i ed to a va r i ety of fi el d s b es i d es a r ch i tectu r e i n cl u d i n g ob j ect-or i en ted p r og r a m m i n g [ 19] , p r oj ect s tr u ctu r e [ 8 ] a n d h u m a n -com p u ter i n ter a cti on [ 5] . Ty p i ca l l y , a P a tter n L a n g u a g e i s d evel op ed b y a com m u n i ty of p r a cti ce a s a wa y to cr ea te, or g a n i z e a n d r eu s e k n owl ed g e. Ou r a ttem p ts to p r ovi d e a d d i ti on a l k n owl ed g e s ou r ces a r e focu s ed m a i n l y on tea ch i n g s tor i es [ 3 4 ] , p a r ti cu l a r l y d u r i n g s p eci fi c s ta g es of p r ob l em s ol vi n g . F or ex a m p l e, th e s tor y " Wh o Sp ea k s for Wol f" b y P a u l a Un d er wood [ 3 6] i s a s tor y es p eci a l l y wel l -s u i ted to ei th er p r ob l em for m u l a ti on or to a l a s t m i n u te ch eck th a t a l l s ta k eh ol d er s ’ con cer n s a r e cover ed b efor e s i g n i fi ca n t r es ou r ces a r e com m i tted to a p a r ti cu l a r p l a n . In oth er ca s es , th e i n d i vi d u a l , tea m , or or g a n i z a ti on wi l l n eed to u s e a s tor y b r ows er wh os e ex p a n d i n g ca p a b i l i ti es a r e ou tl i n es a b ove. In th i s p a p er , we h a ve a ttem p ted to d o th r ee th i n g s . 1. Con vi n ce th e r ea d er th a t i m p r ovi n g a n d u n d er s ta n d i n g th e a b i l i ty of i n d i vi d u a l s , tea m s , a n d or g a n i z a ti on s to i n n ova te m or e effecti vel y i s k ey to ou r col l ecti ve s u r vi va l . 2. Ou tl i n e h ow r ecen t a d va n ces i n s ci en ce a n d tech n ol og y offer a p r om i s e to en h a n ce col l a b or a ti ve i n n ova ti on . 3 . D es cr i b e i n ou tl i n e th e s m a l l con tr i b u ti on s a l on g th es e l i n es of th e IB M R es ea r ch Kn owl ed g e Soci a l i z a ti on G r ou p .

4 . Col l a b or a ti ve In n ova ti on Tool s

3 3

R e fe r e n c e s 4 .1 4 .2 4 .3 4 .4

4 .5 4 .6 4 .7 4 .8 4 .9 4 .10 4 .11

4 .12 4 .13

4 .14

4 .15

4 .16

4 .17

4 .18 4 .19

Al ex a n d er , C., Is h i k a wa , S., Si l ver s tei n , M. Ja cob s on , M., F i k s d a h l -Ki n g , I. a n d An g el , S. A p a t t e r n l a n g u a g e . N ew Yor k : Ox for d Un i ver s i ty P r es s , 1977. Az ech i , S. Soci a l p s y ch ol og i ca l a p p r oa ch to k n owl ed g e-cr ea ti n g com m u n i ti es . In T. N i s h i d a (E d .), D y n a m i c k n o w l e d g e i n t e r a c t i o n . B oca R a ton : CR C P r es s , 2000. B ey er , H. a n d Hol tz b l a tt, K. Con tex tu a l d es i g n : d efi n i n g cu s tom er -cen ter ed s y s tem s . Sa n F r a n ci s co: Mor g a n Ka u fm a n , 1998 . B od k er , SA. Scen a r i os i n u s er -cen ter ed d es i g n : s etti n g th e s ta g e for r efl ecti on a n d a cti on . P r es en ted a t th e 3 2n d a n n u a l Ha wa i i In ter n a ti on a l Con fer en ce on Sy s tem Sci en ce, Ja n u a r y , 1999, Ma u i , Ha wa i i B or ch er s , J. A p a t t e r n s a p p r o a c h t o i n t e r a c t i o n d e s i g n . N ew Yor k : Wi l ey , 2001. Ca r r ol l , J., Th om a s , J.C. a n d Ma l h otr a , A. P r es en ta ti on a n d r ep r es en ta ti on i n d es i g n p r ob l em s ol vi n g . B r i t i s h J o u r n a l o f P s y c h o l o g y , 7 1 (1), . 14 3 -155, 198 0. Col l i n s , J. a n d P or r a s , J. B u i l t t o l a s t . N ew Yor k : Ha r p er , 1994 . Cop l i en , Ja m es . h ttp : //www1.b el l -l a b s .com /u s er /cop e/P a tter n s /P r oces s /i n d ex .h tm l , 2001. D e B on o, E . S i x t h i n k i n g h a t s . B os ton : L i ttl e, B r own , 198 5. D eG eu s , A. T h e l i v i n g c o m p a n y : h a b i t s f o r s u r v i v a l i n a t u r b u l e n t b u s i n e s s e n v i r o n m e n t . B os ton : Ha r va r d B u s i n es s Sch ool P r es s , 1997. D es u r vi r e, H. a n d Th om a s , J. E n h a n ci n g th e p er for m a n ce of i n ter fa ce eva l u a tor s u s i n g n on -em p i r i ca l u s a b i l i ty m eth od s . P r o c e e d i n g s o f t h e 3 7 t h A n n u a l H u m a n F a c t o r s S o c i e t y M e e t i n g , 113 2-113 6, Sa n ta Mon i ca , CA: Hu m a n F a ctor s Soci ety , 1993 . D r u ck er , P . Ma n a g i n g i n a ti m e of g r ea t ch a n g e. Tr u m a n Ta l l ey B ook s : N ew Yor k , 1995. E r i ck s on , T., Sm i th , D . Kel l og g , W., L a ff, M., R i ch a r d s , J. a n d B r a d n er , E . Soci a l l y tr a n s l u cen t s y s tem s : Soci a l p r ox i es , p er s i s ten t con ver s a ti on a n d th e d es i g n of " B a b b l e." In H u m a n F a c t o r s a n d C o m p u t i n g S y s t e m s : T h e p r o c e e d i n g s o f C H I ’ 9 9 . N ew Yor k : ACM P r es s , 1995. E r i ck s on , T. & Kel l og g , W. " Soci a l Tr a n s l u cen ce: An Ap p r oa ch to D es i g n i n g Sy s tem s th a t Mes h wi th Soci a l P r oces s es ." T r a n s a c t i o n s o n C o m p u t e r - H u m a n I n t e r a c t i o n , 7 (1), 59-8 3 , 2000. F a r n h a m , S. et. a l s . Str u ctu r ed on l i n e i n ter a cti on s : Im p r ovi n g th e d eci s i on -m a k i n g of s m a l l d i s cu s s i on g r ou p s . P r o c e e d i n g s o f C S C W 2 0 0 0 . 299-3 08 . N ew Yor k : ACM, 2000. F i n h ol t, T.A., a n d Ol s on , G .M., " F r om L a b or a tor i es to Col l a b or a tor i es : A N ew Or g a n i z a ti on a l F or m for Sci en ti fi c Col l a b or a ti on , " Un i ver s i ty of Mi ch i g a n , An n Ar b or , Ja n u a r y 1997. F i s ch er , G . D om a i n -Or i en ted D es i g n E n vi r on m en ts : Su p p or ti n g In d i vi d u a l a n d Soci a l Cr ea ti vi ty " , i n J. G er o a n d M.L . Ma h er (ed s ): " Com p u ta ti on a l Mod el s of Cr ea ti ve D es i g n IV" , Key Cen tr e of D es i g n Com p u ti n g a n d Cog n i ti on , Sy n d n ey , Au s tr a l i a , 8 3 111, 1999. F r ey , J. H o w t o w r i t e a d a m n e d g o o d n o v e l I I . N ew Yor k : St. Ma r ti n ’ s P r es s , 1994 . G a m m a , E ., Hel m , R ., Joh n s on , R ., a n d Vl i s s i d es , J. D e s i g n P a t t e r n s : E l e m e n t s o f R e u s a b l e O b j e c t O r i e n t e d S o f t w a r e . R ea d i n g , MA: Ad d i s on -Wes l ey , 1995.

3 4

J.C. Th om a s

4 .20 G or d on , A. Th e r ep r es en ta ti on a l r eq u i r em en ts of s tr a teg i c p l a n n i n g . F i fth s y m p os i u m on l og i ca l for m a l i z a ti on s of com m on s en s e r ea s on i n g . N ew Yor k Un i ver s i ty , Ma y 2022, 2001. 4 .21 McKee, R . S t o r y : S u b s t a n c e , s t r u c t u r e , s t y l e a n d t h e p r i n c i p l e s o f s c r e e n w r i t i n g . N ew Yor k : Ha r p er , 1997. 4 .22 N ewel l , A. a n d Si m on , H. H u m a n p r o b l e m s o l vi n g . Up p er Sa d d l e R i ver , N J: P r en ti ce-Ha l l , 1972. 4 .23 N i s h i d a , T. D y n a m i c k n o w l e d g e i n t e r a c t i o n . B oca R a ton : CR C P r es s , 2000. 4 .24 Ol s on , M. a n d B l y , S. Th e P or tl a n d ex p er i en ce: A r ep or t on a d i s tr i b u ted r es ea r ch g r ou p . I n t e r n a t i o n a l J o u r n a l o f M a n - M a c h i n e S t u d i e s , 3 4 , 211-228 , 1991. 4 .25 Ol s on , G . a n d Ol s on , J. D i s ta n ce m a tter s . H u m a n - C o m p u t e r I n t e r a c t i o n . 1 5 , 2-3 , 10713 7, 2000. 4 .26 P ol tr ock , S. a n d E n g l eb eck , G . R eq u i r em en ts for a vi r tu a l col l oca ti on en vi r on m en t, I n f o r m a t i o n a n d S o f t w a r e T e c h n o l o g y , 4 1 (6), 3 3 1-3 3 9, 1999. 4 .27 Sob el , D . L o n g i t u d e : T h e t r u e s t o r y o f a l o n e g e n i u s w h o s o l v e d t h e g r e a t e s t s c i e n t i f i c p r o b l e m o f h i s t i m e . N ew Yor k : P en g u i n , 1995. 4 .28 Stei n , M. Sti m u l a ti n g cr ea ti vi ty . N ew Yor k : Aca d em i c P r es s , 1974 . 4 .29 Th om a s , J. An a n a l y s i s of b eh a vi or i n th e h ob b i ts -or cs p r ob l em . Cog n i ti ve P s y ch ol og y , 6, 257-269, 1974 . 4 .3 0 Th om a s , J. , L y on , D ., a n d Mi l l er , L . Ai d s for p r ob l em s ol vi n g . IB M T. J. Wa ts on R es ea r ch R ep or t, R C-64 68 . N ew Yor k : IB M, 1977. 4 .3 1 Th om a s , J. a n d Ca r r ol l , J. Th e p s y ch ol og i ca l s tu d y of d es i g n . D es i g n Stu d i es , 1 (1), 511, 1979. 4 .3 2 Th om a s , J. C. Th e com p u ter a s a n a cti ve com m u n i ca ti on s m ed i u m . P r o c e e d i n g s o f t h e 1 8 t h A n n u a l M e e t i n g o f t h e A s s o c i a t i o n f o r C o m p u t a t i o n a l L i n g u i s t i c s , 8 3 -8 6, N ew Yor k : ACL , 198 0. 4 .3 3 Th om a s , J. P r ob l em s ol vi n g b y h u m a n -m a ch i n e i n ter a cti on . In G i l h ool y K.J., (E d ). H u m a n a n d m a c h i n e p r o b l e m s o l v i n g . L on d on : P l en u m P u b l i s h i n g , 198 9. 4 .3 4 Th om a s , J. N a r r a ti ve tech n ol og y a n d th e n ew m i l l en n i u m . K n o w l e d g e M a n a g e m e n t J o u r n a l , 2 (9), 14 -17, 1999. 4 .3 5 Th om a s , J. An HCI a g en d a for th e n ex t m i l l en n i u m : E m er g en t g l ob a l i n tel l i g en ce. In R . E a r n s h a w, R . G u ed y , A. va n D a m a n d J. Vi n ce (E d s .), F r o n t i e r s o f H u m a n C e n t e r e d C o m p u t i n g , O n l i n e C o m m u n i t i e s a n d V i r t u a l E n v i r o n m e n t s , 2001. 4 .3 6 Un d er wood , P . T h r e e N a t i v e A m e r i c a n l e a r n i n g s t o r i e s . G eor g etown , Tex a s : Tr i b e of Two P r es s , 1994 . 4 .3 7 Va n D er Hei j d en , K. S c e n a r i o s : T h e a r t o f s t r a t e g i c c o n v e r s a t i o n . N ew Yor k : Wi l ey , 1996. 4 .3 8 Z h en g , J., B os , N . Ol s on , J. a n d Ol s on , G . Tr u s t Wi th ou t Tou ch : Ju m p -Sta r t Tr u s t Wi th Soci a l Ch a t. P r o c e e d i n g s o f C H I 0 1 (Con fer en ce Com p a n i on ), N ew Yor k : ACM, 2001.

5 . B r ic k s &

B its &

I n te r a c tio n

R . F r u ch ter D i r ector of P r oj ect B a s ed L ea r n i n g L a b or a tor y , Ci vi l a n d E n vi r on m en ta l E n g i n eer i n g D ep a r tm en t, Sta n for d Un i ver s i ty , Sta n for d , CA 94 3 05-4 020, USA

5 .1 I n t r o d u c t io n In tod a y ’ s i n for m a ti on tech n ol og y (IT) a n d com m u n i ca ti on i n ten s i ve en vi r on m en t p eop l e, tech n ol og y a n d b u i l d en vi r on m en t d es i g n er s , a n d or g a n i z a ti on s a r e ch a l l en g ed to u n d er s ta n d th e i m p a cts on th e wor k s p a ce, con ten t th a t i s cr ea ted a n d s h a r ed , a n d s oci a l , b eh a vi or a l a n d cog n i ti ve a s p ects of wor k , p l a y , l ea r n i n g , a n d com m u n i ty . Th e s tu d y i s a t th e i n ter s ecti on of th e d es i g n of p h y s i ca l s p a ces , i .e., b r i c k s , r i ch el ectr on i c con ten t s u ch a s vi d eo, a u d i o, s k etch i n g , CAD , i .e., b i t s , a n d n ew wa y s p eop l e b eh a ve i n com m u n i ca ti ve even ts u s i n g a ffor d a n ces of IT a u g m en ted s p a ces a n d con ten t, i .e., i n t e r a c t i o n . Th e s tu d y p r op os es two h y p oth es es . B r i c k & B i t s & I n t e r a c t i o n H y p o t h e s i s : If we u n d er s ta n d i n g th e r el a ti on s h i p b etween b r i c k s , b i t s , a n d i n t e r a c t i o n we wi l l b e a b l e to 1. d es i g n s p a ces th a t b etter a ffor d com m u n i ca ti ve even ts , 2. d evel op col l a b or a ti on tech n ol og i es b a s ed on n a tu r a l i d i om s th a t b es t s u p p or t th e a cti vi ti es p eop l e p er for m , 3 . en g a g e p eop l e i n r i ch com m u n i ca ti ve ex p er i en ces th a t en a b l e th em to i m m er s e i n th ei r a cti vi ty a n d for g et a b ou t th e tech n ol og y th a t m ed i a tes th e i n ter a cti on . C h a n g e H y p o t h e s i s : An y n ew i n for m a ti on a n d col l a b or a ti on tech n ol og y wi l l r eq u i r e ch a n g e a n d r eth i n k i n g of: 1. th e d es i g n a n d l oca ti on of s p a ces i n wh i ch p eop l e wor k , l ea r n , a n d p l a y . 2. th e con ten t p eop l e cr ea te i n ter m s of r ep r es en ta ti on , m ed i a , i n ter r el a ti on a m on g th e d i ffer en t m ed i a , th e con ten t’ s evol u ti on over ti m e s o th a t i t p r ovi d es con tex t a n d s ets i t i n a s oci a l com m u n i ca ti ve p er s p ecti ve. 3 . th e i n ter a cti on s a m on g p eop l e i n ter m s of th e i n d i vi d u a l ’ s b eh a vi or , i n ter a cti on d y n a m i cs , n ew com m u n i ca ti on p r otocol s , col l a b or a ti on p r oces s es ; r el a ti on b etween p eop l e a n d a ffor d a n ces of th e s p a ce; a n d i n ter a cti vi ty wi th th e con ten t. Th e p a p er u s es s cen a r i os a n d two col l a b or a ti on tech n ol og y ex a m p l es to d i s cu s s th e B r i c k s & B i t s & I n t e r a c t i o n p er s p ecti ve a n d h i g h l i g h ts th e b eh a vi or a l a n d s oci a l ch a n g es th a t h a ve to b e a cq u i r ed a s p eop l e i n ter a ct wi th a n d i n th e con tex t of n ew com m u n i ca ti on tech n ol og i es a n d IT a u g m en ted s p a ces . Th e two i n for m a ti on a n d col l a b or a ti on tech n ol og i es a r e: 1. MS N etm eeti n g , a col l a b or a ti on tech n ol og y for vi d eocon fer en ci n g [ 1] , TM 2. R E CAL L a r es ea r ch p r ototy p e d evel op ed a t th e P B L L a b a t Sta n for d [ 2] . Th e two s cen a r i os took p l a ce i n th e con tex t of th e ed u ca ti on tes tb ed focu s ed on G l ob a l Tea m wor k i n Ar ch i tectu r e, E n g i n eer i n g , Con s tr u cti on (A/E /C) offer ed a t Sta n for d Un i ver s i ty [ 3 ] . Th e A/E /C p r og r a m en g a g es s tu d en ts fr om u n i ver s i ti es wor l d wi d e, i .e., Sta n for d Un i ver s i ty , UC B er k el ey , G eor g i a Tech , Ka n s a s Un i ver T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 5-4 2, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

3 6

R . F r u ch ter

s i ty , Ca l P ol y Sa n L u i s Ob i s p o, fr om th e US, TU D el ft fr om N eth er l a n d s , B a u h a u s Un i ver s i ty fr om G er m a n y , E TH Z u r i ch a n d F HA fr om Swi tz er l a n d , Un i ver s i ty of L j u b l j a n a fr om Sl oven i a , Aoy a m a G a k u i n Un i ver i s ty , Ja p a n , i n g l ob a l tea m wor k .

5 .2 V is ib ilit y , A w a r e n e s s , a n d I n t e r a c t io n in V id e o c o n f e r e n c e S p a c e S c e n a r i o . Sy n ch r on ou s m u l ti -m od a l col l a b or a ti on i n a vi d eocon fer en ce m ed i a ted tea m m eeti n g b etween a n a r ch i tectu r e s tu d en t a t B er k el ey a n d two s tu d en ts a t Sta n for d , a s tr u ctu r a l en g i n eer a n d a n u n d er g r a d u a te a p p r en ti ce, i s u s ed a s a s cen a r i o to d i s cu s s th e m eth od a n d fi n d i n g s of th e s tu d y . Th e s tu d y ca p tu r ed th e i n ter a cti on a m on g th e th r ee a ctor s b y vi d eo ta p i n g b oth s i tes . Ab ou t 4 0 h ou r s of i n ter a cti on s wer e r ecor d ed a n d a n a l y z ed u s i n g vi d eo p r otocol a n a l y s i s m eth od s . Two k ey a s p ects wer e s tu d i ed : th e wor k s p a ce a n d con ten t a s p ects p r es en t i n th e p r oces s , a n d th e i n ter a cti on r el a ted to th e s oci a l p r oces s a n d th e d i s cou r s e. B r i c k s . F r om th e p oi n t of vi ew of b r i c k s th e s tu d y a n a l y z es th e a ffor d a n ces a n d l i m i ta ti on s of ty p i ca l p r es et p h y s i ca l vi d eocon fer en ce wor k s p a ces , e.g ., l a b s , or cu b i cl es . Mor e th a n th a t th e l oca ti on of th e P C a n d a u d i o/vi d eo d evi ces i s fi x ed . In s u ch a vi d eocon fer en ce s etti n g on e or m or e p a r ti ci p a n ts m ove, i n ter a ct, a n d u s e th e a ffor d a n ce of th e tech n ol og y a n d th e s p a ce to com m u n i ca te wi th r em ote tea m m em b er s . Th e r es ea r ch a n d p r a g m a ti c q u es ti on i s h o w c a n a f l e x i b l e s p a c e b e d e s ig n e d to a c c o m m o d a te th e c h a n g in g n e e d s o f th e in te r a c tio n , a w a r e n e s s a n d v is ib ility o f th e d is tr ib u te d p e o p le e n g a g e d in th e c o m m u n ic a tiv e e v e n t? Th e a n a l y s i s focu s ed on th e en vi r on m en ta l a s p ects p r es en t i n th e i n ter a cti on , e.g ., th e a n a l y s i s of th e p a r ti ci p a n ts ’ m ovem en ts i n th e s p a ce. F r om th e ob s er va ti on s , on e i m p or ta n t a s p ect i n u n d er s ta n d i n g p a r t of th e b eh a vi or a l p a tter n s of th e p a r ti ci p a n ts wa s th e s tu d y of th e wor k s p a ce u s ed i n th e i n ter a cti on s . It wa s r ed u ced to th e a r ea s u r r ou n d i n g th e P C, cr ea ti n g a r es tr i cted i n ter a cti on s p a ce. Th e a ffor d a n ces of th e eq u i p m en ts u s ed a l s o d eter m i n ed th e wa y i n wh i ch th e p a r ti ci p a n ts u s ed th e s p a ce. P a r ti cu l a r l y , two a s p ects wer e r el eva n t i n th e wa y i n wh i ch p eop l e u s e th e wor k s p a ce: th e l oca ti on s of th e m on i tor a n d th e vi d eo ca m er a . B oth Mon i tor s a n d Vi d eo ca m er a s d efi n e p r efer r ed l oca ti on s for p a r ti ci p a n ts th a t n a r r ow th e p os s i b i l i ti es for u s i n g a cer ta i n a r ea i n th e wor k i n g s p a ce. Wh en a n a l y z i n g th e m ovem en ts of th e p a r ti ci p a n ts i n r el a ti on to th e l oca ti on of th e eq u i p m en t, we ca n s ee th a t th e m ovem en ts a r e r es tr i cted to a tr i a n g u l a r a r ea , th a t we h a ve ca l l ed C o n e o f I n t e r a c t i o n (C O I ). Th e F i g . 1 s h ows th e m ovem en ts of two p a r ti ci p a n ts d u r i n g a r ea l i n ter a cti on . We i d en ti fi ed fou r m a j or a r ea s i n th e COI: C o m m a n d a r e a (A) th e a r ea i n wh i ch th e p er s on th a t l ea d s th e i n ter a cti on i s l oca ted . Th e p os i ti on h a s to d o m os t l i k el y wi th th e u s e of th e i n p u t d evi ce, a n d i t i s a l l th e wa y a r ou n d i n th e ca s e of l eft-h a n d ed u s er s ; S e c o n d a r y a r e a (B ) i s th e a r ea occu p i ed b y d efa u l t b y th e oth er p er s on or p eop l e i n vol ved i n th e i n ter a cti on ; P o i n t i n g d e v i s e (p ), M i c r o p h o n e a r e a (m ). Key a s p ects h a ve to b e con s i d er ed i n r el a ti on to th e COI. On on e h a n d , th e over l a p p i n g of fu n cti on a l a r ea s cr ea ted b y th e vi d eo ca m er a a n d th e m on i tor . Th i s over l a p p i n g cr ea tes th r ee z on es a s s h own i n F i g . 1: th e s ector (1) d efi n es th e a r ea

5. B r i ck s &

B i ts &

In ter a cti on

3 7

i n wh i ch th e u s er of th e com p u ter ca n h a ve vi s i b i l i ty of th e s cr een a n d b e ca p tu r ed b y th e l en s of th e vi d eo ca m er a ; th e s ector (3 ) i s th e a r ea i n wh i ch n o vi s i b i l i ty ca n h a p p en , b oth for th e u s er of th e com p u ter a n d for th e r ecei ver of th e i m a g e ca p tu r ed b y th e vi d eo ca m er a . However , th e s ector (2) i s p oten ti a l l y m os t p r ob l em a ti c of a l l , b eca u s e wh en b ei n g i n th i s a r ea th e u s er of th e s y s tem ca n h a ve vi s i b i l i ty of th e com p u ter s cr een , b u t a t th e s a m e ti m e b e ou t of th e ca m er a r a n g e wi th ou t n oti ci n g i t, cr ea ti n g a vi s u a l con ta ct fa i l u r e i n th e com m u n i ca ti on a l p r oces s .

A 3

2

p m

B

3

2

1

F i g . 1 . Con e of In ter a cti on a n d Ar ea s of Vi s i b i l i ty , Awa r en es s , a n d In ter a cti on

Th e COI con tr i b u tes to a fa l s e s en s e of a wa r en es s of th e p a r ti ci p a n t i n th e vi d eo i n ter a cti on , b y cr ea ti n g th e wr on g b el i ef th a t b y b ei n g i n th e vi s i b i l i ty r a n g e of th e s cr een , th e a cti on s p er for m ed wi l l b e tr a n s m i tted to th e n on -col l oca ted p a r ti ci p a n t. An ex a m p l e of th i s wr on g s en s e of a wa r en es s i s a n a cti on th a t h a p p en s often wh en p a r ti ci p a n ts a r e d es cr i b i n g i n for m a ti on th a t i n vol ves p oi n ti n g wi th th ei r fi n g er a t g r a p h i ca l i n for m a ti on on th e m on i tor , coi n ed i n th e s tu d y “ F a k ed p oi n ti n g .“ “ F a k ed p oi n ti n g ” ca n b e con s i d er ed a com m u n i ca ti on fa i l u r e s i tu a ti on , b eca u s e i t ca n l ea d to m i s u n d er s ta n d i n g a n d d el a y s i n th e com m u n i ca ti on a l p r oces s . Th e u s e of b od y g es tu r es for con vey i n g th e d i s cou r s e a r e d r a s ti ca l l y con s tr a i n ed b y th e a ffor d a n ces of th e vi d eo d evi ces i n u s e, a n d th e l a ck of a wa r en es s of th i s fa ct b y th e s p ea k er ca n l ea d to i m p or ta n t l os s es i n th e com m u n i ca ti on a l p r oces s . B i t s . D u r i n g a vi d eocon fer en ce m eeti n g , p a r ti ci p a n ts m a n i p u l a te, a n d ed i t r i ch con ten t s u ch a s tex t, 3 D m od el s i n CAD , a n d s k etch es on wh i teb oa r d s th r ou g h a p p l i ca ti on s h a r i n g . Th e a d va n ta g e offer ed b y a p p l i ca ti on s h a r i n g p r ovi d es th e p a r ti ci p a n ts wi th i n ter a cti vi ty , vi s i b i l i ty of i d ea s a n d a cti on s i n th e a p p l i ca ti on , m a k i n g th ei r th ou g h t p r oces s vi s i b l e, a s wel l a s m a n i p u l a ti on of ea ch oth er ’ s con ten t a n d ex p l or e a l ter n a ti ves , i .e. “ Wh a t i f” s cen a r i os . I n t e r a c t i o n . Th e vi d eocon fer en ce tech n ol og y r eq u i r es th e p a r ti ci p a n ts to a cq u i r e n ew com m u n i ca ti on s k i l l s a n d ch a n g e th ei r i n ter a cti on h a b i ts to b en efi t fr om th e m u l ti -m od a l com m u n i ca ti on en vi r on m en t. N ew com m u n i ca ti on p r otocol em er g e, for i n s ta n ce, s i n ce vi d eocon fer en ce s etti n g s l a ck th e col l oca ted r i ch q u eu es , e.g ., p a r ti ci p a n ts s p en d l on g er ti m e i n ter va l s a t th e s ta r t of th e m eeti n g to es ta b l i s h a fr a m ewor k a n d a r a p p or t. Th e a ffor d a n ce or l i m i ta ti on of th e d es i g n ed wor k s p a ce a n d h a r d wa r e con fi g u r a ti on ca n l ea d to m i s com m u n i ca ti on s for m a l i z ed a s com m u n i ca ti on fa i l u r es g a z e, fa k e p oi n ti n g , vi s i b i l i ty of a l l p a r ti ci p a n ts a n d a wa r en es s of a cti on s ta k en i n th e s h a r ed a p p l i ca ti on s i n th e d i ffer en t s i tu a ti on s of s p a ce u s a g e b y on e, two, or m or e p a r ti ci p a n ts .

3 8

R . F r u ch ter

Th e vi d eo p r otocol a n a l y s i s of th e 4 0 h ou r s of i n ter a cti on ca p tu r ed l ea d to th e i d en ti fi ca ti on of p a tter n s of i n ter a cti on b a s ed on th e a n a l y s i s of th e ver b a l a n d n on -ver b a l d i s cou r s es . Th e s m a l l es t u n i t of com m u n i ca ti on for th i s l evel of a n a l y s i s wa s th e t u r n d efi n ed b y ea ch i n ter ven ti on p r od u ced b y a n y of th e s p ea k er s i n th e con tex t of a n i n ter a cti on . Th e tu r n s wer e g r ou p ed a n d s tr u ctu r ed i n to l a r g er u n i ts , con for m i n g th r ee d i ffer en t l evel s i n s i d e th e d i s cou r s e’ s s tr u ctu r e: 1. T o p i c s : th e top i cs cor r es p on d to th os e i d en ti fi a b l e th em es r a i s ed b y th e s p ea k er s d u r i n g th e con ver s a ti on . D i ffer en t tu r n s ca n s h a r e th e s a m e top i c. 2. E p i s o d e s : ep i s od es a r e s er i es of tu r n s th a t s h a r e s om e s p eci fi c fu n cti on a l con ten t i n th e con tex t of th e d i s cou r s e. Th es e tu r n s i n th e ep i s od e ca n b el on g to d i ffer en t k i n d s of top i cs . 3 . P r o t o c o l s : p r otocol s p oi n t ou t th e ex i s ten ce of p a tter n s i n th e com m u n i ca ti on b etween th e p a r ti ci p a n ts th a t h a p p en i n th e i n n er s tr u ctu r e of th e E p i s od es . A p r otocol i s s h a p ed b y a p a r ti cu l a r s er i es of tu r n s , wh i ch con for m s tr u ctu r es of ver b a l a n d /or b eh a vi or a l a cti on s th a t ca n b e i d en ti fi ed a s h a vi n g a p a r ti cu l a r p u r p os e i n th e con tex t of th e i n ter a cti on . In or d er to eva l u a te th i s i n n er s tr u ctu r e of th e ep i s od es , two k i n d s of a n a l y s i s wer e a p p l i ed to th e tex ts . Th e fi r s t on e wa s a tech n i q u e ca l l ed l i n k og r a p h y [ 4 ] , wh i ch s h ows g r a p h i ca l l y th e r el a ti on s h i p a m on g th e d i ffer en t top i cs p r es en t i n th e d i s cou r s e. L i n k og r a p h y i s u s efu l to i d en ti fy ch a r a cter i s ti cs of th e ver b a l i n ter a cti on a s th e a r ea s en cl os ed b y ea ch E p i s od e, th e con n ecti on s b etween th e Top i cs , a n d th e r ecu r r en ce of th em . F i g . 2 s h ows on e of th e l i n k og r a p h y g r a p h i cs p r od u ced . It i s p os s i b l e to i d en ti fy th e r ecu r r en ce of th e top i cs , r ep r es en ted b y s ever a l tr i a n g l es ; th e b i g g er th e tr i a n g l e, th e fa r th er th e a p p ea r a n ce of a top i c i s fr om th e l a s t ti m e i t a p p ea r ed i n th e i n ter a cti on . Th i s i n for m a ti on wa s m a d e m or e ex p l i ci t b y a u g m en ti n g th e l i n k og r a p h y m eth od b y col or -cod i n g th e tr i a n g l es . Us e of col or s m a d e i t p os s i b l e to i d en ti fy i m p or ta n t ch a r a cter i s ti cs of th e top i cs a s th e r ecu r r en ce of th em . Th os e i n ter a cti on s th a t a r e m or e fr eq u en t - r ep r es en ted b y s m a l l s er i es of p y r a m i d s - a r e th e top i cs th a t con s ti tu te th e cor e of th e d i s cu s s i on . Th e d i ffer en t ep i s od es con ta i n ed i n th e i n ter a cti on wer e r ep r es en ted i n a b a r g r a p h i n wh i ch th e h or i z on ta l a x i s r ep r es en ts th e tem p or a l d u r a ti on of th e ep i s od es b y u s i n g p r op or ti on a l s ca l e. Th i s i s i m p or ta n t to es ta b l i s h r a ti os b etween th e wei g h ts of th e d i ffer en t ep i s od es i n th e con tex t of th e wh ol e con ver s a ti on . Th i s a n a l y s i s en a b l es th e i d en ti fi ca ti on of th e com m u n i ca ti on p r otocol s p r es en t i n th e d i ffer en t ep i s od es (F i g . 2). E x a m p l es of i t a r e th e s tr i p s p oi n ted ou t b y th e a r r ows i n th e g r a p h i c, wh i ch r ep r es en t th e p r otocol for p r od u ci n g th e tr a n s i ti on b etween ep i s od es . Th e s tu d y th en l i n k ed th e tem p or a l g r a p h i c r ep r es en ta ti on of th e ep i s od es wi th th e m ovem en t of th e p a r ti ci p a n ts i n th e vi d eocon fer en ce s p a ce (F i g . 3 ) to b etter u n d er s ta n d th e a ffor d a n ces a n d l i m i ta ti on s of th e b r i ck s a n d b i ts d u r i n g th e com m u n i ca ti ve even t a n d m a k e p r el i m i n a r y r ecom m en d a ti on s a s to h ow to ch a n g e a n d i m p r ove th e wor k s p a ce, a cces s to con ten t, a n d i n ter a cti on a m on g th e p a r ti ci p a n ts . In th i s ex a m p l e, th e occu r r en ce of th e m ovem en ts wa s a n a l y z ed i n ti m e, cor r el a ti n g i t to th e ver b a l i n ter a cti on , a n d p h y s i ca l i n r el a ti on to th e com p u ter ’ s l oca ti on . B y cr os s i n g th e i n for m a ti on a b ou t ep i s od es a n d m ovem en ts i n s p a ce, th e s tu d y s h owed th a t th er e i s a n i n cr em en t i n th e p h y s i ca l m ovem en ts of th e p a r ti ci p a n ts on ce a s econ d p a r ti ci p a n t a r r i ves . N ever th el es s , th e wor k s p a ce a n d th e h a r d wa r e con fi g u r a ti on d o n ot a d a p t i n a fl ex i b l e m a n n er to r es p on d to th e

5. B r i ck s &

B i ts &

In ter a cti on

3 9

ch a n g e i n th e n u m b er of p a r ti ci p a n ts or th ei r l oca ti on i n r el a ti on to th e h a r d wa r e. Th i s ca n p oten ti a l l y l ea d to l i m i ted s oci a l i n ter a cti on s a n d cog n i ti ve ex p er i en ces . Th e s p a ti a l a n d m ovem en t a n a l y s i s i n d i ca ted th e effect of th e l oca ti on of th e vi d eo ca m er a i n th e u s e of s p a ce, i .e., th e p r es en ce of th e ca m er a for ces a d i a g on a l d i s p os i ti on of th e p a r ti ci p a n ts d u r i n g m os t of th e i n ter a cti on .

(a )

(b )

F i g . 2 . Meeti n g D i s cou r s e An a l y s i s : (a ) L i n k og r a p h i c R ep r es en ta ti on a n d (b ) An a l y s i s of Tem p or a l D u r a ti on of E p i s od es

F i g . 3 . Movem en t &

In ter a cti on i n th e Vi d eocon fer en ce Wor k s p a ce

5 .3 M o b ile L e a r n e r s in E - le a r n in g S p a c e s Th i s s ecti on tu r n s th e a tten ti on fr om th e a n a l y s i s of fi x ed s etti n g s for i n ter a cti on i n wh i ch a r ch i ved k n owl ed g e, i n for m a ti on , a n d p r od u ct m od el s a r e s h a r ed d u r i n g a com m u n i ca ti ve even t to fa ci l i ta te tea m wor k a n d b u i l d com m on g r ou n d , to th e n eed s of m ob i l e l ea r n er s to ca p tu r e i n for m a l k n owl ed g e i n d i ver s e for m a l a n d i n for m a l e-l ea r n i n g s p a ces . Th e s p eci fi c col l a b or a ti on tech n ol og y th e s tu d y focu s ed TM TM on wa s R E CAL L [ 2] . R E CAL L i s a l ea r n i n g a n d col l a b or a ti on tech n ol og y th a t fa ci l i ta tes tr a n s p a r en t a n d cos t effecti ve ca p tu r e, s h a r i n g , a n d r e-u s e of k n owl ed g e. TM R E CAL L i s a d r a wi n g Ja va a p p l i ca ti on th a t ca p tu r es k n owl ed g e i n i n for m a l m ed i a s u ch a s s k etch es , a u d i o a n d vi d eo. S c e n a r i o s . Two s cen a r i os a r e offer ed to d i s cu s s th e u s e of R E CAL L tech n ol og y a n d i ts r el a ti on to b r i c k s & b i t s & i n t e r a c t i o n : I n t e r a c t i v e L e c t u r e s a n d T e a m w o rk . Th e q u es ti on s r a i s ed b y th e b r i c k s & b i t s & i n t e r a c t i o n p er s p ecti ve a r e TM - How d oes R E CAL L i m p a ct th e wor k s p a ce a n d th e p l a ce th e i n ter a cti on a cti vi ty ca n ta k e p l a ce? - How d oes r i ch con ten t i m p a ct th e l evel of r eten ti on i n th e i n ter a cti ve l ectu r e s cen a r i o, a n d th e q u a l i ty of th e com m u n i ca ti on i n th e tea m wor k s cen a r i o? TM - How d oes R E CAL L ch a n g e th e fl ow of com m u n i ca ti ve even ts ?

4 0

R . F r u ch ter

B r i c k s . Sp a ce d es i g n h a s to ta k e i n to con s i d er a ti on th a t b r a i n s tor m i n g for n ew i d ea s a n d tea m i n ter a cti on d o n ot n eces s a r i l y h a ve to ta k e p l a ce i n th e offi ce, cl a s s r oom , or l a b , i n fa ct often th ey ta k e p l a ce a t th e coffee h ou s e, a i r p or t g a tewa y , etc. Th e p a p er i d en ti fi es th e fol l owi n g wor k a n d l ea r n i n g s p a ces : e - S p a c e ( e l e c t r o n i c s p a c e ) – a for m a l a n d fl ex i b l e P B L L a b th a t s u p p or ts th e d i ver s e a cti vi ti es of m ob i l e l ea r n er s , s u ch a s l ectu r e, p r es en ta ti on s , tea m wor k , i n d i vi d u a l wor k . An ex a m p l e of th e m ob i l e, wi r el es s , fl ex i b l e P B L L a b s p a ce th a t i s TM a u g m en ted wi th R E CAL L wa s b u i l d a t Sta n for d . Th e d es i g n of th e P B L l a b wa s g r ou n d ed i n cog n i ti ve a n d s i tu a ti ve l ea r n i n g th eor y . Th e cog n i ti ve p er s p ecti ve ch a r a cter i z es l ea r n i n g i n ter m s of g r owth of con cep tu a l u n d er s ta n d i n g a n d g en er a l s tr a teg i es of th i n k i n g a n d u n d er s ta n d i n g [ 5] . Th e d es i g n of th e P B L L a b --to p r ovi d e tea m i n ter a cti on wi th th e p r ofes s or , wi th i n d u s tr y m en tor s a n d tea m own er s -p r ovi d es a s tr u ctu r e for m od el i n g a n d coa ch i n g wh i ch s ca ffol d s th e l ea r n i n g p r oces s , b oth i n th e d es i g n a n d con s tr u cti on p h a s es , a s wel l a s for tech n i q u es s u ch a s a r ti cu l a ti n g a n d r efl ecti n g on cog n i ti ve p r oces s es . Th e s i tu a ti ve p er s p ecti ve s h i fts th e focu s of a n a l y s i s fr om i n d i vi d u a l b eh a vi or a n d cog n i ti on to l a r g er s y s tem s th a t i n cl u d e i n d i vi d u a l a g en ts i n ter a cti n g wi th ea ch oth er a n d wi th oth er s u b s y s tem s i n th e en vi r on m en t [ 6] . Th e P B L L a b i s b u i l t a s a fl ex i b l e l ea r n i n g s p a ce th a t ca n b e r econ fi g u r ed b y fa cu l ty or s tu d en ts on a n a s -n eed ed b a s i s to a ccom m od a te th e d i ffer en t l ea r n i n g a n d tea ch i n g a cti vi ti es (F i g . 4 a ).

(a )

(b )

(c)

F i g . 4 . E x a m p l es of e - S p a c e , d - S p a c e , a n d g - S p a c e

d - S p a c e (d i s t r i b u t e d s p a c e ) – a n i n for m a l wor k s p a ce th a t s u p p or ts th e m ob i l e l ea r n er wi th wi r el es s con n ecti vi ty to th e i n s tr u ctor s , tea m m em b er s , m en tor s . F i g . 4 b i l l u s tr a tes a n ex a m p l e of th e P B L L a b wi r el es s coffeeh ou s e d - S p a c e a t Sta n for d a s a s oci a l wor k , a n d l ea r n i n g s p a ce wh er e l ea r n er s g et tog eth er a n d u s e th ei r m oTM b i l e l a p top s a u g m en ted wi th R E CAL L , vi d eocon fer en ce, a n d oth er s ta n d a r d a p p l i ca ti on s u s ed i n p r oj ects . g - S p a c e (g l o b a l s p a c e ) – a for m a l a n d fl ex i b l e P B L L a b th a t s u p p or ts l a r g e g r ou p i n ter a cti on s i n b oth col l oca ted a n d g l ob a l g eog r a p h i ca l l y d i s tr i b u ted vi d eoTM con fer en ce a n d R E CAL L con n ecti vi ty (F i g . 4 c). In s u ch a b r oa d s p a ce, i .e., e - S p a c e , d - S p a c e , a n d g - S p a c e , th a t p r ovi d es s m ooth tr a n s i ti on s b etween for m a l a n d i n for m a l s etti n g s , l ea r n i n g a n d wor k occu r s a n y wh er e. Con s eq u en tl y , con ten t, k n owl ed g e, a n d p eop l e wa l k wi th th e i n d i vi d u a l l i k e a vi r tu a l k n owl ed g e b u b b l e (k - b u b b l e ). TM B i t s . Th e R E CAL L a p p l i ca ti on en cod es a n d s y n ch r on i z es a u d i o/vi d eo a n d s k etch . P r od u cti on a n d r ep l a y u s es a cl i en t-s er ver a r ch i tectu r e. On ce a s es s i on i s com p l ete, th e d r a wi n g a n d vi d eo/a u d i o i n for m a ti on i s a u tom a ti ca l l y i n d ex ed a n d p u b l i s h ed on a web s er ver th a t a l l ows d i s tr i b u ted a n d s y n ch r on i z ed p l a y b a ck of

5. B r i ck s &

B i ts &

In ter a cti on

4 1

th e s es s i on a n d fr om a n y wh er e a t a n y ti m e. Th e u s er i s a b l e to n a vi g a te th r ou g h th e s es s i on b y s el ecti n g i n d i vi d u a l d r a wi n g el em en ts a s a n i n d ex a n d j u m p to th e TM p a r t of i n ter es t. Th e R E CAL L tech n ol og y i n ven ti on i s cu r r en tl y b ei n g p a ten ted . Th i s r i ch a n d i n for m a l con ten t, i .e., s k etch , a u d i o, a n d vi d eo en a b l es th e p a r ti ci p a n ts to com m u n i ca te th e r a ti on a l e a n d con tex t i n wh i ch th ei r con cep ts , p r op os ed ch a n g es , or q u es ti on s ca m e u p . Th e i n ter a cti vi ty wi th th e con ten t en a b l es u s er s to a cces s th e con ten t p a r t of i n ter es t a n d m a n a g e i n for m a ti on over l oa d . I n t e r a c t i o n s . Th e s k etch i s a n a tu r a l m od e for d es i g n er s , i n s tr u ctor s , or l ea r n er s to com m u n i ca te i n h i g h l y i n for m a l a cti vi ti es s u ch a s b r a i n s tor m i n g s es s i on s , l ectu r es , or Q & A s es s i on s . Often a s k etch i ts el f i s m er el y th e veh i cl e th a t s p a wn s d i s cu s s i on s a b ou t a p a r ti cu l a r d es i g n i s s u e. Th u s , fr om a d es i g n k n owl ed g e ca p tu r e p er s p ecti ve; ca p tu r i n g b oth th e s k etch i ts el f a n d th e d i s cu s s i on th a t p r ovi d es th e con tex t b eh i n d th e s k etch a r e i m p or ta n t. It i s i n ter es ti n g to n ote th a t tod a y ’ s s ta te-of-p r a cti ce n ei th er i s ca p tu r ed a n d k n owl ed g e i s l os t wh en th e wh i teb oa r d s i s TM er a s ed . R E CAL L a ct a s a n ex p l or a ti on en vi r on m en t th a t ca p tu r es b oth a n i n d i vi d u a l m em or y of i d ea s a n d r a ti on a l e i - m e m o , a n d tea m m em or y t - m e m o . TM R E CAL L offer s s om e k ey b en efi ts for p r od u cer s a n d con s u m er s of r i ch con ten t, s u ch a s , z er o over h ea d cos t for i n d ex i n g a n d p u b l i s h i n g on th e Web r i ch con ten t i n th e for m of s k etch es , a u d i o a n d vi d eo, a s wel l a s r ea l -ti m e i n ter a cti vi ty . TM In ter m s of i n ter a cti on a m on g tea m m em b er s R E CAL L en a b l es a fa s ter tu r n over of i n for m a ti on a n d tea m feed b a ck ; i n s tr u ctor s ca n h a ve a n i n s i g h t i n to l ea r n er ’ s th ou g h t p r oces s b ey on d th e ex er ci s e r es u l t/a n s wer or q u es ti on ; s i m i l a r b en efi ts ca n b e ob s er ved i n p l a y m od e or i n cu s tom er r el a ti on m a n a g em en t. Si n ce th e k n owl ed g e i s i n con tex t, p a r ti ci p a n ts ca n m a k e i n for m ed d eci s i on s .

5 .4 E m e r g in g C h a n g e s I n f lu e n c e d b y B r ic k s & I n te r a c tio n

B its &

B oth s tu d i es offer i n s i g h ts i n ter m s s oci o-tech n i ca l -en vi r on m en ta l ch a n g es th a t n eed to b e con s i d er ed fr om a l l th r ee a s p ects , en vi r on m en ta l – b r i c k s , tech n i ca l – b i t s , a n d s oci a l – i n t e r a c t i o n . Al l th r ee a s p ects con s ta n tl y i n fl u en ce ea ch oth er . Th e i n fl u en ce of b r i c k s on b i t s i n d i ca tes th a t th e wor k s p a ce con fi g u r a ti on ca n en h a n ce or l i m i t th e vi s i b i l i ty of p a r ti ci p a n ts i n a m u l ti -m od a l vi d eocon fer en ce a n d th e a wa r en es s of s h a r ed con ten t d i s p l a y ed on a m on i tor or s cr een . Con s eq u en tl y , b etter s oftwa r e a n d h a r d wa r e th a t s u p p or ts z oom i n g of th e COI wou l d i m p r ove th e com m u n i ca ti ve even t th a t h a s to ta k e p l a ce i n a fi x ed a n d con fi n ed wor k s p a ce. A s i m p l e s ol u ti on th a t ca n h el p i m p r ove th e vi s i b i l i ty a n d a wa r en es s i n a vi d eo con fer en ce s etti n g wou l d b e to ch a n g e th e l oca ti on of th e vi d eo ca m er a s o th a t th er e i s a n over l a p of th e a r ea s (1) a n d (2) s h own i n F i g . 1. Th e i n fl u en ce of b i t s on b r i c k s l ea d s to ch a n g es s u ch a s d evel op m en t of fl ex i b l e s tr u ctu r a l el em en ts a n d m ob i l e d evi ces i n th e wor k s p a ce th a t a d a p t a n d a d j u s t to a d d r es s th e n eed s for vi s u a l i z a ti on , com p os i ti on a n d m a n i p u l a ti on of r i ch con ten t, or em b ed d ed m u l ti -m ed i a d evi ces i n wa l l s a n d fu r n i tu r e. Th e i n fl u en ce of b r i c k s on i n t e r a c t i o n r eq u i r es p a r ti ci p a n ts to ch a n g e th ei r b eh a vi or , a cq u i r e n ew h a b i ts a s th ey m ove a n d i n ter a ct i n th e wor k s p a ce, a s wel l a s

4 2

R . F r u ch ter

s h a r e th e wor k s p a ce to a l l ow a wa r en es s a n d vi s i b i l i ty i n d i ffer en t s cen a r i os , i .e., i n d i vi d u a l p r es en ce, s m a l l or l a r g e col l oca ted tea m s l i n k ed to g l ob a l p a r tn er s . Th e i n fl u en ce of i n t e r a c t i o n on b r i c k s l ea d s u s to r eth i n k i n g th e d es i g n of s p a ces a s a d j u s ta b l e wor k s p a ces , e.g ., m ob i l e p a r ti ti on wa l l s , fl ex i b l e fu r n i tu r e, n etwor k a n d p ower i n fr a s tr u ctu r e th a t a l l ows con n ecti vi ty a n y wh er e a n y ti m e, to a d d r es s i n d i vi d u a l a n d tea m wor k . In a d d i ti on , b r i c k s i n th e for m of for m a l a n d i n for m a l wor k , l ea r n i n g , p l a y , a n d com m u n i ty s p a ces h a ve to fa ci l i ta te s m ooth tr a n s i ti on s a m on g e - S p a c e s , d - S p a c e s , a n d g - S p a c e s . Th e i n fl u en ce of i n t e r a c t i o n on b i t s d i r ects ou r th i n k i n g towa r d s th e d es i g n a n d d evel op m en t of n ew s oftwa r e a n d h a r d wa r e tool s th a t ca n for i n s ta n ce r es ol ve vi s i b i l i ty com m u n i ca ti on p r ob l em s s u ch a s “ fa k e p oi n ti n g ” i .e., p oi n ti n g a t i n for m a ti on on th e s cr een wi th h a n d g es tu r es , “ th e g a z e” i .e., p r ovi d i n g ey e-con ta ct of r em ote p a r ti ci p a n ts i n a vi d eocon fer en ce, n o m a tter wh er e th ey l ook . Th e i n fl u en ce of b i t s on i n t e r a c t i o n r eq u i r es i n d i vi d u a l s ’ b eh a vi or a n d tea m d y n a m i cs ch a n g e, a s n ew p r otocol s a r e for m a l i z ed a n d a d op ted b y th e p a r ti ci p a n ts to b es t ta k e a d va n ta g e of em er g i n g col l a b or a ti on tech n ol og i es . Soci a l i n tel l i g en ce evol ves a s p a r ti ci p a n ts l ea r n h ow to s h a r e a n d i n ter a cti vel y m a n i p u l a te r i ch con ten t, i .e., b i t s . Th i s p r oces s en a b l es a g l ob a l l y d i s tr i b u ted or col l oca ted p r oj ect tea m to b u i l d a “ com m on g r ou n d ” or s h a r ed u n d er s ta n d i n g of th e g oa l s , con s tr a i n ts , a n d s ol u ti on a l ter n a ti ves . Th e a va i l a b i l i ty of con tex t i n wh i ch con ten t wa s cr ea ted op en s n ew d i m en s i on s i n th e u n d er s ta n d i n g of d es i g n d eci s i on s . Mor e th a n th a t s h a r ed r i ch con ten t i m p a cts th e l evel of r eten ti on , a tten ti on , a n d th e q u a l i ty of th e com m u n i ca ti on . F i n a l l y , i n b u i l d i n g n ew s oci a l i n tel l i g en ce i n d i vi d u a l s l ea r n to s h a r e m or e i n for m a ti on i n a ti m el y fa s h i on a t a fa s ter r a te, a s wel l a s b ecom e r es p on s i ve to r eq u es ts for i n for m a ti on . Th i s p r oces s l ea d s to fa s ter d es i g n -b u i l d d eci s i on i ter a ti on s a n d s h or ter ti m e-to-m a r k et s ol u ti on s i n a n i n d u s tr y en vi r on m en t, a n d s u p p os ed l y a m or e i n ten s e s oci a l i n ter a cti on i n a n y com m u n i ty .

R e fe r e n c e s 5.1 F r u ch ter , R ., Ca va l l i n , H. a n d Ma r ti n v is ib ility a n d a w a r e n e s s in v id e o c o ICCCB E -VIII Con fer en ce, ed . R . Au g u s t 14 -16, 2000, CA. 5.2 F r u ch ter , R . a n d Yen , S., (2000) “ R E Con fer en ce, ed . R . F r u ch ter , K. R od d 5.3 F r u ch ter , R . “ A r c h i t e c t u r e , E n g i n e e D e s i g n a n d L e a r n i n g S p a c e , ” Jou r n 1999, Vol 13 N o.4 , p p 261-270. 5.4 G ol d s ch m i d t, G . “ O n D e s i g n R e a s o n Sci en ce of Con cep ts , F r a n ce 1997. 5.5 D ewey , J. (1928 , 1958 ) “ E x p e r i e n c e 5.6 G r een o, J.G . (1998 ) “ T h e S i t u a t i v i t y P s y c h o l o g i s t , 53 , 5-26.

, W.M., “ T O S E E O R N O T T O S E E : T h e r o l e o f n f e r e n c e - m e d i a t e d t e a m w o r k , “ P r oc. of ASCE F r u ch ter , K. R od d i s , F . P en a -Mor a , Sta n for d , C A L L in i s , F . P en a r in g , C o n a l of Com

A c t i o n , ” P r oc. of ASCE ICCCB E -VIII -Mor a , Sta n for d , Au g u s t 14 -16, CA. s tr u c tio n T e a m w o r k : A C o lla b o r a tiv e p u ti n g i n Ci vi l E n g i n eer i n g , Octob er

i n g : C o n t e n t s a n d S t r u c t u r e , ” Wor k s h op on Th e a n d n a t u r e .” N ew Yor k : D over . o f K n o w in g , L e a r n in g a n d R e s e a r c h .” A m e r ic a n

6 . A D is tr ib u te d M u lti-a g e n t S y s te m S e lf-E v a lu a tio n o f D ia lo g s

fo r th e

Al a i n Ca r d on L IH L e Ha vr e Un i ver s i ty a n d L IP 6 P a r i s VI Un i ver s i ty Al a i n .Ca r d on @ l i p 6.fr A b s t r a c t . We ca n i n ter p r et m ea n i n g of th e com m u n i ca ti on s on tol og y on th e d i s cou r s es . We ca n l i n k th es e s ta ti c on tol og a n d ex p r es s es th e ch a n g es of th e m u l ti -a g en t or g a n i z a ti on s wa y . So, we ca n ex h i b i t th e m ea n i n g of th e d i s cou r s es b y b etween two a g en t or g a n i z a ti on s , a n d i n r ea l ti m e.

b etween u s er s i f we h a ve l a r g e i es wi th m u l ti -a g en ts s y s tem s , i n a g eom etr i ca l a n d d y n a m i c em er g en ce i n a s y s tem i c l oop

K e y w o r d s : com m u n i ca ti on , m u l ti -a g en t s y s tem , em er g en ce, m or p h ol og y , s h a r ed k n owl ed g e, m ea n i n g .

6 .1 I n t r o d u c t io n A com m u n i ca ti on a l s y s tem u s i n g l a r g e n etwor k s a n d i n vol vi n g m a n y u s er s ca n b e s een i n two wa y s . Th e fi r s t i s p oi n t of vi ew of th e ex ch a n g e of k n owl ed g e a n d of s h a r ed k n owl ed g e b etween th e u s er s , i n a cog n i ti ve wa y . Th e s econ d i s th e p oi n t of vi ew of th e con tr ol i n a n d of th e s y s tem , i n a s oci a l wa y . In fa ct, u s i n g n etwor k s , u s er s h a ve to com m u n i ca te a n d u s e l a r g e k i n d of k n owl ed g e: th e ex ch a n g e of i n for m a ti on i s a l wa y s a n ex ch a n g e of k n owl ed g e. Wi th th i s p r a cti ce, u s er s m a k e u p a n ew d y n a m i c s oci a l s p a ce wh er e p r ob l em s of cu l tu r e, of p ower a n d of s oci a l tr a n s for m a ti on s s p r i n g u p . An d th e q u es ti on of th e con tr ol i s i n h er en t i n s u ch fi el d s . So, we ca n h a ve a ver y d eep or a s oft con tr ol b u t a l wa y s we h a ve con tr ol ex p r es s ed b y th e s oci ety i ts el f a b ou t th e g oa l s of p eop l e u s i n g n etwor k s . Com m u n i ca ti on i s , u p p er th e tech n i ca l a s p ect, a s oci a l a ct i n vol ved a p os s i b l e tr a n s for m a ti on of th e s oci a l s tr u ctu r es . Th e ex p r es s i on of th e con tr ol of u s er s i n com m u n i ca ti on a l n etwor k s i s a n a tu r a l ten d en cy i n ou r s oci eti es , for m a i n ta i n coh es i on a n d a voi d b r ea k i n g . E x ch a n g e of i n for m a ti on b etween u s er s i s ex ch a n g e of k n owl ed g e a n d i m p l i es th e d evel op m en t a n d th e m od i fi ca ti on of th e u s er s ’ g r ou p s . Th i s s tr u ctu r e, th es e or g a n i z a ti on a l m od i fi ca ti on s m u s t b e k n own for s om e s oci a l a n d p ol i ti c s tr u ctu r es p u tti n g i n p l a ce th e n etwor k s a n d th ei r s fa ci l i ti es . B u t th es e or g a n i z a ti on a l m od i fi ca ti on s ca n a l s o b e k n own b y u s er s th em s el ves a n d th en ta ck l e a n ew s oci a l s p a ce. In th i s ca s e, th e ex ch a n g ed k n owl ed g e i s a u tom a ti ca l l y a u g m en ted wi th i ts i n ter p r eta ti on a n d i ts s oci a l i m p l i ca ti on s . We p r es en t th e a r ch i tectu r e of s u ch a s y s tem a l l owi n g th e r ep r es en ta ti on of th e m ea n i n g of com m u n i ca ti on s b etween u s er s . Th e a r ch i tectu r e s tr on g l y u s es th e m u l ti -a g en t p a r a d i g m .

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 4 3 -50, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

4 4

A. Ca r d on

6 .2 S y s t e m

G e n e r a l A r c h ite c tu r e

As we g en er a l l y b el i eve th e wor l d to b e con s i s ten t, we g en er a l l y ex p ect th e s a m e fr om ou r r ep r es en ta ti on of i t. Wh i l e th e fa ct we k n ow th i s i s n ’ t tr u e, we u s u a l l y th eor eti ca l l y con s i d er ou r a b i l i ti es to p er cei ve i t to b e r el i a b l e a n d con s i s ten t, a n d we d own p l a y th e p os s i b l e m i s ta k es we ca n com m i t wh en d oi n g s o. F or ex a m p l e, cl a s s i ca l Com m u n i ca ti on a n d In for m a ti on Sy s tem s u s u a l l y s u p p os e p eop l e u s i n g i t s a y ’ ’ th e tr u th ’ ’ , th a t i s th a t th ey k n ow wh a t th ey ’ r e s a y i n g for s u r e a n d d on ’ t l i e [ 7] . Com m u n i ca ti on Sy s tem s often d on ’ t d ea l a l l too wel l wi th con tr a d i ctor y k n owl ed g e s ou r ces , b eca u s e of th e l a ck of cor r ect i n for m a ti on or th r ou g h m a l evol en ce. An oth er ex a m p l e i s th e i n p u ts fr om r ob ot s en s or s : th es e s en s or s a r en ’ t p er fect a n d s o n ei th er a r e th e d a ta th ey tr a n s m i t. Th es e ca n th er efor e b e con tr a d i ctor y . An d we m u s t th er efor e d ea l wi th th es e con tr a d i cti on s . An oth er i m p or ta n t ch a r a cter i s ti c of k n owl ed g e i n com m u n i ca ti on a l s i tu a ti on s i s i ts i n h er en tl y d y n a m i c n a tu r e. Wh en we con s i d er a s y s tem th a t h a s to h el p p eop l e i n th ei r d eci s i on m a k i n g p r oces s i n a r ea l ti m e fr a m ewor k , wh a t i s r i g h t a t on e m om en t m i g h t p r ove i n cor r ect m i n u tes l a ter [ 2] . How th en ca n a s y s tem cop e wi th s u ch a fl u ctu a ti n g k n owl ed g e a n d i n wh i ch wa y i t ca n ex p r es s th e n a tu r e a n d for m of th e con tr ol ? It ob vi ou s l y h a s to k eep i n m i n d m a n y p os s i b l e s cen a r i os , i n oth er wor d s , i t h a s to con cei ve m a n y p os s i b l e fu tu r e wor l d s i n or d er to m a tch th em to r ecor d ed p l a n s s o th a t i t ca n effi ci en tl y h el p i n th e d eci s i on m a k i n g p r oces s . However , on ce th e s y s tem h a s ch os en s om e cu r r en t wor l d r ep r es en ta ti on , i t h a s to r eta i n th e oth er p os s i b l e r ep r es en ta ti on s s o a s to b e a b l e to a l ter i ts cu r r en t s ta te i n ca s e th e a ctu a l s i tu a ti on s h i fts . We ca n focu s on th e s i x l evel s of m od el of Com m u n i ca ti on a n d In for m a ti on Sy s tem (CIS) wh i ch a r e th e or g a n i z a ti on l evel s for com p l ex s y s tem s [ 5] : 1. P h y s i ca l wor l d , ob j ecti ve en ti ti es , 2. Sp a ce of d evel op m en t of th e en ti ti es , 3 . Movem en t, or g a n i z a ti on s , p l a n n i n g 4 . Com m u n i ca ti on of i n for m a ti on 5. Va l u es , s y m b ol s , m ea n i n g of th e p h en om en on , i n ten ti on s , 6. R u l es of th e s oci a l g a m e, p ower r el a ti on s , em er g en ce of th e g l ob a l m ea n i n g of th e p h en om en on Th e th r ee fi r s t l evel s b el on g to th e fi el d of th e cl a s s i ca l In for m a ti on Sy s tem , th e fou r th a l l ows th e d y n a m i c or g a n i z a ti on of th e th r ee p r evi ou s . Th e l evel s fi ve a n d s i x b el on g to th e s oci a l , p s y ch ol og i ca l a n d cu l tu r a l fi el d . Th ey ca n n ot b e r ep r es en ted b y a -p r i or i d efi n ed s tr u ctu r es u s i n g fi x ed p r i m a l com p on en ts : th e i m p or ta n ce a n d k i n d of p s y ch ol og i ca l a n d s oci a l ca teg or i es th ey r ep r es en t d ep en d on th e cu r r en t s i tu a ti on i ts el f. Th ey ca n n ot b e d ecom p os ed i n to fi x ed com p on en ts , for th e s a m e r ea s on . L i k e th i s , th es e l evel s b el on g to a ver y com p l ex d om a i n . We a r e i n ter es ted i n th es e l a s t l evel s to ta k e i n to a ccou n t i n ten ti on s , op i n i on s a n d j u d g m en ts i n th e com m u n i ca ti on p r oces s , i n or d er to d efi n e th e g ood k n owl ed g e d el i ver ed to th e a ctor s .

6. A D i s tr i b u ted Mu l ti -a g en t Sy s tem

for th e Sel f-E va l u a ti on of D i a l og s

4 5

6 .3 R e p r e s e n t a t io n o f t h e S e m a n t ic o f t h e C o m m u n ic a t io n A c t Th e a p p r oa ch con s i s ts i n th e k n owl ed g e of th e s i tu a ti on of com m u n i ca ti on , of th e r ea l a n d ob j ecti ve fa cts a n d a l s o of th e m en ta l r ep r es en ta ti on s of th e s i tu a ti on b y a ctor s th em s el ves . So, on e i n cl u d es th e fa ctu a l i n for m a ti on a n d th e el a b or a ti on of th e p r oces s of d eci s i on , th e op i n i on s a n d j u d g m en ts of th e d i ffer en t a ctor s a b ou t th e d i ffer en t s i tu a ti on s a n d a b ou t th em s el ves . In th i s a p p r oa ch , th e i n ten ti on a l i ty i n th e a ct of i n for m a ti on ex ch a n g e ta k es p r eced en ce over th e tr a n s m i s s i on of n eu tr a l i n for m a ti on , a s i n th e cl a s s i ca l In for m a ti on Sy s tem s . We u s e a n oti on of a g en t a s a s oftwa r e en ti ty [ 8 ] . Th i s n oti on p u ts a g en t’ s n oti on l i k e a n a cti on en ti ty d efi n ed a t th e con s tr u cti on s tep of a s oftwa r e s y s tem a n d op er a ti n g i n th e s etti n g of a n op en p r ob l em to s ol ve. A m u l ti -a g en t s y s tem (MAS) i s con s ti tu ted of a s et of a g en t or g a n i z a ti on s a n d i s s i tu a ted i n a n en vi r on m en t com p os ed of m a n y ob j ects th a t a r e n ot a g en ts th a t a r e es s en ti a l l y r ea cti ve i n a p er m a n en t wa y . Th i s s y s tem com m u n i ca tes wi th i ts en vi r on m en t b y th e a cti on of s p eci fi c a g en ts s o-ca l l ed i n ter fa ci n g a g en ts . Th e a g en ts of th e MAS u s e ob j ects of th ei r wor l d a s wel l a s a cti on s of th e oth er a g en ts to a ch i eve s om e va r i ou s a cti on s . Th ey u n i te th ei r a cti on s to d efi n e s om e col l ecti ve b eh a vi or s . Th e effi ci en t, vi s i b l e b eh a vi or of MAS wi l l es s en ti a l l y b e a ch i eved b y th e b eh a vi or of th e a g en ts a n d wi l l b e con s tr u cted th er efor e of d i s tr i b u ted m a n n er . Th i s i s i n th e a g en ts , a n d es s en ti a l l y i n th e a g en ts th a t wi l l b e d i s tr i b u ted th e ch a r a cter s of a cti on , th e effects th e s y s tem i n wh ol e p r od u ces on th e en vi r on m en t [ 1] . We s a w i n th e d efi n i ti on of CIS th a t th e th r ee fi r s t l evel s d es cr i b e th e ob j ecti ve s i tu a ti on . Th es e l evel s a r e p r oces s ed b y th e com m u n i ca ti n g i n for m a ti on l evel (s o n a m ed L evel 4 ). We m a k e th e h y p oth es i s th a t s om e a g en ts ca n a l s o r ep r es en t th e l evel s 5 a n d 6. Th es e l evel s con s ti tu te a s p eci fi c d om a i n , ex p r es s i n g eva l u a ted k n owl ed g e, s u b j ecti ve, s oci a l a n d cu l tu r a l a s p ects a b ou t th e s i tu a ti on i n p r og r es s . Th ey a r e a b ove th e fou r p r evi ou s on es a n d a l ter th ei r s tr u ctu r e. Th i s i s th e fi r s t h y p oth es i s of s el f-r efer en ce. Th ey ca n n ot b e r ep r es en ted , i n th e s y s tem , b y fu n cti on a l a n d s ta ti c p r e-d efi n ed ca teg or i es : ea ch ch a r a cter i n th es e l evel s , i s m a i n l y a n a c t o f c o m m u n i c a t i o n . It m ea n s , th a t ea ch com m u n i ca ti on i s wr a p p ed b y a l ot of a g en ts r ep r es en ti n g th e ca teg or i es of m ea n i n g of th e eva l u a ted com m u n i ca ti on . Th i s s et of en ti ti es q u a l i fi es th e com m u n i ca ti on a n d m od i fi es p h y s i ca l l y a p a r t of th e s tr u ctu r e of th e s y s tem i ts el f: th ey a r e effecti ve s oftwa r e a cti on s . So we ex p r es s ca teg or i es i n l evel s 5 a n d 6, a t th e on tol og i ca l l evel , wi th a cts of com m u n i ca ti on [ 3 ] . Th e ch a r a cter i z a ti on of th e s i tu a ti on a ccor d i n g to th e d i ffer en t a ctor s i s r ep r es en ted b y th e va r i a b i l i ty of s i tu a ti on s , op i n i on s , j u d g m en ts , p oi n ts of vi ew. Th e r ep r es en ta ti on of th i s ch a r a cter i z a ti on i n th e s y s tem wi l l b e a s tr u ctu r a l m od i fi ca ti on i n s p a ce a n d ti m e, wr a p p i n g ever y com m u n i ca ti on . Th e m a i n h y p oth es i s i s th a t p l a s ti c m od el a n d p l a s ti c s oftwa r e s tr u ctu r es a r e wel l a d a p ted to r ep r es en t a ver y evol vi n g p h en om en on .

4 6

6 .4

A. Ca r d on

S e m a n tic T r a its a n d A g e n ts

Th e on l y m od el , wh i ch a l l ows s u ch a p l a s ti c r ep r es en ta ti on , u s es th e Mu l ti -Ag en t Sy s tem s : we r ep r es en t th e d i ffer en t ch a r a cter s of th e com m u n i ca ti on b y a l ot of s oftwa r e a g en ts (c.f. F i g . 1). Th e s en ten ces ex ch a n g ed b etween u s er s a r e com p os ed of s p eci fi c wor d s com i n g fr om th e d i ffer en t on tol og i es of th e d i s cou r s e d om a i n . E a ch wor d or s et of wor d s i n ea ch m es s a g e a r e l oca ted i n on e or m or e on tol og i es [ 6] . We ca l l s u ch a wor d , or g r ou p of wor d s , a s e m a n t i c t r a i t . It ex p r es s es a ch a r a cter of th e cu r r en t s i tu a ti on . Con cr ete Actor

Con cr ete Actor

MAS of a s p ectu a l a g en ts Com m u n i ca ti on s

As p ectu a l a g en ts Con cr ete Actor

Con cr ete Actor

F i g . 1 . Softwa r e a g en ts wr a p p i n g th e com m u n i ca ti on a l s y s tem .

F or ea ch s em a n ti c tr a i t, we a s s oci a te a l ot of s oftwa r e a g en ts , th e s o-ca l l ed a s p e c t u a l a g e n t s . An a s p ectu a l a g en t i s a wea k a g en t r ei fy i n g a s em a n ti c tr a i t. F or ea ch s em a n ti c tr a i t, we ca n a s s oci a te s ever a l a s p ectu a l a g en ts , s p eci fy i n g th e s em a n ti c tr a i t, i ts con tr a r y , i ts op p os i te, th e d er i ved tr a i ts … So, we ob ta i n , for ea ch s em a n ti c tr a i t, a l ot of a s p ectu a l a g en ts th a t m u s t cor r es p on d . F or a l l th e s em a n ti c tr a i ts ex p r es s i n g th e wh ol e of on tol og y of th e d om a i n , we h a ve a l a r g e s et of a s p ectu a l a g en ts , th a t a r e n or i n d ep en d en ts . Th e a g en ts a r e l i n k ed b y th ei r a cq u a i n ta n ces , th ey ca n com m u n i ca te, th ey ca n a wa k e or k i l l oth er s a g en ts , co-op er a te a n d for m g r ou p s ex p r es s i n g com p l ex a s s oci a ti on s of s em a n ti c tr a i ts [ 3 ] .

6 .5 A s p e c t u a l A g e n t O r g a n iz a t io n Th en , for ea ch s en ten ce ex ch a n g ed b etween u s er , we h a ve a l ot of s em a n ti c tr a i ts ex p r es s ed i n a s et of a cti va ted a s p ectu a l a g en ts , th e a g en t th a t m a tch on th e d i ffer en t s em a n ti c tr a i ts , a g r ou p for th e s en d er a n d a n oth er for th e r eci p i en ts . We a u g m en t th es e s em a n ti c tr a i ts wi th s om e s u b j ecti ve a s p ects a b ou t th e p er cep ti on of th e s i tu a ti on th e u s er s ca n h a ve, l i k e j u d g m en ts or feel i n g s l i k e fea r , d r ea d , s a ti s fa cti on , l i e … An d we r ei fy th es e s u b j ecti ve s em a n ti c tr a i ts wi th oth er s a s p ectu a l a g en ts . Th es e a g en ts a wa k e oth er s i n ob s er vi n g a s p ectu a l a g en ts a n d s o m a k e em er g en ce of th e s em a n ti c tr a i ts th ey m a tch (c.f. F i g . 2). L i k e th i s , we ca n ex p r es s

6. A D i s tr i b u ted Mu l ti -a g en t Sy s tem

for th e Sel f-E va l u a ti on of D i a l og s

4 7

b y a g en ts th e s i x l evel s of th e CIS, i n cl u d i n g j u d g m en ts a n d feel i n g ex p r es s ed b y u s er s . Th e a s p ectu a l a g en ts a wa k e or k i l l oth er s a g en ts , s tr u g g l e wi th s om eon e, coop er a te wi th oth er s a n d for m th a t we ca l l a n a g e n t l a n d s c a p e , a ver y d y n a m i c a g en t or g a n i z a ti on ex p r es s i n g wi th a u g m en ta ti on th e s em a n ti c of ea ch com m u n i ca ted s en ten ce. Mor e th a n , th e a s p ectu a l a g en ts ta k e i n to a ccou n t th e or g a n i z a ti on a l s ta te of th e cu r r en t a s p ectu a l or g a n i z a ti on of ea ch cu r r en t u s er r ecei vi n g a n ew m es s a g e. Th ey " s et i n s i tu a ti on " th e cu r r en t m es s a g e, ta k i n g a ccou n t of th e p r evi ou s : th ey con s ti tu te a n or g a n i z a ti on a l m em or y .

Sem a n ti c tr a i t In p u t of th e s y s tem : com m u n i ca ti on s

Sem a n ti c tr a i t

A c tiv e a sp e c tu a l a g e n ts

Sem a n ti c tr a i t

F i g . 2 . Th e a s p ectu a l or g a n i z a ti on op er a ti n g on th e s em a n ti c tr a i ts

Th e b eh a vi or of th es e a g en ts , th ei r i n ter n a l tr a n s for m a ti on a n d th ei r com m u n i ca ti on r ea l i z e s p a ti a l a n d tem p or a l or g a n i z a ti on of l evel 5 a n d 6. Th e g l ob a l ch a r a cter s , wh i ch ca n b e fou n d i n th e m u l ti -a g en t s y s tem , a r e e m e r g i n g c h a r a c t e r s . Th u s , th os e a g en ts wi th th ei r own p a r ti cu l a r b eh a vi or m a y d i s tu r b th e or g a n i z a ti on of th e s y s tem a n d m a k e i t s el f-r eor g a n i z e to ex h i b i t n ew em er g i n g ch a r a cter s . In MAS, ex p ected or u n ex p ected s tr u ctu r es m a y a p p ea r . We m a k e th e h y p oth es i s th a t em er g i n g s tr u ctu r es ex p r es s th e m ea n i n g of th e com m u n i ca ti on s b etween u s er s d es cr i b i n g th em on l y i n a g eom etr i ca l wa y we ca l l a m o r p h o l o g y . Th i s em er g i n g s tr u ctu r e r ep r es en ts th e a ccu r a te vi ews a b ou t th e d i ffer en t p er cep ti on s of th e p h en om en on el a b or a ted d u r i n g com m u n i ca ti on . B eca u s e th e s y s tem i s d y n a m i c, th e wh ol e em er g i n g s tr u ctu r es ch a n g e a ccor d i n g to th e evol u ti on of th e u s er s ’ p er cei ved p h en om en on . So th e a g en t s tr u ctu r e a n d i ts evol u ti on r efl ects th e or g a n i z a ti on a n d th e evol u ti on of th e p er cei ved p h en om en on i ts el f.

4 8

A. Ca r d on

Th i s a s p ectu a l or g a n i z a ti on wi l l g r a s p th e com m u tr a ct th ei r ch a r a cter i s ti cs . Th e a s p ectu a l a g en ts r ep b eh a vi or a n d i n n er s ta tes , th e em er g en ce of s em a n p r ox i m i ty wi th th e oth er s p r evi ou s l y ex p r es s ed s em a

n i ca ti on a l d a ta i n or d er to ex r es en t, b y th ei r a cti on s , th ei r ti c tr a i ts i n a ccou n t wi th th e n ti c tr a i ts .

6 .6 T h e E m e r g in g M e a n in g o f t h e C o m m u n ic a t io n : T h e M o r p h o lo g ic a l A g e n t O r g a n iz a tio n Th e p r evi ou s l y d efi n ed a s p ectu a l a g en ts a l l ow th e ex p r es s i on of th e m ea n i n g of ea ch s em a n ti c tr a i t of th e com m u n i ca ti on i n a n a ct of com m u n i ca ti on . Th e s et of th e a l l MAS wr a p p ed to ea ch con cr ete a ctor a l l ows th e ex p r es s i on of th e wh ol e m ea n i n g of th e com m u n i ca ti on a l s i tu a ti on . Th i s m ea n i n g i s g en er a ted b y e m e r g i n g s tr u ctu r es , ex p r es s i n g th e m or p h ol og y of th e s et of MAS. F or th i s , we h a ve d efi n ed th e n oti on of f o r m o f t h e a g e n t l a n d s c a p e [ 4 ] , th a t i s th e tr a n s for m a ti on of th e a g en t l a n d s ca p e i n a g eom etr i ca l wa y . Th i s i s a n i m p or ta n t p oi n t of ou r wor k , wh er e we s tu d y th e coh er en ce a n d s ta b i l i ty of MAS ex p r es s i n g g l ob a l s en s e b u t u s i n g g eom etr i ca l ch a r a cter s of th e MAS. Th e g oa l i s to b u i l d a s tr u ctu r a l a n d i m m ed i a te con n ecti on b etween th e s et of a ctor s ’ i d ea s a n d th e l a n d s ca p e of a g en ts . Th i s n oti on i s cen tr a l i n th e m od el a n d u n d er s tood a s a r ea l n ew for m of m ea n i n g , ex p r es s i n g wi th a l ot of a g en ts th e s y n th es i s of p a r ti cu l a r for m s (th e a s p ectu a l a g en ts ) a r ou n d th e d i ffer en t con cr ete a ctor s . G i ven th e ver y g r ea t n u m b er of a s p ectu a l a g en ts , i t i s n ’ t p os s i b l e to fol l ow th em i n d i vi d u a l l y . We th er efor e s tu d y th em a s a wh ol e, d i s ti n g u i s h i n g s h a p es a n d for m s i n th e i n ter a cti on s . We a p p r eci a te a for m i n a g eom etr i ca l wa y , u s i n g th e s p eci fi c or g a n i z a ti on of th e m or p h ol og i ca l a g en ts . We ca l l th i s vi ew of th e a s p ectu a l a g en ts or g a n i z a ti on , con s i d er ed a s a p op u l a ti on , a n a g e n t l a n d s c a p e [ 3 ] , [ 7] . An a g en t l a n d s ca p e i s s p a ce ex p r es s i n g th e a cti ve a s p ectu a l a g en ts , con s i d er ed a s wel l u n d er s ta n d a b l e In th e s y s tem a n a g en t l a n d s ca p e i s r ep r es en ted b y s p eci fi c p r oj ecti on s of th e s tu d i ed a g en t or g a n i z a ti on a ccor d i n g to h ei g h t a x es . Su ch a r ep r es en ta ti on d efi n es i n fa ct a n ew s p a ce of d y n a m i c d es cr i p ti on of a n y a g en t or g a n i z a ti on . Th e h ei g h t s p a ce d i m en s i on s a r e th e fol l owi n g : • or g a n i z a ti on a l d i s ta n ce: th e s ta te of th e a g en t com p a r ed wi th th e s ta te of th e wh ol e a g en t or g a n i z a ti on , • vel oci ty : th e s p eed wi th wh i ch a n a s p ectu a l a g en t h a s d evel op ed s o fa r , • fa ci l i ty : th e ea s e wi th wh i ch a n a s p ectu a l a g en t h a s d evel op ed s o fa r , • s u p r em a cy : a m ea s u r e of th e r a ti o en em y a l l i ed of ea ch a s p ectu a l a g en t, • com p l ex i fi ca ti on : a m ea s u r em en t of th e evol u ti on of th e i n n er s tr u ctu r e of th e a g en t i n u s • i n ten s i ty of th e i n ter n a l a cti vi ty : th e ex p r es s i on of th e ex ch a n g es b etween th e i n n er com p on en ts of a s p ectu a l a g en t b efor e a cti on , • p er s i s ten ce: a m ea s u r em en t of th e ti m e of l i fe of th e a g en t, • d ep en d en cy : th e fa ct th e a g en t i s or i s n ot fr ee or d ep en d en t.

6. A D i s tr i b u ted Mu l ti -a g en t Sy s tem

for th e Sel f-E va l u a ti on of D i a l og s

4 9

We ex p r es s th e ch a r a cter s (th e d i m en s i on s ) of th i s s p a ce u s i n g s p eci fi c a g en ts . Th e m or p h ol og i ca l a g en ts a r e th e ex p r es s i on of th e a g g r eg a ti on of a s p ectu a l a g en ts i n th e l a n d s ca p e m a d e wi th th es e a g en ts , a ccor d i n g to th os e h ei g h t cr i ter i a . Th e s et of m or p h ol og i ca l a g en t’ s for m a k i n d of d y n a m i c s p a ce, ea ch p oi n t i n th i s s p a ce i s i n fa ct a m or p h ol og i ca l a g en t. Su ch a n ex p r es s i on of a m a s s i ve s et of a g en t i s th e fu n d a m en ta l r es u l t a l l owi n g th e d evel op m en t of th e s y s tem .

6 .7 I n t e r p r e t a t io n o f t h e M o r p h o lo g ic a l O r g a n iz a t io n : T h e E v o c a tio n A g e n ts Th e m or p h ol og i ca l a g en ts p r ovi d e th e s ta b i l i z ed s ta te of th e a s p ectu a l or g a n i ti on th a t cor r es p on d s to a fi x ed p oi n t of th e m i r r or i n g p r oces s . Th e r ea d i n g of e m or p h ol og y , th a t i s th e r ep r es en ta ti on of ca l cu l a ti on s d on e b y a s p ectu a l a g en t g r eg a ti on s , p r ovi d es th e em er g en ce of th e s en s e of th a t h a s b een effecti vel y ca l l a ted wi th th e a s p ectu a l a g en t, wh i l e ta k i n g a ccou n t of m or p h ol og i ca l a g en ts of g a g em en t. Th i s n oti on of em er g en ce h a s a s tr i ctl y or g a n i z a ti on a l ch a r a cter wel l . B u t we won ' t r em a i n a t th e l evel of th e s i m p l e ex p r es s i on of m or p h ol og i ca l a g en ts i n g r ou p s . Th e s y s tem m u s t ta k e a ccou n t of th e s i g n i fi ca n ce of th i s m or p h ol og y , to fea r i t, wou l d b e th a t to m em or i z e i t i n a n or g a n i z a ti on a l wa y , th a t i s to ta k e a ccou n t i m p l i ci tl y i n i ts fu tu r e a cti va ti on , i n i ts fu tu r e en g a g em en ts . Th e s y s tem h a s th a t to b e-to-s a y i t fu n cti on s l i k e a n or g a n i z a ti on a l m em or y . An d a n oth er or g a n i z a ti on of a g en ts , a fter th e a s p ectu a l a g en ts a n d th os e of m or p h ol og y b e g oi n g to ta k e i n con s i d er a ti on th e s ta te of th e l a n d s ca p e of m or p h ol og i ca l a g en ts to a ch i eve a n a n a l y s i s of i ts own m or p h ol og y . It i s a b ou t r ep r es en ti n g th e s en s e of th e a cti va ti on of th e a s p ectu a l a g en t or g a n i z a ti on , fr om i ts ch a r a cter s of a s p ect ex p r es s ed b y m or p h ol og i ca l a g en ts . An or g a n i z a ti on of a g en ts , th e a g en ts of evoca ti on , b e g oi n g to p r ovi d e a cog n i ti ve vi ew of th a t th a t h a s b een ex p r es s ed b y th e g eom etr i c a n d s em a n ti c i n for m a ti on com i n g fr om th e l a n d s ca p e of m or p h ol og i ca l a g en ts , a b ove of th e a s p ectu a l a g en t l a n d s ca p e. Ag en ts of evoca ti on , th a t h a ve a cl a s s i ca l s tr u ctu r e, a r e g oi n g to r ep r es en t ca teg or i es of s i g n i fi ca n ce b etween th e a cti on of th e r ob ot, th e a cti vi ty of i ts i n ter fa ci n g a g en ts , th e com p u ta ti on a l d evel op m en t of i ts b eh a vi or a n d th e r ep r es en ta ti on of th i s d evel op m en t b y m or p h ol og i ca l a g en ts . Th ey ex p r es s th e g l ob a l con s i s ten cy of a cti va ti on wh i l e d oi n g ch oi ces a n d d eci s i on s of g l ob a l b eh a vi or , wh i l e k eep i n g s tr a teg i es of i n h i b i ti on of a cti on for cer ta i n i n ter fa ci n g , a s p ectu a l or m or p h ol og i ca l a g en ts a n d wh i l e con tr ol l i n g s o th e g en er a l l i n e of or g a n i z a ti on a l em er g en ce a ch i eved i n th e s y s tem . L et' s n oti ce th a t th es e s tr a teg i c a cti on s wi l l b e i n d i r ect, i n r el a ti on to ever y a g en t' s b eh a vi or , p er m i tti n g to con s ti tu te a s y s tem wi th em er g en ce of s en s e wi th i ts i n tr i n s i c ch a r a cter s of n on -s ta b i l i ty a n d l ea r n i n g b y s tr u ctu r a l d i s tor ti on on l y . Th e s y s tem i c l oop i s n ow cl os ed . z a th a g cu en

50

A. Ca r d on

6 .8 C o n c lu s io n We h a ve a p p l i ed s u ch a s y s tem for th e m a n a g em en t of cr i s i s s i tu a ti on i n i n d u s tr i a l d i s a s ter s a n d we h a ve d evel op ed a p r ototy p e for s i m u l a ti on of com m u n i ca ti on s TM b etween a ctor s , cod ed i n D i s tr i b u ted Sm a l l ta l k [ 7] . Th e ta s k a t h a n d wa s tr y i n g to b u i l d a n u n d er s ta n d i n g of a d y n a m i c, con fl i cti n g s i tu a ti on , p er cei ved b y th e E va l u a ti on Sy s tem th r ou g h ex ch a n g ed m es s a g es p oten ti a l l y i n coh er en t or con fl i cti n g . To r ea ch th i s d i ffi cu l t g oa l , we h a ve p r op os ed a r ch i tectu r e for a n E va l u a ti on Sy s tem b a s ed on th e m or p h ol og y of th e b eh a vi or of a s p ectu a l a g en t or g a n i z a ti on s . We ca n tr a n s p os e a n d a p p l y th i s m od el to th e d i a l og s b etween ever y vi r tu a l u s er ’ s com m u n i ty . Th i s i s a r es ea r ch p r og r a m wh er e we h a ve to ex p r es s th e on tol og i es a b ou t th e ex ch a n g ed a n d s h a r ed k n owl ed g e u s ed b y u s er s a n d a d a p t th e E va l u a ti on Sy s tem for th e ca s e i n th e en vi r on m en t of ea ch u s er , a b ove h i s u s u a l com m u n i ca ti on a l i n ter fa ce.

R e fe r e n c e s 6.1 Ax el r od R ., Th e com p l ex i ty of Coop er a ti on : Ag en t-b a s ed Mod el of Com p eti ti on a n d Coop er a ti on , P r i n ceton Un i ver s i ty P r es s , 1997. 6.2 Ca r d on A., D u r a n d S., A Mod el of Cr i s i s Ma n a g em en t Sy s tem In cl u d i n g Men ta l R ep r es en ta ti on s , AAAI Sp r i n g Sy m p os i u m , Com m u n i ca ti on p u b l i é e d a n s l es a ctes , Sta n for d Un i ver s i ty , Ca l i for n i e, USA, 23 -26 m a r s 1997. 6.3 Ca r d on , A., L es a g e F ., Towa r d a d a p ti ve i n for m a ti on s y s tem s : con s i d er i n g con cer n a n d i n ten ti on a l i ty , P r occ; KAW' 98 , B a n ff, Ca n a d a . 6.4 Ca r d on A., Con s ci en ce a r ti fi ci el l e et s y s tè m es a d a p ta ti fs , E y r ol l es , P a r i s , 1999. 6.5 L a p i er r e, J.W. , L ’ An a l y s e d es Sy s tem es , Sy r os , 1992. 6.6 L en a t D ., G u h a R .V., B u i l d i n g L a r g e Kn owl ed g e-B a s ed Sy s tem s , R ep r es en ta ti on a n d In fer en ce i n th e Cy c P r oj ect. Ad d i s on Wes l ey P u b l i s h i n g Co. , 1990. 6.7 L es a g e F ., In ter p r é ta ti on a d a p ta ti ve d u d i s cou r s d a n s u n e s i tu a ti on m u l ti p a r ti ci p a n ts : m od é l i s a ti on p a r a g en ts . Th è s e d e l ' Un i ver s i té d u Ha vr e, D é cem b r e 2000. 6.8 Wool d r i d g e M., Jen n i n g s N .R ., Ag en t Th eor i es , Ar ch i tectu r es a n d L a n g u a g es : a Su r vey ; L ectu r es N otes i n A.I., 8 90, Sp r i n g er Ver l a g , 1994 .

7. Public Opinion Channel: A System for Augmenting Social Intelligence of a Community Tomohiro Fukuhara13 , Toyoaki Nishida2 , and Shunsuke Uemura3 1

2 3

Synsophy Project, Communications Research Laboratory, Kyoto 619-0289, Japan E-mail: [email protected] School of Engineering, The University of Tokyo, Tokyo 113-8656, Japan E-mail: [email protected] Graduate School of Information Science, Nara Institute of Science and Technology, Nara 630-0101, Japan E-mail: {tomohi-f, uemura}@is.aist-nara.ac.jp

7.1 Introduction The Internet has became a social place. It allows us to exchange our thoughts and opinions with other people who have similar interests or goals. However, existing communication systems such as e-mail, BBS (Bulletin Board System), chat and instant messaging systems have limitations on eliciting and circulating opinions in a community1 because of communication costs that block talking various opinions between community members. We consider that social intelligence is a property of a community that enables the members to exchange and evolve their implicit knowledge. To augment social intelligence of a community, facilitating elicitation and circulation of hidden opinions of the members by reducing the communication costs are required. We have developed the Public Opinion Channel (POC) prototype system that reduces the communication costs. POC is a concept of an automatic community broadcasting system[7.1][7.2]. POC elicits and circulates community members’ opinions by providing a story to the members. A story is a digest of opinions in the community. Although the members have their opinions, they often hesitate to say their opinions to others. By providing the story to the members, they can easily find implicit opinions including not only major but also minor opinions in their community, and are encouraged to say their opinions. The POC prototype system allows members to listen to the stories as radio program, viewing various opinions passively, and send their opinions as anonymous short messages. 1

community here is a group of peoples who have the same interests and goals, and discussing and working together on the Internet.

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 51− 58 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

52

T. Fukuhara, T. Nishida, and S. Uemura

Table 7.1. Comparison of costs for receiving, sending, selecting a message between an e-mail system and other communication systems. E-mail (baseline) BBS Network news Chat system Instant messaging

Receiving — High High Low Low

Sending — Medium Medium Low Low

Selection — High High Low Low

7.2 Communication Costs The communication costs referred to here are expenses of cognitive resources for receiving, sending, and selecting a message by using communication tools on the Internet. There are three kinds of the communication costs: (1) receiving cost, which is the cost of a user receiving and comprehending a message by using the communication system, (2) sending cost, which is the cost of a user preparing and sending a message, and (3) selection cost, which is the cost of a user selecting a message to read. Table 7.1 compares the communication costs between several communication systems and an e-mail system, such as Eudora2 and Outlook3 that receives and sends only a plain text message. A message referred to here is a unit of information such as an article on BBS or from network news, one or several lines of texts for chat systems and instant messaging tools4 . BBS and network news incur high costs for receiving and selecting a message. This is because a user has to keep track of messages in order to partake in discussions. When BBS and network news are updated, it becomes difficult to follow discussions. Furthermore, selecting messages from a large number of messages from BBS and network news is difficult. A chat system and an instant messaging tool require all costs to be low. This is because these systems treat short messages consisting of one or several lines of text. Thus, a user can receive and comprehend the contents of the message easily and instantly. In effect, they can send their thoughts just like talking by using these systems. From this comparison, a communication system should be designed to meet three requirements: (1) it should allow a user to attend discussions without requiring them to keep track of discussions, and (2) it must help a user to find or select a message they actually wants to read, and (3) it allows a user to send short message. In addition to these requirements, we added the following to the requirements in order to facilitate community members to acquire stories and 2 3 4

http://www.eudora.com/ http://www.microsoft.com/office/outlook/default.htm including Yahoo Messenger and AOL Instant Messenger

7. Public Opinion Channel

53

Fig. 7.1. An overview of POC prototype system.

encouraging them to talk their opinions: (4) anonymous messaging, which allows community members to send their opinions without revealing their personal information such as their names, (5) passive viewing of opinions, which enables the members to view opinions without any operations, and (6) continuous broadcasting, in which a POC broadcasts stories at all times by generating new stories or rebroadcasting existing stories.

7.3 POC Prototype System The POC prototype system consists of a POC server (community broadcasting server) and several POC clients. Figure 7.1 shows an overview of the POC prototype system. A POC server is a broadcasting system that provides (1) opinions for supporting discussions between community members, and (2) stories for notifying picked out opinions to the members. A POC client is a tool for (1) listening to stories, which are provided as radio program by the POC server, and for (2) exchanging opinions between the members for discussion. In this section, we describe the story broadcasting function of the POC server, and the discussion support function of the POC client. 7.3.1 POC Server The POC server has two roles: (1) discussion server, which provides opinions to the POC clients for facilitating discussions between community members, and (2) broadcasting server, which generates and broadcasts stories as radio program. We describe the latter function in this subsection.

54

T. Fukuhara, T. Nishida, and S. Uemura

Table 7.2. Example of a story. DJ Opinion 1 DJ Opinion 2 DJ Opinion 3 DJ

Next opinion is “affordance”. Does anyone know about affordance? Related to this opinion, there is another opinion. There is a workshop on designing intelligent artifacts. This is a good introduction to affordance. This is the last opinion. I found a good page on cognitive psychology when I was searching affordance. Thanks all. We’re waiting for your opinions.

Generating stories. The POC server generates a story that has a context. Context here means the semantic relationship between sentences. The context is made by linking related opinions and is generated as follows. 1. Pick out an opinion (source opinion) from an opinion database. 2. Retrieve opinions from the opinion database by using title of the first opinion. 3. Sort retrieval results by date order, and first n opinions are added to the source opinion. (n is threshold) An example of a story is shown in Table 7.2. In this example, a DJ who plays a role of a disc jockey in a radio program introduces three opinions related to “affordance”. These opinions are sorted by date. Broadcasting stories. The POC server broadcasts stories as radio programs on the Internet. This is done by MP3 audio stream. The POC server generates audio files by using a text-to-speech system (TTS), and broadcasts them via MP3 streaming server. The POC server uses CHATR5 for TTS, and icecast6 for the MP3 streaming server. A user can listen to the stories via MP3 players such as WinAmp7 . We regard MP3 players as the POC client for listening to the stories. 7.3.2 POC Client: POCViewer In this subsection, we describe the discussion support function of the POC client. We have developed an implementation of POC client named POCViewer that supports exchanging opinions between community members. With the POCViewer, users can view opinions passively, and compose and send their opinions to the POC discussion server. Figure 7.2 shows an image of the POCViewer. POCViewer shows opinions in the Telop style, i.e., each character of a story appears one by one. The POCViewer has several functions for facilitating the discussions. 5 6 7

http://results.atr.co.jp/products e/frame9.html http://www.icecast.org/ http://www.winamp.com/

7. Public Opinion Channel

55

Fig. 7.2. A screen image of the POCViewer.

Table 7.3. An example of a message.  ?xml version=”1.0” encoding=”Shift JIS” ?  opinion name=”tem imf” date=”2001/5/6 20:52:18” host=”192.168.31.163” reference=”comment5.xml”  title ATM service in Japan /title  comment I think ATM services in Japan are inconvenient. Banks should run their ATMs for 24 hours. /comment  url http://www.japanese-online.com/language/bank.html/url /opinion

Opinion composer. A user can compose, edit, and send their opinion to the POC server. The user can save their opinion as a local opinion, which is stored in the local hard disc, and modify or browser it later. An example of an opinion is shown in Table 7.3. The opinion consists of a title, a comment, and a reference URL. When the user sends he opinion, she inputs title and comment from the POC client. The POC client inserts XML tags to the opinion, and sends it to the server. Local mode and network mode. A user can select the mode of the POCViewer as either local mode or network mode. In local mode, the user can compose and store their opinions into local hard disk. In network mode, the user can not only send their opinion but also view and capture opinions of their community. Local mode is suitable for composing and viewing personal opinions. By separating the local and network modes, the user can store their

56

T. Fukuhara, T. Nishida, and S. Uemura

Fig. 7.3. Continuous retrieval. Retrieval is made by extracting keywords from previous retrieval results, and retrieving them continuously.

tentative opinions on their local hard disc, and send the mature opinions to the server. Capturing opinions. A user can capture opinions on the local hard disk. The user can view captured opinions in local mode. And they can also edit and modify the captured opinions, and send them to a POC server. Opinion retrieval. A user can retrieve opinions in network mode. The actual retrieval process is run on the POC server. The POC server uses the n-gram search method which searches messages according to pieces of queries consisting of one or two characters[7.3]. This method has the advantage that various texts that include queries partially are retrieved. Thus, the user can view various stories. Continuous retrieval. A user can view set of similar opinions continuously. The POCViewer can retrieve opinions continuously. When a user retrieves via a keyword, the POCViewer gets another keyword from the retrieval results, and retrieves a set of opinions by using that keyword. Figure 7.3.2 shows an overview of the continuous retrieval. The user can view a set of opinions based on the retrieval results. In Figure 7.3.2, opinions related to a keyword “Agent” are retrieved. When continuous retrieval mode is off, further retrievals are not perform. When continuous retrieval mode is on, further retrievals based on previous retrieval are performed. The retrievals are performed by extracting a keyword from previous retrieval results. The keyword is picked out according to the feature value of a word. In the implementation, we use the frequency of word as the feature value. Retrievals continues according to previous retrieval results. The user can view another opinions originating from initial keyword given by the user.

7. Public Opinion Channel

57

7.4 Evaluation We performed two preliminary experiments of applying the POC prototype system to a practical community. One is a long term observation of opinions in a group[7.4], and the other is a short term observation in a group thinking situation. The first was on the evaluation of exchanging implicit opinions in a group. This experiment was made for three months. The group consists of eight members, all Japanese, and each member is familiar to the others. 1,329 opinions were collected during this experiment. The members exchanged their opinions about various including not only their business but also movies and TV programs. Some opinions are referring to other members’ opinions, and the others are monologues. Although the members post many opinions, we found a point that discussions did not last for a long time. We consider the reason is that the members had got used to the “couch potato” style of viewing of the opinions because the POCViewer shows the opinions automatically. To facilitate discussions in the POCViewer is our future work. The second was on the evaluation of creativity support by POC[7.5]. Miura argued that POC enabled group members to find an opinion to which they have not paid attention. In this experiment, members discussed on demands or requests from their university using the POC system. The POC server broadcasts opinions in order to provide various viewpoints for the members periodically. In this experiment, circulating opinions enabled members to recognize importance of previous opinions. We will continue evaluation of creativity support by POC.

7.5 Discussion 7.5.1 Automatic Broadcasting System Tanaka et al. proposed information visualization tools using a TV program metaphor[7.6]. By using these visualization tools, the user can view Web documents or retrieval results from a database in passive viewing style like viewing a TV program. One of major differences between POC and the information visualization tools is the source of the story. We treat community members’ opinions as the source. This is different in story generation method from the visualization tools because identifying minor opinion from major ones is required. In the concept of POC, POC takes up not only major opinions but also minor ones. This requirement is inevitable for fair discussions in a community. Although we have not implemented this function yet, we consider it is important to find minor opinions for the automatic broadcasting system for a community.

58

T. Fukuhara, T. Nishida, and S. Uemura

7.5.2 POC and Narrative Intelligence Lawrence et al. proposed to use storytelling to exchange knowledge in a group[7.7]. They argued that there is a function for collecting and sharing knowledge in storytelling. One of points of POC in regard to narrative intelligence is that opinions in the POCViewer become seeds of narratives. In the long term experiment, we found that several opinions becomes the seeds of narratives, i.e., community members replied to the opinions by adding their thoughts or memories related to them. However, what kind of opinions are suitable for seeds of narratives that cause further replies. To analyze this kind of opinions is future work.

7.6 Conclusion We have developed a POC prototype system for eliciting and circulating opinions in a community. The system augments social intelligence by reducing the communication costs. From the experiments, we found availabilities of the POC prototype system on (1) eliciting and circulating various implicit opinions in a community, and (2) creativity support in a community.

References 7.1 Nishida, T. et al.: “Public Opinion Channel for Communities in the Information Age”; New Generation Computing, Vol. 14, No. 4, pp. 417-427(1999). 7.2 Azechi, S. et al: “Public Opinion Channel: A challenge for interactive community broadcasting”; Digital Cities: Experiences, Technologies and Future Perspectives, Lecture Notes in Computer Science, 1765, Springer-Verlag, pp. 427441(2000). 7.3 Sato, S & Kawase, T.: “A high-speed best match retrieval method for Japanese text”; Japan Advanced Institute of Science and Technology (JAIST) Research Report, IS-RR-94-9I(1994). 7.4 Fujihara, N.: “How to evaluate social intelligence design”; Workshop notes on JSAI-Synsophy International Workshop on Social Intelligence Design,(2001). (http://www.synsophy.go.jp/sid2001/papers/preprints/fujihara sid200111.pdf) 7.5 Miura, A. et al.: “Support for idea creation in groups: A social psychological approach”; Proc. of the 15th Annual Conference of JSAI (JSAI2001), 2E2-06, (2001), [in Japanese]. 7.6 Tanaka, K. et al.: “Back to the TV: Information visualization interfaces based on TV-program metaphors”; Proc. of IEEE International Conference on Multimedia and Expo2000(ICME2000), New York, pp. 1229-1232(2000). 7.7 Lawrence, D. et al.: “Social dynamics of storytelling: Implications for story-base design”; AAAI 1999 Fall Symposium on Narrative Intelligence, Massachusetts, (1999). (http://www.cs.cmu.edu/afs/cs/user/michaelm/www/nidocs/ LawrenceThomas.pdf)

8 . E n a b lin g P u b lic D is c o u r s e Kei i ch i N a k a ta In s ti tu te of E n vi r on m en ta l Stu d i es , G r a d u a te Sch ool of F r on ti er Sci en ces , Th e Un i ver s i ty of Tok y o, 7-3 -1 Hon g o, B u n k y o-k u , Tok y o 113 -003 3 Ja p a n

8 .1 I n t r o d u c t io n In cr ea s i n g con cer n s for en vi r on m en ta l p r ob l em s h a ve con tr i b u ted to th e g en er a l a wa r en es s r eg a r d i n g th e i m p or ta n ce a n d d i ffi cu l ty of en g a g i n g a r a n g e of s ta k eh ol d er s i n th e d eci s i on -m a k i n g p r oces s . Th os e wh o a r e i n vol ved i n m a k i n g d eci s i on s , a n d a ffected b y th e d eci s i on s m a d e, fr om th e a u th or i ti es to th e m em b er s of p u b l i c, s h ou l d b e offer ed a n op p or tu n i ty to en g a g e i n i n for m ed d el i b er a ti on , i n wh i ch vi ews fr om va r i ou s p er s p ecti ves a r e r a i s ed , ex a m i n ed , d i s cu s s ed a n d ta k en i n to con s i d er a ti on b efor e a n a ttem p t to r ea ch a con s en s u s i s m a d e. Ou r u l ti m a te g oa l i s to m a k e s u ch a p r oces s of p u b l i c d el i b er a ti on , i .e., a “ for m a l or i n for m a l p r oces s for com m u n i ca ti on a n d for r a i s i n g a n d col l ecti vel y con s i d er i n g i s s u es ” [ 1] , a s effecti ve a n d m ea n i n g fu l a s p os s i b l e b y s u p p or ti n g th e com m u n i ty of s ta k eh ol d er s wi th i n for m a ti on tool s [ 2] . Th i s ta k es i n to a ccou n t th e i n cr ea s i n g ca s es of i n i ti a ti ves fr om n a ti on a l a n d l oca l a u th or i ti es over th e wor l d to m a k e en vi r on m en ta l d a ta a n d i n for m a ti on el ectr on i ca l l y a va i l a b l e a n d th e ever i n cr ea s i n g p op u l a r a cces s to th e In ter n et, wh i ch offer a p oten ti a l for n ew k n owl ed g e to em er g e th r ou g h i n ter a cti on s of p eop l e over th e n etwor k [ 3 ] . In th i s con tex t, we vi ew “ s oci a l i n tel l i g en ce” a s th e p oten ti a l ca p a b i l i ty of a com m u n i ty to en g a g e i n i n for m ed d el i b er a ti ve d eci s i on -m a k i n g p r oces s , a n d th e “ tr a ces ” l eft b eh i n d s u ch a cl a s s of coop er a ti ve a cti vi ty .1 Th e i m p or ta n ce of b r i n g i n g tog eth er th e wi d e s p ectr u m of con cer n ed g r ou p s a n d i n d i vi d u a l s i n r ea l i z i n g a s u s ta i n a b l e s oci ety ca n n ot b e u n d er s ta ted . Th e n eces s i ty of p a r tn er s h i p b etween con cer n ed m em b er s of th e com m u n i ty , s u ch a s i n d i vi d u a l r es i d en ts , p ol i cy m a k er s , i n d u s tr y a n d N P Os , a n d th e es ta b l i s h m en t of com m u n i ca ti on ch a n n el s for s u ch a col l a b or a ti ve en ter p r i s e a r e r ep ea ted l y em p h a s i z ed [ 4 ] . Mer e d i s s em i n a ti on of i n for m a ti on , h owever i t m a y b e d es i g n ed to ca ter for th e p r es u m ed i n ter es ts of oth er p a r ti es , i s n ot en ou g h : th er e m u s t b e a p l a ce for a d i a l og u e b a s ed on th e i n for m a ti on m a d e a va i l a b l e. Su ch a vi s i on of p a r tn er s h i p i n s oci a l d eci s i on -m a k i n g a m on g va r i ou s coh or ts wi th i n a com m u n i ty i s , h owever , n ot s i m p l e to a ch i eve. In r ea l i ty , th er e a r e n u m b er of fa ctor s th a t wou l d h a m p er a n d p r even t i ts effecti ve i m p l em en ta ti on . Am on g th em i s th e d i ffi cu l ty for p u b l i c to p a r ti ci p a te i n d i s cu s s i on s d u e to th e p r es u m ed l a ck of con fi d en ce con cer n i n g tech n i ca l i s s u es th a t m i g h t a r i s e. Th er e cou l d b e a b r ea k d own i n com m u n i ca ti on d u e to j a r g on s a n d tech n i ca l ter m i n ol og y . 1

Th i s i s b y n o m ea n s m ea n t to b e th e d efi n i ti on of “ s oci a l i n tel l i g en ce” , b u t a n ex a m p l e of ci r cu m s ta n ces u n d er wh i ch i t wou l d m a n i fes t i ts el f.

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 59-66, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

60

K. N a k a ta

At th e s a m e ti m e, wh i l e i n for m a ti on con cer n i n g p u b l i c p ol i cy cou l d b e d i s s em i n a ted , i t i s n ot th e ca s e th e oth er wa y a r ou n d ; i t i s d i ffi cu l t for a u th or i ti es , to ob ta i n feed b a ck fr om m em b er s of th e com m u n i ty , a n d wh en th ey d o, th a t m i g h t ca r r y a h os ti l e ton e. We b el i eve th a t s om e a s p ects of th e p r ob l em s r a i s ed a b ove ca n b e ta ck l ed b y p r ovi d i n g s u p p or t for m em b er s of th e com m u n i ty to h a ve th ei r voi ces h ea r d m or e effecti vel y . In ou r on g oi n g r es ea r ch , we a ttem p t to a d d r es s th i s i s s u e b y en a b l i n g i n d i vi d u a l s i n i n for m ed d i s cu s s i on s th r ou g h th e p r ovi s i on of s u p p or ti n g tool s a i m ed to en h a n ce p u b l i c d i s cou r s e. Ou r cu r r en t r es ea r ch effor t i s th e d evel op m en t of a n etwor k -b a s ed d i s cu s s i on s y s tem th a t i s a i m ed to s u p p or t p u b l i c d el i b er a ti on on en vi r on m en ta l i s s u es [ 2] . In d el i b er a ti on , i t i s a s s u m ed th a t th er e a r e b oth con s en s u a l a n d a d ver s a r i a l p r oces s es , s i n ce th e p a r ti ci p a tor y col l ecti ve wou l d often i n cl u d e th os e wi th op p os i n g a g en d a a n d d i ffer en t va l u es . It i s n ot m er el y a d i s cu s s i on for u m ; i t s h ou l d b e a n i ter a ti ve p r oces s i n wh i ch th e over a l l a i m i s ei th er or b oth to r ea ch a con s en s u s (c l o s u r e ), a n d /or to i n cr ea s e p a r ti ci p a n ts ’ u n d er s ta n d i n g of th e i s s u es r a i s ed a n d d i ffer en t p os i ti on s a s s u m ed b y oth er p a r ti ci p a n ts . Th u s , d el i b er a ti ve p r oces s es a r e often s een a s es s en ti a l i n m a k i n g i n for m ed , col l ecti ve d eci s i on s . Th e i n ten d ed u s e of th e s y s tem i s p r i m a r i l y for a s y n ch r on ou s d i s cu s s i on s i n l i m i ted d om a i n s for wh i ch s i m u l a ti on -b a s ed a n a l y ti ca l tool s a r e a va i l a b l e. Th e fea tu r e th a t i s r eq u i r ed i n ter m s of i n ter a cti on d es i g n i s th e fa ci l i ty for th e p a r ti ci p a n ts to con tr i b u te to th e d i s cu s s i on wi th ou t too m u ch over h ea d a n d p s y ch ol og i ca l b a r r i er s th a t h a m p er th e r ep r es en ta ti on of n on -tech n i ca l or n ovi ce vi ews . In th e p a s t, s u ch a tech n i ca l d i vi d e h a s often l ed to th e tota l b r ea k d own of ex ch a n g e of vi ews b etween ex p er ts a n d n on -ex p er ts i n th e en vi r on m en ta l for u m , r es u l ti n g i n ty p i ca l s ta n d off s i tu a ti on s b etween th e two a n ta g on i s ti c ca m p s . In s tea d , i n or d er to p r om ote p u b l i c d el i b er a ti on a n d s h a r i n g of th e r es p on s i b i l i ty of col l ecti ve d eci s i on -m a k i n g b y th e com m u n i ty , voi ces s h ou l d b e h ea r d a n d a r g u m en ts s h ou l d b e u n d er s tood .

8 .2 E n a b lin g I n d iv id u a ls t o C o lle c t a n d E x c h a n g e I n f o r m a t io n a n d O p in io n s R ecen t i n ter es ts i n com m u n i ty -or i en ted (i n tel l i g en t) i n for m a ti on s y s tem s p r oj ects g r ou p ed u n d er l a b el s s u ch a s “ com m u n i ty com p u ti n g ” a n d “ com m u n i ty wa r e” h i g h l i g h t th e focu s on com m u n i ti es wi th th e a i m of s u p p or ti n g th ei r for m a ti on a n d th ei r a cti vi ti es . D es p i te th e em p h a s i s on th e “ com m u n i ty ” , we ob s er ve th a t a n es s en ti a l el em en t of th es e com m u n i ty -or i en ted s y s tem s i s th e en h a n cem en t of i n ter a cti ve ca p a b i l i ti es of i n d i vi d u a l m em b er s of a com m u n i ty . Th i s i s a n a tu r a l con s eq u en ce s i n ce th e com m u n i ty a cti vi ti es a r e often d ecen tr a l i z ed a n d b ottom -u p , wi th th e s tr en g th b ei n g th e ca p a b i l i ty to g en er a te em er g en t s ol u ti on s to i l l s tr u ctu r ed p r ob l em s for wh i ch i n d i vi d u a l p a r ti ci p a ti on i s es s en ti a l . We s ee th e p oten ti a l i n th e com m u n i ty -or i en ted i n fr a s tr u ctu r e s u ch a s N i s h i d a ’ s “ P u b l i c Op i n i on Ch a n n el ” (P OC) [ 5] a s a veh i cl e for i n cr ea s i n g s oci a l a wa r en es s i n ter m s of i n for m a ti on ex ch a n g e a n d p er s p ecti ve s h a r i n g . F or s u ch p u r p os es , a con ver s a ti on a l i n ter fa ce s u ch a s E g oCh a t [ 6] ca n b e con s i d er ed to b e a n a tu r a l

8 . E n a b l i n g P u b l i c D i s cou r s e

61

ca n d i d a te for p eop l e to i n ter a ct th r ou g h a p l a tfor m s u ch a s P OC. In E g oCh a t, con ver s a ti on s b etween p er s on a l a g en ts a r e g en er a ted a n d s u s ta i n ed th r ou g h k ey wor d m a tch i n g i n a con ver s a ti on a l d a ta b a s e (“ con ver s a ti on b a s e” ) th a t s tor es con ver s a ti on a l fr a g m en ts (u tter a n ces ) of th e u s er s th es e a g en ts r ep r es en t. A h u m a n u s er ob s er ves th e con ver s a ti on th a t u n fol d s b etween th es e a g en ts a n d ca n i n ter r u p t th e con ver s a ti on a n d “ ta l k ” to th e a g en ts th er eb y i n cr ea s i n g th e d a ta i n th e con ver s a ti on b a s e of h i s p er s on a l a g en t. Th e con tex t of th e u tter a n ce i s d eci d ed b y th e top i c th ey (i .e., th e u s er a n d th e a g en ts ) a r e “ ta l k i n g ” a b ou t. Th e s i g n i fi ca n ce of s u ch a s y s tem i s th a t i t u s es th e ever y d a y for m of i n for m a ti on ex ch a n g e i n com m u n i ti es , vi z . con ver s a ti on s , a s th e m ea n s to el i ci t i n for m a ti on fr om th e h u m a n p a r ti ci p a n t. Th i s ta k es a d va n ta g e of th e n a tu r e of h u m a n con ver s a ti on s u g g es ted b y Sch a n k th a t h u m a n s d o n ot n eces s a r i l y cr ea te n ew k n owl ed g e th r ou g h con ver s a ti on s b u t p r es en t wh a t th ey h a ve a l r ea d y th ou g h t a b ou t, r efor m u l a ted i n th e for m a p p r op r i a te for th e con ver s a ti on [ 7] . Con ver s a ti on b a s es h el d b y ea ch p er s on a l a g en t for m em b er s of a com m u n i ty tog eth er s tor e a r i ch s ou r ce of op i n i on s a n d i n for m a ti on h el d i n th a t com m u n i ty . However , th e con ver s a ti on a l fr a g m en ts s tor ed a s tex t l os es th e n on -ver b a l i n for m a ti on s u ch a s s h a r ed vi s u a l i n for m a ti on a n d g es tu r es th a t a r e p r es en t i n or d i n a r y fa ce-to-fa ce con ver s a ti on s . In ter a cti on s wi th s u ch a con ver s a ti on b a s e s h ou l d i d ea l l y b e m u l ti -m od a l , i n cor p or a ti n g vi s u a l a n d a u d i o d a ta to a ccom p a n y i n for m a ti on b r oa d ca s t. We a r e cu r r en tl y ex p er i m en ti n g th e u s e of wea r a b l e d evi ces for i n for m a ti on g a th er i n g , wi th th e p os s i b l e effect of g r ou n d i n g i n for m a ti on to en a b l e i n d i vi d u a l s to s h a r e wi th th e com m u n i ty wh a t th ey s a w a n d h ea r d , wh i ch wou l d p r ovi d e a fi r m er con tex t i n wh i ch on e’ s op i n i on s wer e r a i s ed . F i g u r e 1 i l l u s tr a tes a con ver s a ti on a l a g en t i n ter fa ce m od el ed a fter E g oCh a t, wh i ch s ta g es con ver s a ti on s b etween u s er s ’ p er s on a l a g en ts . In th i s d em on s tr a tor , a n en tr y i n th e con ver s a ti on b a s e con ta i n s n ot on l y th e tex t of a n u tter a n ce b u t a l s o th e a ccom p a n y i n g vi s u a l i n for m a ti on a t th e ti m e i t wa s m a d e, ta k en th r ou g h a h ea d -m ou n t ca m er a a n d s tor ed i n a wea r a b l e P C, a n d th e g es tu r e i n for m a ti on th a t i s ca p tu r ed th r ou g h a m oti on ca p tu r e d evi ce wi th s en s or s a tta ch ed th e s p ea k er ’ s a r m s a n d i n ter p r eted b y a s i m p l e g es tu r e r ecog n i ti on s y s tem . On ce u p l oa d ed to th e con ver s a ti on s er ver , vi s u a l a n d g es tu r e i n for m a ti on i s s h own i n a ccor d a n ce wi th ver b a l s p eech . Ou r i n i ti a l ex p er i en ce wi th th e s y s tem s u g g es ts th a t, i f n on ver b a l i n for m a ti on i s s een a s th e a u g m en ta ti on of ver b a l i n for m a ti on , i .e., a s m od i fi er s of k ey wor d s , s u ch a m u l ti -m od a l con ver s a ti on b a s e i s u s efu l wh en a n u tter a n ce i n vol ves th e u s e of i n d i ca ti ves s u ch a s “ s u ch ” a n d “ l i k e th i s ” , s i n ce s u ch i n for m a ti on i s s om eti m es n ot ea s y to el a b or a te i n wor d s . Wh i l e th e u s e of m ob i l e d evi ces m a y s eem r a th er cu m b er s om e a t th i s p oi n t i n ti m e, we b el i eve th a t i m a g e ca p tu r e a n d tr a n s m i s s i on wou l d b e i n th e n ea r fu tu r e a s or d i n a r y a s voi ce tr a n s m i s s i on . In for m a ti on a b ou t th e com m u n i ty , s u ch a s tr a ffi c s i tu a ti on a n d r i s i n g wa ter l evel s i n n ea r b y s tr ea m s ca n b e col l ected b a ck ed u p wi th i m a g es . Us i n g com m u n i ty m em b er s a n d th ei r ever y d a y a wa r en es s a b ou t th e en vi r on m en t offer th e p os s i b i l i ty of col l ecti n g en vi r on m en ta l i n for m a ti on b i a s ed towa r d s con cer n s of th e com m u n i ty m em b er s .

62

K. N a k a ta

F i g . 1 . P r ovi d i n g m u l ti -m od a l i n for m a ti on for th e con ver s a ti on b a s e

8 .3 R a is in g S o c ia l A w a r e n e s s t h r o u g h P o s it io n - O r ie n t e d D is c u s s io n s Wh en a g r ou p of con cer n ed i n d i vi d u a l s s h a r e th e over a l l g oa l of r ea ch i n g s om e for m of con s en s u s on a n i s s u e th r ou g h d el i b er a ti on , d i s cu s s i on s ca n b e s een a s coop er a ti ve wor k . D i s cu s s i on s , h owever , con ta i n on l y a wea k r ep r es en ta ti on of i ts com m on fi el d of wor k . In oth er wor d s , th e ob j ect of coop er a ti on , i n th i s ca s e d el i b er a ti on a n d p os s i b l y con s en s u s for m a ti on , i s often p oor l y r ep r es en ted a n d n ot ea s y for th e coop er a ti ve en s em b l e (i .e., th e p a r ti ci p a n ts ) to m on i tor i ts p r og r es s . Ta k e for ex a m p l e, th r ea d ed d i s cu s s i on s , wh i ch a r e on e of th e m os t com m on for m s of el ectr on i c b u l l eti n b oa r d s y s tem s (B B Ss ), a n d i n cor p or a ted i n s om e g r ou p wa r e a p p l i ca ti on s a s th e i s s u e-b a s ed i n for m a ti on s y s tem s (IB IS) s ty l e of d i s cu s s i on th r ea d s . Wh en th e d i s cu s s i on i s r el a ti vel y s m a l l , a u s er ca n ea s i l y m on i tor wh a t i s g oi n g on b y s k i m m i n g th r ou g h th e con ten ts of con tr i b u ti on s . As i t g r ows , u n l es s th e u s er i s a ver y a cti ve p a r ti ci p a n t of th e d i s cu s s i on , i t wi l l n ot on l y b e d i ffi cu l t for h er to m on i tor th e d evel op m en t of a r g u m en ta ti on a n d s u b -top i cs , b u t a l s o to p a r ti ci p a te i n i t. In com p u ter -m ed i a ted d i s cu s s i on s , th e vi s u a l i z a ti on of a r g u m en ta ti on a d d r es s es th i s i s s u e, a n d s y s tem s s u ch a s Con k l i n ’ s g I B I S [ 9] h a ve b een p r op os ed . In th e a r ea of s ch ol a r l y d i s cou r s e, col l a b or a ti ve a r g u m en ta ti on [ 8 ] i s

8 . E n a b l i n g P u b l i c D i s cou r s e

p r op os p r oa ch “ l ou d ” fi cu l ty

ed offer i n g es i s a d eq u voi ces a n d i n a s s es s i n g

h y p er tex t-b a s ed s ol u ti on s . a te i n d ea l i n g wi th p oten ti a l effects of i n a cti ve b u t es s en ti a h ow on e i s u n d er s tood b y oth

However p r ob l em s l p a r ti ci p a er p a r ti ci p

63

, n ei th er of th es e a p s u ch a s d om i n a n ce of n ts (“ l u r k er s ” ), a n d d i fa n ts .

8 .3 .1 P o s it io n in g - O r ie n t e d D is c u s s io n I n t e r f a c e We a d d r es s th es e p r ob i n ter fa ce th a t en cou r a g op i n i on s p a ce, th er eb y s h ows a n ex a m p l e of th

−

l em es p vi s u e op

s th r ou a r ti ci p a a liz in g i n i on s

g h a for m of i n ter a cti on i n vol vi n g a g r a p h i ca l n ts to d i r ectl y m a n i p u l a te th ei r p os i ti on s i n th e p a r ti ci p a n ts ’ p os i ti on s i n a d i s cu s s i on . F i g u r e 2 p a ce:

−

It d es cr i b es a two-d i m en s i on a l s p a ce (th e “ b oa r d ” ) wi th h or i z on ta l a n d ver ti ca l a x es r ep r es en ti n g two of th e fa ctor s (i s s u es ) i n th e d i s cu s s i on . We b el i eve th a t or d i n a r y u s er s wou l d n ot b e a b l e to cop e wi th m or e th a n two d i m en s i on s , es p eci a l l y wh en i t com es to p os i ti on i n g th em s el ves i n th e op i n i on s p a ce. E a ch p a r ti ci p a n t i s a s s i g n ed a n i con , wh i ch ca n b e a p i ece wi th a d es i g n a ted col ou r , or h er own i m a g e s u ch a s a p h oto or ca r toon . To m a k e a con tr i b u ti on , th e u s er “ m oves ” h er p i ece to th e p os i ti on i n th e b oa r d s h e th i n k s th a t d es cr i b es h er s ta n ce i n th e d i s cu s s i on , a n d ty p es i n th e a r g u m en t or j u s ti fi ca ti on for h er m ove. N a tu r a l l y , s h e d oes n ot h a ve to a ctu a l l y “ m ove” — s h e ca n r em a i n i n th e s a m e p os i ti on a n d con tr i b u te h er op i n i on s . In a d d i ti on to th e g r a p h i ca l i n ter fa ce, ea ch con tr i b u ti on i s l i s ted i n a ta b l e, a l on g wi th i n for m a ti on a b ou t th e con tr i b u tor , d i r ecti on of m ove, a n d a ti m e s ta m p . L a b el s of ea ch a x i s th a t d efi n e th e op i n i on s p a ce a r e ch a n g ed a ccor d i n g to th e d evel op m en t of th e d i s cu s s i on . On ce th e l a b el s a r e ch a n g ed , th e p os i ti on s a r e r es et to th e n eu tr a l p os i ti on (i .e., th e or i g i n ).

−

− −

In th i s wa y , u s er s ca n p os i ti on th em s el ves wi th r es p ect to th ei r p er cep ti on of oth er p a r ti ci p a n ts ’ p os i ti on s , r evea l i n g h ow con tr i b u ti on s a r e p er cei ved a n d i n ter p r eted a m on g p a r ti ci p a n ts . In ou r p r el i m i n a r y ex p er i m en t u s i n g th i s i n ter fa ce, a m on g th e com m en ts we r ecei ved a fter th e s es s i on i n cl u d ed th e cl a r i ty of m u tu a l p os i ti on s con cer n i n g i s s u es wi th r es p ect to r el a ti ve p os i ti on s wi th oth er p a r ti ci p a n ts , a n d th e effect of i n ter fa ce for focu s i n g on i s s u es wi th ou t d i ver g i n g too m u ch . Som e of th e effects of vi s u a l i z i n g p os i ti on s we i d en ti fi ed wer e a s fol l ows :

− − −

B y “ p l a y i n g b a ck ” th e ch a n g es i n th e b oa r d , th e p a r ti ci p a n ts wer e a b l e to r eca l l th e fl ow of d i s cu s s i on a n d h ow i t u n fol d ed . P a r ti ci p a n ts s eem ed to h a ve r eta i n ed i n for m a ti on a s to h ow oth er p a r ti ci p a n ts ch a n g ed th ei r op i n i on s . B y ob s er vi n g th e ch a n g e i n op i n i on , es p eci a l l y a t “ cr os s i n g th e a x i s ” , we ca n a n a l y s e wh a t m a d e th e p a r ti ci p a n t ch a n g e th ei r vi ews a n d i ts j u s ti fi ca ti on s .

Mor eover , i t p r ovi d es vi s u a l i n for m a ti on a s to h ow d i ver s e ex i s ti n g op i n i on s a r e a n d h ow th e d i s cu s s i on h a s con tr i b u ted to p a r ti ci p a n ts cl os i n g i n (or g r owi n g fa r th er a p a r t) on i s s u es b ei n g d i s cu s s ed .

64

K. N a k a ta

F i g . 2 . D i s cu s s i on i n ter fa ce th r ou g h p os i ti on i n g

Wh en s een a s th e ex p l i ci t i n d i ca ti on of on e’ s p r efer en ces , p os i ti on i n g i n th e op i n i on s p a ce ca n b e s een a s a for m of i n f o r m a l voti n g , a n d th e d i s tr i b u ti on of p i eces a s th e ta l l y of th e vote. Wh i l e ta k i n g a for m a l vote i s often a voi d ed for th e fea r of m a k i n g p r em a tu r e d eci s i on s , ta k i n g i n for m a l votes i s con s i d er ed to b e u s efu l i n d i s cu s s i on s [ 10] . Th er efor e, we ex p ect th e a d va n ta g es (a n d d i s a d va n ta g es ) of ta k i n g i n for m a l votes d u r i n g d i s cu s s i on s to b e i n h er i ted i n th i s i n ter a cti on . On e of th e i m p or ta n t i s s u es i n th e p os i ti on -or i en ted d i s cu s s i on i n ter fa ce i s th e ch oi ce of l a b el s for ea ch a x i s . Th e s el ecti on of th es e l a b el s en ta i l s th e g en er a ti on of op i n i on s p a ce, a n d th i s i n i ts el f i s often th e p oi n t of con tr over s y . To a d d r es s th i s p r ob l em , a h i er a r ch i ca l i s s u e s tr u ctu r e m a y b e cr ea ted , a s a n i n i ti a l r oa d m a p for th e d i s cu s s i on , fr om wh i ch l a b el s for th e a x es ca n b e s el ected b a s ed on a n i s s u e a n d on e of i ts s u b -i s s u es , r ecor d i n g th e ou tcom es a s th e d i s cu s s i on p r oceed s . Th e d i s cu s s i on s y s tem i ts el f i s d es i g n ed to i n cl u d e fea tu r es s u ch a s top i c ex tr a cti on , p a r ti ci p a n t cl u s ter i n g , p a r ti ci p a ti on i n d u cti on , a n d l i n k s to com m u n i ty i n for m a ti on . We b el i eve th es e fea tu r es wou l d en h a n ce th e a cces s i b i l i ty to d i s cu s s i on s wh en th ey g r ow l a r g e. In con tr a s t to th e IB IS-fa m i l y of d i s cu s s i on s y s tem s , th i s i n ter fa ce g u a r a n tees th e s i m p l e s n a p s h ot of th e s ta te of d i s cu s s i on s i n ter m s of op i n i on s p a ce, r a th er th a n a n ever -ex p a n d i n g l i s t of tex t or tr ee.

8 .4 T o w a r d s “ S o c ia l I n t e llig e n c e D e s ig n ” Th e th em e of th i s p a p er i s s u p p or ti n g p u b l i c to p a r ti ci p a te i n d i s cou r s e con cer n i n g en vi r on m en ta l i s s u es for th e a ch i evem en t of a s u s ta i n a b l e com m u n i ty . Th e u n d er l i n i n g a s s u m p ti on i s th a t m em b er s of th e com m u n i ty a r e m oti va ted a n d en cou r a g ed to d o s o, b u t m a y l a ck th e m ea n s a n d op p or tu n i ti es — h en ce th e d evel op m en t of s y s tem s s u p p or t for en a b l i n g p u b l i c d i s cou r s e. Wh en we ob s er ve th e ex i s ten ce of n u m er ou s d i s cu s s i on g r ou p s a n d m a i l i n g l i s ts i n th e In ter n et, i t m i g h t a p p ea r th a t p eop l e a l r ea d y d o h a ve th e m ea n s a n d op p or tu n i ty to ex p r es s th ei r

8 . E n a b l i n g P u b l i c D i s cou r s e

65

vi ews a n d ca r r y ou t d i s cu s s i on s . However , i f we a r e a i m i n g to s u p p or t d el i b er a ti ve p r oces s es for com m u n i ty -b a s ed d eci s i on -m a k i n g , we m u s t a t th e s a m e ti m e con s i d er h ow a com m u n i ty ca n b e en a b l ed to ca r r y ou t s u ch a p r oces s . Am on g th e p r oj ects th a t s h a r ed th e s i m i l a r a i m i s R u l eN et [ 11] , a n ex p er i m en t i n s u p p or ti n g con s en s u s b u i l d i n g u s i n g a n el ectr on i c con fer en ce com m i s s i on ed b y U.S. N u cl ea r R es ea r ch Cou n ci l (N R C) i n vol vi n g n on -tech n i ca l m em b er s of th e p u b l i c on a top i c wh i ch wa s u n ti l th en th ou g h t to b e h i g h l y tech n i ca l a n d ou t of l i m i ts to th em . P a r ti ci p a n ts ’ eva l u a ti on s wer e r ep or ted l y h i g h l y p os i ti ve, p r i m a r i l y b eca u s e i t m a d e th em feel th a t th ei r voi ces wer e h ea r d a n d th ei r con tr i b u ti on s h a d a n effect. However , N R C i ts el f q u es ti on ed th e cr ed i b i l i ty of d i s cu s s i on s a n d th e p a r ti ci p a n ts ’ q u a l i fi ca ti on s . Th er e wa s a l s o a s en s e of m i s tr u s t a m on g s om e p a r ti ci p a n ts towa r d s N R C con cer n i n g i ts m oti va ti on , a n d th e tech n i ca l s ta ff d ou b ted th a t a n y th i n g n ew h a s b een r a i s ed . In ter es ti n g l y , th i s r ep r es en ts ty p i ca l s ta n ces of p a r ti es fr om d i ffer en t s ector s i n vol ved i n s oci a l d eci s i on -m a k i n g — r eg a r d l es s of wh eth er i s i t con d u cted wi th or wi th ou t el ectr on i c con fer en ci n g . Th er efor e i t i s n ot th e p r ob l em of th e m ea n s , b u t th e ca p a ci ty of th e com m u n i ty to a ttem p t a con s en s u s -or i en ted d eci s i on -m a k i n g . Wh a t i s r eq u i r ed i s “ s oci a l i n tel l i g en ce” , a s th e p oten ti a l ca p a b i l i ty of a com m u n i ty to en g a g e i n i n for m ed d el i b er a ti ve d eci s i on -m a k i n g p r oces s . To d es i g n s oci a l i n tel l i g en ce, th en , wou l d b e to en a b l e p u b l i c to en g a g e i n d i s cou r s e. F or th i s p u r p os e, we a r e a ttem p ti n g to d evel op m ea n s to ta p i n to com m u n i ty i n for m a ti on h el d b y i n d i vi d u a l m em b er s , a n d ex p er i m en ti n g on a n ew for m of i n ter a cti on i n ca r r y i n g ou t d i s cu s s i on s . Th es e two a p p r oa ch es b oth r eq u i r e th e s en s e of coop er a ti on ; i n th e for m er , th er e m u s t b e a m oti va ti on a m on g m em b er s of th e com m u n i ty to g a th er a n d s h a r e i n for m a ti on ; i n th e l a tter , th e p a r ti ci p a n ts a r e ex p ected to en g a g e i n a con s tr u cti ve d el i b er a ti ve p r oces s , a n d fi t i n to th e a s s u m p ti on th a t d el i b er a ti on i s coop er a ti ve wor k . Id ea l l y , m em b er s of th e com m u n i ty s h ou l d b e en g a g ed i n p u b l i c d i s cou r s e a s a n ever y d a y a cti vi ty wi th ou t b ei n g con s ci ou s of i ts coop er a ti ve n a tu r e. On e wa y to a ch i eve i t, we b el i eve, i s th r ou g h p r ovi d i n g a n a cces s i b l e i n ter fa ce to ea s e i n d i vi d u a l i n ter a cti on , a n d d es i g n of i n ter a cti on th a t en h a n ces a wa r en es s a b ou t th e oth er s , fos ter i n g r econ ci l i a ti on of d i ffer en ces . Ou r cu r r en t r es ea r ch r es u l t i s s ti l l too p r em a tu r e to j u d g e wh eth er s u ch s oci a l i n tel l i g en ce d es i g n i s p os s i b l e, a n d ex ten s i ve ex p er i m en ts a n d eva l u a ti on s a r e r eq u i r ed to d r a w a n y con cr ete con cl u s i on s . Th i s wi l l b e ou r focu s u p on th e i m p l em en ta ti on of p r ototy p e s y s tem s .

8 .5 C o n c lu d in g R e m a r k Ad m i tti n g th a t vi s i on s a n d p r oj ects d es cr i b ed i n th i s p a p er 2 a r e r a th er ex p l or a tor y a n d s p ecu l a ti ve, we b el i eve th a t fos ter i n g p u b l i c d i s cou r s e a d d r es s es a wi d e r a n g e 2

Th e p r oj ects d es cr i b ed i n th i s p a p er a r e fu n d ed b y JSP S G r a n t-i n -Ai d for Sci en ti fi c R es ea r ch a n d JSP S R es ea r ch for th e F u tu r e P r og r a m . We a ck n owl ed g e Ak i r a Ka wa g u ch i a n d Tos h i y a s u Mu r a y a m a wh o a r e d evel op i n g p a r ts of th e s y s tem s th a t a r e r efer r ed to i n th i s p a p er .

66

K. N a k a ta

of i s s u es i n s u p p or ti n g i n ter a cti on s i n a com m u n i ty . In th e m or e p r a cti ca l s i d e, we h a ve i n i ti a ted a p i l ot p r oj ect th a t a ttem p ts to d evel op a s y s tem th a t s u p p or ts a n a l y ti c-d el i b er a ti on b y i n teg r a ti n g a s et of a n a l y ti ca l tool s i n to a n etwor k ed d i s cu s s i on s y s tem , i n cl u d i n g a cces s to s i m u l a tor s th a t m od el en vi r on m en ta l effects to b e u s ed a s j u s ti fi ca ti on s i n th e d i s cu s s i on . It i s en vi s i on ed th a t a p p r oa ch es d es cr i b ed i n th i s p a p er en h a n ce s u ch a s y s tem a n d con tr i b u te to i t a s en a b l i n g tech n ol og i es for p u b l i c d i s cou r s e i n a ch i evi n g en vi r on m en ta l l y s u s ta i n a b l e com m u n i ti es . It i s often s a i d th a t cr ea ti on of a s u s ta i n a b l e com m u n i ty i n vol ves ca p a ci ty b u i l d i n g . We b el i eve th a t “ s oci a l i n tel l i g en ce d es i g n ” i s a for m of ca p a ci ty b u i l d i n g th a t en a b l es p u b l i c to en g a g e i n d i s cou r s e a m on g va r i ou s va l u e j u d g m en ts a n d p er s p ecti ves , a n d i n cr ea s e a wa r en es s a b ou t th e en vi r on m en t i n cl u d i n g i ts i n h a b i ta n ts a n d p ol i cy m a k er s i n or d er to a ch i eve a r a ti on a l con s en s u s -b a s ed s oci a l d eci s i on -m a k i n g .

R e fe r e n c e s 8 .1 Ster n , P .C. a n d F i n eb er g , H.V. (ed s ): U n d e r s t a n d i n g R i s k : I n f o r m i n g D e c i s i o n s i n a D e m o c r a t i c S o c i e t y . Wa s h i n g ton D .C.: N a ti on a l Aca d em y P r es s (1996). 8 .2 N a k a ta , K. a n d F u r u ta , K.: In for m a ti on Sy s tem s R eq u i r em en ts for Su p p or ti n g P u b l i c D el i b er a ti on . In P r o c . I n t e r n a t i o n a l C o n f e r e n c e o n A d v a n c e s i n I n f r a s t r u c t u r e f o r E l e c t r o n i c B u s i n e s s , S c i e n c e , a n d E d u c a t i o n o n t h e I n t e r n e t ( S S G R R 2 0 0 0 ) , L ’ Aq u i l a , Ita l y (2000) CD -R OM. 8 .3 N a k a ta , K.: Kn owl ed g e a s a Soci a l Med i u m . N e w G e n e r a t i o n C o m p u t i n g 17 (1999) 3 95-4 05. 8 .4 Ja p a n es e G over n m en t E n vi r on m en t Ag en cy : K a n k y o H a k u s h o ( S o s e t s u ) (E n vi r on m en t Wh i te P a p er ). Tok y o: G y os ei (2000) (i n Ja p a n es e). 8 .5 N i s h i d a , T: P u b l i c op i n i on ch a n n el . N e w G e n e r a t i o n C o m p u t i n g 17 (1999) 4 17-4 27. 8 .6 Ku b ota , H. a n d N i s h i d a , T.: E g oCh a t Ag en t: A Ta l k i n g Vi r tu a l i z ed Mem b er for Su p p or ti n g Com m u n i ty Kn owl ed g e Cr ea ti on . In T h e P r o c e e d i n g s o f t h e A A A I F a l l S y m p o s i u m o n “ S o c i a l l y I n t e l l i g e n t A g e n t s — T h e H u m a n i n t h e L o o p ” (2000). 8 .7 Sch a n k , R .: T e l l M e A S t o r y . N or th wes ter n Un i ver s i ty P r es s (1990). 8 .8 Con k l i n , J. a n d B eg em a n , M.L .: g IB IS: A h y p er tex t tool i n g for ex p l or a tor y p ol i cy n d d i s cu s s i on . In Ta ta r , D . (ed .): P r o c e e d i n g s o f t h e 2 C o n fe r e n c e o n C o m p u te r S u p p o r t e d C o o p e r a t i v e W o r k . N ew Yor k : ACM (198 8 ) 14 0-152. 8 .9 B u ck i n g h a m Sh u m , S.: C o m p u t e r - S u p p o r t e d C o l l a b o r a t i v e A r g u m e n t a t i o n R e s o u r c e S i t e , h ttp : //k m i .op en .a c.u k /s b s /cs ca . 8 .10 Wh i twor th , B . a n d McQ u een , R .J.: Voti n g b efor e d i s cu s s i n g : Com p u ter voti n g a s s on d ci a l com m u n i ca ti on . In P r o c e e d i n g s o f t h e 3 2 H a w a i i C o n f e r e n c e o n S y s t e m S c i e n c e s (1999). 8 .11 F er en z , M. a n d R u l e, C.: R UL E N E T: An E x p er i m en t i n On l i n e Con s en s u s B u i l d i n g . In Su s s k i n d , L ., McKea r n a n , S. a n d Th om a s -L a r m er , J. (ed s ): T h e C o n s e n s u s B u i l d i n g H a n d b o o k . Th ou s a n d Oa k s , CA: Sa g e P u b l i ca ti on s (1999) 8 79-8 99.

9 . I n te r n e t, D is c o u r s e s , a n d D e m o c r a c y 1

1

R . L u eh r s , T. Ma l s ch , a n d K. Vos s 1

2

2

Tech n i ca l Un i ver s i ty of Ha m b u r g -Ha r b u r g , D ep a r tm en t for Tech n ol og y As s es s m en t, Sch wa r z en b er g s tr . 95, 21071 Ha m b u r g , G er m a n y P i x el p a r k AG , In s ti tu te for Med i a D evel op m en t, R oth er s tr . 8 , 1024 5 B er l i n , G er m a n y

9 .1 I n t r o d u c t io n It i s th e ver y p u r p os e of th e D E MOS1 p r oj ect – th e s u b j ect of th i s p a p er 2 – to ex p l oi t n ovel for m s of com p u ter m ed i a ted com m u n i ca ti on i n or d er to s u p p or t d em ocr a cy on -l i n e (' e-d em ocr a cy ' ) a n d to en h a n ce ci ti z en p a r ti ci p a ti on i n m od er n s oci eti es . In th e fol l owi n g p a p er we wi l l fi r s tl y p oi n t ou t h ow D E MOS a i m s to s u p p or t th e d em ocr a ti c p r oces s b y ex p l oi ti n g th e com m u n i ca ti ve p oten ti a l of th e In ter n et. Secon d l y , we wi l l i n tr od u ce a n ovel p a r ti ci p a ti on m eth od ol og y wh i ch i s d er i ved fr om d i ffer en t s oci a l s ci en ce a p p r oa ch es . Th i r d l y we wi l l b r i efl y d es cr i b e th e over a l l d es i g n a p p r oa ch .

9 .2 O n lin e S u p p o r t f o r D e m o c r a t ic P r o c e s s e s Si n ce th e n eol og i s m ' e-d em ocr a cy ' r efer s to b oth com p u ter m ed i a ted com m u n i ca ti on a n d d em ocr a cy wi th ou t s p eci fy i n g th e u n d er l y i n g con cep ts , th er e i s a n eed to ex p l a i n wh a t ex a ctl y we m ea n wh en u s i n g th e ter m . To s ta r t wi th th e ‘ d em ocr a cy ’ p a r t of th e ter m , th er e a r e d i ffer en t con cep ti on s of d em ocr a cy a n d d ep en d i n g on th e p er s p ecti ve, d i ffer en t p er cep ti on s of h ow th e i n ter n et cou l d s u p p or t, r efor m or even r evol u ti on i s e th e wa y d em ocr a cy wor k s . Th e m os t com m on d i s ti n cti on i n th e d efi n i ti on of d em ocr a cy r efer s to th e wa y s ci ti z en s p a r ti ci p a te i n th e d eci s i on m a k i n g p r oces s a n d th e r es p ecti ve ty p es a r e ca l l ed d i r ect or r ep r es en ta ti ve d em ocr a cy . However , th es e a p p r oa ch es h a ve to b e u n d er s tood n ot a s a l ter n a ti ve, op p os i n g s y s tem s of d em ocr a ti c g over n a n ce b u t a s two com p l em en ta r y for m s of p a r ti ci 1

2

D E MOS (D el p h i Med i a ti on On l i n e Sy s tem ) i s fu n d ed a s a s h a r ed -cos t R TD p r oj ect u n d er th e 5th F r a m ewor k P r og r a m m e of th e E u r op ea n Com m i s s i on (IST) a n d i s b ei n g d evel op ed b y a r es ea r ch con s or ti u m com p r i s i n g ei g h t or g a n i s a ti on s fr om fi ve d i ffer en t E u r op ea n cou n tr i es , r ep r es en ti n g th e fi el d s of a ca d em i c r es ea r ch , m u l ti m ed i a , s oftwa r e, m a r k et r es ea r ch a n d p u b l i c a d m i n i s tr a ti on . Th e D E MOS P r oj ect (IST-1999-2053 0) com m en ced Sep tem b er 2000 a n d i s g oi n g on for 3 0 m on th s . F or m or e i n for m a ti on s ee th e p r oj ect web s i te: h ttp : //www.d em os -p r oj ect.or g Th i s r ep or t d es cr i b es th e en ti r e s p r ea d of th e on g oi n g p r oj ect a n d h a s to b e s een a s a n s h or t i n tr od u cti on to th e p a r ti cu l a r fi el d s of r es ea r ch a n d d evel op m en t wh i ch a r e p u l l ed tog eth er i n D E MOS.

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 67-74 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

68

R . L u eh r s , T. Ma l s ch , a n d K. Vos s

p a ti on wh i ch ex i s t s i d e b y s i d e i n ever y m od er n s oci ety . It wou l d s i m p l y n ot b e fea s i b l e i n con tem p or a r y s oci eti es to a s k p eop l e for th ei r a p p r ova l b efor e com i n g to a n y d eci s i on s , l i k e th e a n ci en t G r eek s d i d , n or cou l d r ep r es en ta ti ve d em ocr a cy d i s p en s e wi th th e ci vi l en g a g em en t of th e ci ti z en s . Mos tl y , ‘ e-d em ocr a cy ’ i n th i s con tex t ca l l s i n to q u es ti on th e a p p r op r i a te m i x tu r e of b oth ty p es of p a r ti ci p a ti on , n ot r ep r es en ta ti ve d em ocr a cy a s s u ch . Wh eth er or n ot m or e d i r ect p a r ti ci p a ti on i s p er cei ved a s b ei n g d es i r a b l e, d ep en d s on th e u n d er l y i n g n or m a ti ve m od el of d em ocr a cy . F or th e l i b er a l , d em ocr a cy op er a tes b y a r r a n g i n g com p r om i s es b etween ci ti z en s wi th d i ffer en t i n ter es ts on th e b a s i s of fa i r p r oced u r es s u ch a s eq u a l voti n g r i g h ts . Th e n or m a ti ve i m p l i ca ti on s a r e l ow a n d th e l i b er ti es of th e ci ti z en s a r e a b ove a l l d efi n ed a s ' n eg a ti ve l i b er ti es ' i n th e s en s e of th em n ot b ei n g too m u ch d i r ected b y th e s ta te. F r om th i s p oi n t of vi ew, m or e d i r ect p a r ti ci p a ti on i s on l y wor th wh i l e – i f a t a l l – i n ter m s of p l eb i s ci tes b u t n ot i n ter m s of i n ten s i fi ed p u b l i c d eb a te. In th i s ca s e e-d em ocr a cy wou l d on l y m ea n s u b s ti tu ti n g p a p er -b a s ed p r oced u r es wi th el ectr on i c on es i n or d er to i n cr ea s e con ven i en ce a n d effi ci en cy . B y con tr a s t th e r ep u b l i ca n a p p r oa ch to d em ocr a cy b el i eves th a t " th e for m a ti on of th e ci ti z en ' s op i n i on a n d wi l l for m s th e m ed i u m th r ou g h wh i ch s oci ety con s ti tu tes i ts el f a s a p ol i ti ca l wh ol e" (Ha b er m a s 1996, 26). E s p eci a l l y i n i ts com m u n i ta r i a n r ea d i n g th e r ep u b l i ca n vi ew ten d s to over -con cep tu a l i s e eth i ca l va l u es a n d th e n eed a n d ch a n ces for eth i ca l l y i n teg r a ted s oci eti es . Al th ou g h , h er e th e In ter n et cou l d b e p oten ti a l l y u s ed i n i ts en ti r e d i ver s i ty i n or d er to s u p p or t p u b l i c d em ocr a ti c p r oces s es , th e ex p ecta ti on th a t el ectr on i c n etwor k s wi l l l ever a g e th e eth i ca l i n teg r a ti on of s oci ety s eem s to b e fa r too i d ea l i s ti c. Th ou g h th er e m i g h t b e a " tr en d towa r d s m or e a u ton om ou s l oca l u n i ts a n d th e em er g en ce of m u l ti cu l tu r a l a n d m or e eg a l i ta r i a n p ol i ti cs , (...) s tr on g cou n ter -ten d en ci es a r e a t wor k . Th e In ter n et i s i n vol ved i n th i s p r oces s b y b oth i n fl u en ci n g th e d es i r ed en d s a n d th ei r op p os i tes " (Sa s s i 1997, 4 3 6). In s tea d of i d en ti fy i n g d em ocr a cy m er el y wi th voti n g l i k e l i b er a l d em ocr a ts ten d to or r ed u ci n g p ol i ti ca l to eth i ca l q u es ti on s , l i k e r ep u b l i ca n d em ocr a ts a r e s u p p os ed to d o, a th i r d va r i a n t, th e d i s cou r s e th eor eti c (d el i b er a ti ve) m od el , focu s es on th e p r o c e d u r e s of p u b l i c wi l l for m a ti on . Th es e p r oced u r es a r e con s i d er ed to g en er a te l eg i ti m a cy a n d p r a cti ca l r a ti on a l i ty (B en h a b i b 1996, 71). " In a g r eem en t wi th r ep u b l i ca n i s m , i t g i ves cen ter s ta g e to th e p r oces s of p ol i ti ca l op i n i on - a n d wi l l for m a ti on " (Ha b er m a s 1996, 27) b u t wi th ou t b u r d en i n g th i s p r oces s wi th th e i d ea l i s ti c ex p ecta ti on of en a b l i n g th e p u b l i c s p h er e i ts el f to a ct. Accor d i n g to th e d i s cou r s e th eor y th i s q u a l i ty b el on g s ex cl u s i vel y to th e r ea l m of th e s p eci a l i s ed s u b s y s tem ca l l ed a d m i n i s tr a ti on . Th e p u r p os e of th e d el i b er a ti ve p r oces s i s , th ou g h , to i n fl u en ce th e ex er ci s e of p ower b y th e a d m i n i s tr a ti on . " Th e p ower a va i l a b l e to th e a d m i n i s tr a ti on ch a n g es i ts a g g r eg a te con d i ti on a s s oon a s i t em er g es fr om p u b l i c u s e of r ea s on s a n d com m u n i ca ti on s th a t d o n ot j u s t m on i tor th e ex er ci s e of p ol i ti ca l p ower r etr os p ecti vel y , b u t m or e or l es s p r og r a m i t a s wel l " (Ha b er m a s 1996, 24 ). In t h i s s e n s e , th e p r oj ect s tr i ves to s tr en g th en th e l eg i ti m a cy a n d r a ti on a l i ty of d em ocr a ti c d eci s i on m a k i n g p r oces s es b y u s i n g D E MOS to i n s p i r e a n d g u i d e l a r g e s ca l e p ol i ti ca l d eb a tes , to cl os e th e d i s ta n ce b etween p ol i ti ca l r ep r es en ta ti ves a n d ci ti z en s , ex p er ts a n d l a y m en .

9. In ter n et, D i s cou r s es , a n d D em ocr a cy

69

9 .3 A N o v e l P a r t ic ip a t io n M e t h o d o lo g y Th e s p eci fi c com m u n i ca ti on p oten ti a l of th e In ter n et ca n b e ch a r a cter i s ed b y th e th r ee ter m s - i n ter a cti vi ty , s p eed a n d s cop e. Tog eth er , th es e ch a r a cter i s ti cs a l l ow n ovel for m s of i n ter a cti ve com m u n i ca ti on b etween l a r g e n u m b er s of p a r ti ci p a n ts . On th e on e h a n d , i t i s th eor eti ca l l y p os s i b l e for a n u n l i m i ted n u m b er of p eop l e to d i s cu s s a com m on s u b j ect– a l l ‘ ta l k i n g a t th e s a m e ti m e’ a n d con tr i b u ti n g to th e s a m e d i s cu s s i on . On th e oth er h a n d , th e s a m e p a r ti ci p a n ts cou l d a l s o p oten ti a l l y u s e el ectr on i ca l l y a va i l a b l e i n for m a ti on to d eep en th ei r k n owl ed g e, to g i ve m or e evi d en ce to th ei r a r g u m en ts or to con vi n ce oth er p a r ti ci p a n ts . F u r th er m or e, p eop l e cou l d for m coa l i ti on s b y g etti n g i n tou ch wi th l i k e-m i n d ed p eop l e effor tl es s l y or th ey cou l d g r ou p a r ou n d a n d d i s cu s s cer ta i n top i cs or s u b top i cs of m u tu a l i n ter es t. To r ea l i s e th i s p oten ti a l h owever , th er e i s a n eed for m eth od ol og i es th a t m a tch th e m ed i a . Th ey n eed to b e a b l e to a g g r eg a te a n d i n ter r el a te th e i n d i vi d u a l con tr i b u ti on s , to i d en ti fy a n d fos ter th e m os t p r om i s i n g a s p ects of th e d i s cu s s i on , to p r ofi l e d i ffer en t p os i ti on s a n d to s tr i ve for con ver g en ce b etween th em or a t l ea s t to fi g u r e ou t wh a t a r e th e tr u l y d i s p u ted a s p ects wh er e n o com p r om i s e ca n b e a ch i eved . In th e l a tter ca s e, we a r e a l wa y s l ook i n g for a r es u l t fr om th e d i s cu s s i on - wh eth er i t i s a con s en s u a l s ta tem en t s u p p or ted b y a m a j or i ty of th e p a r ti ci p a n ts or wh a t i s ca l l ed a ‘ r a ti on a l d i s s en t’ 3 . On l y i f th e d i s cu s s i on l ea d s to a r es u l t i s th e d i s cu s s i on l i k el y to h a ve a n y i n fl u en ce on p ol i ti ca l d eci s i on -m a k i n g p r oced u r es . Th i s i m p a ct, of cou r s e, ca n b e m a n i fol d : i f th e ou tcom e i s a cl ea r s ta tem en t s u p p or ted b y th e b r oa d p u b l i c, i t wi l l n ot b e i g n or ed b y el ected r ep r es en ta ti ves . If th e r es u l t i s m er el y a wi d es p r ea d col l ecti on of d i ffer en t vi ewp oi n ts , i t ca n s er ve a s i n p u t to p r os p ecti ve l a ws or i t ca n a n ti ci p a te fu tu r e ob j ecti on s to p l a n n ed p ol i ci es a n d th e l i k e. Ta k i n g a cl os er l ook a t th i s m eth od ol og y , we a r e b a s i ca l l y p l a n n i n g to a s s em b l e a n d i n teg r a te th r ee wel l -p r oven s oci a l r es ea r ch m eth od s , n a m el y th e Su r vey tech n i q u e, th e D el p h i a p p r oa ch 4 a n d th e Med i a ti on m eth od 5. Th e d i ffi cu l ty h er e i s th a t th es e i d ea s ca n n ot s i m p l y b e a d d ed a n d com p i l ed to for m a n ew m eth od ol og y b eca u s e th ey a r e, a t l ea s t p a r ti a l l y , con tr a d i ctor y . Sta r ti n g wi th th e cl a s s i c Su r vey tech n i q u e, th i s m eth od i s d es i g n ed for r ep r es en ta ti ve op i n i on p ol l s a n d con tr i b u tes to p u b l i c op i n i on for m a ti on on a l a r g e-s ca l e b a s i s b y i n cl u d i n g (vi r tu a l l y ) th e en ti r e p op u l a ti on . However , th i s tech n i q u e i s r a th er u n s u i ta b l e for i n ter a cti ve p a r ti ci p a ti on . D el p h i p ol l s , on th e oth er h a n d , op er a te wi th a cer ta i n a m ou n t of i n ter a cti ve feed b a ck , b u t th i s h a s th e con s eq u en ce of l i m i ted s ca l a b i l i ty . F or D E MOS, D el p h i p ol l s a r e ex tr em el y i n ter es ti n g b eca u s e th ey ca n b e u s ed to ex p l oi t ex p er t k n owl ed g e. Th e b a s i c i d ea i s to g en er a te a con s en s u s a m on g a l i m i ted n u m b er of d om a i n ex p er ts b y a g g r eg a ted feed b a ck . F eed b a ck i s s u p p l i ed b y th e ‘ D el p h i s t’ on a s tr i ctl y a n on y m ou s a n d s ta ti s ti ca l b a s i s to 3

4 5

„ A r a ti on a l d i s s en t (...) i m p l i es th a t, on th e b a s i s of wh a t i s or h a s b een col l ecti vel y a ccep ted , th e p er s on s i n vol ved s u cceed i n u n d er s ta n d i n g p r eci s el y wh a t i s n ’ t col l ecti vel y a ccep ted “ (Mi l l er 1992, 14 ). As a n over vi ew s ee F l or i a n et a l . 1999. Th e m ed i a ti on m eth od i s on e of th e s o-ca l l ed Al ter n a ti ve D i s p u te R es ol u ti on (AD R ) p r oced u r es , wh i ch focu s on ' i n for m a l p a r ti ci p a ti on ' i n th e s en s e th a t th ey a r e n ot r eg u l a ted b y l a w. See Su s s k i n d a n d Cr u i k s h a n k (198 9), Ma er k er a n d Sch m i d t-B el z (2000).

70

R . L u eh r s , T. Ma l s ch , a n d K. Vos s

ex cl u d e d i r ect p er s on a l i n fl u en ce a m on g th e p a r ti ci p a n ts . A D el p h i p r oces s r u n s th r ou g h two (or m or e) cy cl es of i n ter vi ew-feed b a ck -i n ter vi ew. After ea ch cy cl e th e ex p er ts a r e a s k ed to r eth i n k th ei r or i g i n a l a n s wer s i n th e l i g h t of th e s ta ti s ti ca l l y a g g r eg a ted ‘ g r ou p op i n i on ’ th a t h a s em er g ed i n th e p r evi ou s cy cl e, u n ti l a s a ti s fa ctor y l evel of con ver g en ce or (s ta ti s ti ca l ) con s en s u s i s r ea ch ed . Wh er ea s b oth Su r vey a n d D el p h i a r e q u a n ti ta ti ve m eth od s , th e Med i a ti on tech n i q u e i s a q u a l i ta ti ve m eth od u s ed to r evea l p r ob l em s a n d r es ol ve con fl i cts . Th e b a s i c i d ea of Med i a ti on i s th a t con s en s u s i s n ot a s ta ti s ti ca l fi g u r e b u t a n eg oti a ted com p r om i s e. Med i a ti on i s a g r ou p p r oces s wi th a l i m i ted n u m b er of p a r ti ci p a n ts , ch a i r ed b y a n i m p a r ti a l m ed i a tor , a n d often r u n n i n g th r ou g h s ever a l cy cl es of op en d i s cu s s i on . It i s h i g h l y i n ter a cti ve a n d p a r ti ci p a ti ve, b u t b ei n g r es tr i cted to fa ce-tofa ce i n ter a cti on , i t i s u n s u i ted for l a r g e n u m b er s of a cti ve p a r ti ci p a n ts . Th e ch a l l en g e for th e D E MOS p r oj ect i s to ta k e th e a d va n ta g es of a l l th r ee m eth od s a n d com b i n e th em i n to a n ew m eth od ol og y for on -l i n e d em ocr a ti c p a r ti ci p a ti on a n d i n ter a cti ve con fl i ct r es ol u ti on . (1) F r om Su r vey s i t wi l l ta k e th e i d ea of m a s s op i n i on p ol l s on a l a r g e-s ca l e b a s i s , (2) fr om D el p h i i t wi l l ta k e th e i d ea of a cy cl i ca l d eci s i on p r oces s ex p l oi ti n g ex p er t k n owl ed g e, a n d (3 ) fr om Med i a ti on i t wi l l ta k e th e i d ea of a n op en p r oces s of p a r ti ci p a ti ve con fl i ct r es ol u ti on . Th e i n com p a ti b i l i ti es m en ti on ed ea r l i er ca n b e ea s ed b y en r i ch i n g ea ch of th e p a r ti cu l a r m eth od s wi th el em en ts b or r owed fr om th e oth er s . F or ex a m p l e, i n s tea d of con d u cti n g a s ta n d a r d i s ed s u r vey wi th p r e-for m u l a ted q u es ti on s , th e i tem s ca n b e g en er a ted ‘ b ottom u p ’ b y s or ti n g a n d a g g r eg a ti n g q u a l i ta ti ve s em a n ti c con ten t fr om ea r l i er or on g oi n g d i s cu s s i on s . Th e g en er a ti on of th e q u es ti on n a i r e, th en , i s con cep tu a l i s ed a s a n i n ter a cti ve p r oces s . L i k e con ven ti on a l s u r vey s , th e m a i n p u r p os e h er e i s to con d en s e a n d a g g r eg a te i n for m a ti on a n d b ey on d th a t to s u m m a r i s e th e d i s cu s s i on a t a cer ta i n s ta g e. Accor d i n g l y cl a s s i ca l D el p h i s tu d i es ca n b e s u p p l em en ted wi th q u a l i ta ti ve, op en en d ed q u es ti on s a n d ex ten d ed to i n vol ve h i g h er n u m b er s of p a r ti ci p a n ts . On th e oth er s i d e, th e Med i a ti on m eth od h a s to b e a d a p ted to th e s p eci fi c con s tr a i n ts of th e In ter n et, th a t i s m a i n l y to d evel op fu n cti on a l eq u i va l en ts wh i ch tr a n s fer th e m eth od ’ s cor e s tr en g th s , l i k e cr ea ti n g a n a tm os p h er e of con fi d en ce a n d tr u s t fr om fa ce-to-fa ce i n ter a cti on s , to th e on -l i n e d om a i n . Th e th r ee s oci a l r es ea r ch m eth od s (Su r vey , D el p h i a n d Med i a ti on ) wi l l b e a p p l i ed a n d m er g ed tog eth er i n th e s o-ca l l ed ' D E MOS p r oces s ' . Th i s p r oces s i s a l wa y s con cer n ed wi th on e m a i n top i c to b e com m on l y d i s cu s s ed on a l i m i ted ti m el i n e u n d er th e g u i d a n ce of on -l i n e m od er a tor s . To l i m i t th e d eb a te to n ot m or e th a n on e m a i n top i c i s a con cep tu a l d eci s i on d er i ved fr om th e g en er a l ob j ecti ve of th e p r oj ect to con cen tr a te on d el i b er a ti ve d i s cou r s es wi th p oten ti a l i m p a ct on p u b l i c d eci s i on m a k i n g p r oces s . It a l s o s er ves to d i s cou r a g e d eb a tes fr om l os i n g a n y s en s e of d i r ecti on . As a m a tter of cou r s e s ever a l p r oces s es ca n b e con d u cted i n p a r a l l el a n d ea ch of th em wi l l s p l i t u p i n to d i ffer en t s u b top i cs d u r i n g th e cou r s e of th e d eb a te. To focu s on j u s t on e m a i n top i c r eq u i r es a ca r efu l s el ecti on of th e top i c to b e d i s cu s s ed on th e b a s i s of g en er a l cr i ter i a . Wi th i n ou r r es ea r ch p r oj ect we h a ve fou n d th a t a p oten ti a l th em e s h ou l d a t l ea s t m eet cr i ter i a l i k e p op u l a r i ty , com p l ex i ty , con tr over s y a n d p er s i s ten cy . Th e q u es ti on of to wh a t ex ten t a D E MOS p r oces s a ffects ‘ r ea l -wor l d ’ d eci s i on s i m p l i es a d d i ti on a l l y a q u es ti on r el a ti n g to th e g en er a l s u cces s of p u b l i c d i s cou r s es , wh i ch ca n n ot b e ex p a n d ed on h er e.

9. In ter n et, D i s cou r s es , a n d D em ocr a cy

71

Th e b a s i c p r oces s m od el com p r i s es th r ee d i ffer en t p h a s es ea ch wi th s p eci fi c g oa l s . Th e fi r s t p h a s e h a s a b ove a l l to i n i ti a te, fa ci l i ta te a n d b r oa d en th e d eb a te a n d s u b s eq u en tl y to i d en ti fy th e m os t i m p or ta n t a s p ects or s u b top i cs of th e ch os en s u b j ect m a tter . Th er efor e th e m od er a tor s h a ve to a n a l y s e a n d cl u s ter th e fr ee tex t con tr i b u ti on s i n or d er to fi n d ou t th e i s s u es m os t p a r ti ci p a n ts s eem to b e i n ter es ted i n . Th es e ta s k s wi l l b e s u p p or ted b oth on a m eth od ol og i ca l a n d tech n ol og i ca l l evel . Th e m od er a tor s wi l l b e b a ck ed u p b y q u a l i ta ti ve m eth od s of con ten t a n a l y s i s a n d ca n ex p l oi t va r i ou s m ech a n i s m s r el a ti n g to th e s oci a l s y s tem ’ s s el f-or g a n i s a ti on . A g ood ex a m p l e of th e l a tter i s th e d etecti on a n d u s e of th e th r ea d -g en er a ti n g p a r ts of th e d i s cu s s i on . Her e a tex t m i n i n g tool wi l l b e a b l e to a u tom a ti ca l l y g r ou p th e tex t con tr i b u ti on s on ce a s et of ca teg or i es (s u b top i cs ) a r e d efi n ed a n d i l l u s tr a ted b y ex a m p l es . Ad d i ti on a l l y , th e m od er a tor s wi l l h a ve to s u m m a r i s e th e d i s cu s s i on d u r i n g th e cou r s e of th e fi r s t p h a s e fol l owi n g a s p eci fi c p r oced u r e. Th es e s u m m a r i es con s i s t of con ten t a n d p r og r es s r el a ted p a r ts a n d h i g h l i g h t a n d p r ofi l e em er g i n g l i n es of con fl i ct a ccor d i n g to th e Med i a ti on m eth od . Th e fi r s t p h a s e fi n a l l y r es u l ts i n a s et of p r op os ed s u b top i cs th a t ca n b e m or e i n ten s i vel y d i s cu s s ed i n s ep a r a te d i s cu s s i on for u m s i n th e n ex t p h a s e. Si n ce th i s p r oced u r e i s r el y i n g on i n ter p r eta ti on s of th e i n d i vi d u a l p os ti n g s a s wel l a s of th e en ti r e d i s cu s s i on , th e r es u l t m a y n ot ex a ctl y m eet th e p r efer en ces of th e p a r ti ci p a n ts . At th i s p oi n t th e Su r vey m eth od com es i n to p l a y i n or d er to eva l u a te wh eth er or n ot th e p r op os ed s u b -for u m s m eet th e d em a n d s of th e com m u n i ty a n d i f n eces s a r y , to g en er a te i d ea s on h ow to r evi s e th e l i s t of s u b top i cs . In th e s econ d p h a s e a l i m i ted n u m b er of s u b -for u m s wi l l b e offer ed b y th e s y s tem on th e b a s i s of th e p ol l r es u l ts . Th e p u r p os e of th i s p h a s e i s to i n ten s i vel y d i s cu s s s p eci fi c a s p ects i n s m a l l er g r ou p s of i n ter es ted p a r ti ci p a n ts , wh i l e th e m a i n for u m s ti l l ca tch es th os e p a r ti ci p a n ts wh o wa n t to d i s cu s s th e top i c on a m or e g en er a l l evel . Ag a i n th e m od er a tor s wi l l h a ve to s u m m a r i s e th e d evel op i n g d eb a te on a r eg u l a r b a s i s a n d a t th e s a m e ti m e tr y to tea s e ou t a n d m a n a g e em er g i n g con fl i cts . Th i s i s wh er e th e Med i a ti on m eth od com es i n a s p a r t of th e m od er a tor ’ s ta s k wi l l b e to cl a r i fy h ow a n d to wh a t ex ten t p eop l e a r e a g r eei n g or d i s a g r eei n g a n d a t th e s a m e ti m e to r ed u ce th e d i s ta n ce b etween d i ver g i n g p os i ti on s b y d el i b er a ti ve, m od er a ted d i s cou r s es . Th e r es u l ts of th e s econ d p h a s e s h ou l d ei th er b e a g r eem en t (con s en t) or a r a ti on a l d i s s en t i n th e s en s e ex p l a i n ed a b ove. If r eq u i r ed a n d a p p r op r i a te, th i s op i n i on s h a p i n g p r oces s ca n b e en r i ch ed a n d s u p p l em en ted wi th ex p er t k n owl ed g e b y con d u cti n g D el p h i s u r vey s a m on g a p r ed efi n ed s et of d om a i n ex p er ts . D el p h i ty p e s tu d i es ca n ei th er b e a p p l i ed i n th e or i g i n a l fa s h i on e.g . to r ed u ce th e u n cer ta i n ty wi th r es p ect to fu tu r e d evel op m en ts or i n or d er to eva l u a te cer ta i n p os i ti on s of th e com m u n i ty fr om a n ex p er t p oi n t of vi ew. Si n ce even ex p er ts a r e often n ot of th e s a m e op i n i on th e D el p h i m eth od h er e p r ovi d es th e p a r ti ci p a n ts wi th a con d en s ed p i ctu r e of th ei r d eg r ee of a g r eem en t r eg a r d i n g s p eci fi c i s s u es . F i n a l l y th e m od er a tor s wi l l cl os e th i s p h a s e wi th a s u m m a r y of wh a t wa s d i s cu s s ed s o fa r , a n d wi l l on ce a g a i n a s k th e p a r ti ci p a n ts for th ei r a p p r ova l (s u r vey ). Th e th i r d p h a s e r ei n teg r a tes th e s u b -for u m s i n to th e s ti l l ex i s ti n g m a i n for u m b y tr a n s fer r i n g th e s u m m a r i es a n d r el a ted s u r vey r es u l ts . Her e th e p a r ti ci p a n ts h a ve th e op p or tu n i ty to s ee th e p a r ti cu l a r s u b top i c a s p a r t of th e g en er a l s u b j ect m a tter

72

R . L u eh r s , T. Ma l s ch , a n d K. Vos s

a n d a ‘ b i g p i ctu r e’ wi l l em er g e. P a th e m a i n top i c a n d th e a s s em b l ed wi l l b e a s k ed to r a te th e s u b top i cs th e D E MOS p r oces s wa s i n ten ti on d en s ed d ocu m en t d ep i cti n g b oth th s i on a n d th e i m p or ta n ce a ccor d ed p a n ts .

9 .4 S y s t e m

r ti ci p a n ts h a ve th e l a s t ch a n ce to com m en t on r es u l ts of th e s u b -for u m s a n d th e com m u n i ty i n ter m s of i m p or ta n ce for th e m a i n top i c th a t a l l y s et u p for . Th e fi n a l r es u l t wi l l b e a con e r es u l ts of a d y n a m i c a n d d el i b er a ti ve d i s cu s i ts d i ffer en t a s p ects i n th e vi ew of i ts p a r ti ci -

D e s ig n

Th e d es i g n a p p r oa ch for th e D E MOS s y s tem s ta r ted wi th th e d ed u cti on of th e g en er i c D E MOS p r oces s fr om th e p a r ti ci p a ti on m eth od ol og y a s d es cr i b ed i n th e p r evi ou s ch a p ter . Accor d i n g l y th e g r a p h i ca l u s er i n ter fa ce (G UI) d ep i cts th e m a i n ch a r a cter i s ti cs of th i s p r oces s , e.g . vi s u a l i s es th e d i ffer en t p h a s es wi th i n a g i ven ti m e l i m i t, d i ver s e d i s cu s s i on for u m s a n d u s er r ol es . Th e n a vi g a ti on a l con cep t i s b a s ed on a ti m el i n e, wh i ch a l l ows th e u s er to d i s cer n th e cu r r en t p h a s e of th e d i s cu s s i on , a n d th e a ctu a l top i cs . Sta r ti n g fr om th er e, u s er s ca n z oom s u cces s i vel y i n to th e focu s of th ei r i n ter es t, th a t i s , i n to s u b -for u m s a n d p os ti n g s . Th e n u m b er of s u b -for u m s i s l i m i ted b y th e d em a n d s of s cr een d es i g n a n d u s a b i l i ty . In or d er to tech n i ca l l y s u p p or t th e D E MOS p r oces s , th e s y s tem a r ch i tectu r e con s i s ts of fou r m a j or s u p p or t com p on en ts for th e m od u l es : Ar g u m en ta ti on a n d Med i a ti on (A& M), On l i n e D el p h i Su r vey s (OD S), Su b g r ou p F or m a ti on a n d Ma tch m a k i n g (SF M) a n d Kn owl ed g e Ma n a g em en t Sy s tem (KMS). Th e m a i n el em en t of D E MOS i s th e for u m , wh er e top i cs a r e d i s cu s s ed u n d er th e g u i d a n ce of a m od er a tor . Th e d i s cu s s i on for u m s of th e Ar g u m en ta ti on a n d Med i a ti on m od u l e a r e p r ovi d ed b y th e Z en o s y s tem (G or d on et a l . 2001). Z en o p r ovi d es p a r ti cu l a r s u p p or t to tr u s ted th i r d p a r ti es (e.g . th e i m p a r ti a l m ed i a tor ) r es p on s i b l e for m od er a ti n g th e d i s cu s s i on s . Th e Z en o s er ver i s a j a va b a s ed a p p l i ca ti on for th e www, wh i ch en a b l es a n d fa ci l i ta tes m od er a ted , i s s u e b a s ed d i s cu s s i on for u m s i n a s ecu r e en vi r on m en t. Z en o d i s cu s s i on for u m s a r e i n teg r a ted wi th a wor k s p a ce fa ci l i ty for s h a r i n g cl a s s i fi ed d ocu m en ts . Th e On l i n e D el p h i Su r vey m od u l e p r ovi d es th e m od er a tor s wi th m ea n s to g en er a te a n d con d u ct on -l i n e s u r vey s a s p r evi ou s l y d es cr i b ed . In a fi r s t s tep , a d i s cu s s i on wi l l b e a n a l y s ed q u a l i ta ti vel y a n d ca teg or i s ed wi th th e h el p of a tex t d a ta m i n i n g a l g or i th m b a s ed on s ta n d a r d B a y es i a n i n fer en ce m eth od s . Th i s en g i n e i s a b l e to ex tr a ct th e ‘ con cep ts ’ , or m a i n i d ea s ou t of a fr ee tex t a n d to s ea r ch for ‘ s i m i l a r tex ts ’ b a s ed on com p a r i s on of th es e con cep ts . On ce th e m od er a tor h a s cl u s ter ed th e con tr i b u ti on s of th e u s er s a n d s o p r el i m i n a r y s tr u ctu r ed th e d i s cu s s i on , s h e m a y g en er a te a q u es ti on n a i r e a n d con d u ct a d eta i l ed q u a n ti ta ti ve s u r vey i n or d er to va l i d a te h er fi n d i n g s , cl a r i fy p a r ti cu l a r i s s u es or focu s on cer ta i n a s p ects . F u r th er m or e th e OD S com p on en t s u p p or ts D el p h i s u r vey s a n d th e vi s u a l i s a ti on of r es u l ts , wh i ch a r e s u b s eq u en tl y u s ed to fu r th er or g a n i s e th e D E MOS p r oces s a n d a l s o to es ta b l i s h n ew for u m s a n d g r ou p s of u s er s . Th e cl u s ter i n g of u s er s i s cr u ci a l for th e s ca l a b i l i ty of th e s y s tem . It wi l l b e h a n d l ed b y th e Su b g r ou p F or m a ti on a n d Ma tch m a k i n g m od u l e wh i ch m a k es u s e of

9. In ter n et, D i s cou r s es , a n d D em ocr a cy

73

d i ffer en t p r ofi l i n g i n for m a ti on . To m a i n ta i n s ca l a b i l i ty on th e tech n i ca l l evel , SF M i s a l s o b a s ed on th e ca teg or i s a ti on tool . Th e fi r s t, l i m i ted d ep l oy m en t of th e s y s tem wi l l l ea d to a d eep er u n d er s ta n d i n g of th e u s er s b eh a vi ou r i n th e D E MOS en vi r on m en t. On ce, th e b eh a vi ou r of u s er s a n d th e r u l es a r e k n own p r eci s el y , fu r th er ta s k s ca n b e a u tom a ted . It i s p l a n n ed to r ep r es en t u s er s a s wel l a s for u m s wi th s oftwa r e a g en ts . Th es e a g en ts wi l l ca r r y a s et of r u l es d er i ved for m th e fi r s t, ’ m a n u a l ’ d ep l oy m en t of D E MOS, wh i ch wi l l a l l ow for u m a g en ts to m a tch l i k e-m i n d ed u s er s a n d ex p er ts , u s er a g en ts to i d en ti fy a p p r op r i a te for u m s a n d u s er s to s et u p th ei r own g r ou p s a n d for u m s i n l i n e wi th th e p r og r es s of th e m a i n p r oces s . In oth er wor d s m or e a n d d en s er i n ter a cti on b etween a l a r g e n u m b er of p a r ti ci p a n ts ca n b e r ea l i z ed b y th e h el p of s oftwa r e a g en ts i n th e con tex t of D E MOS th a n i n a n y r ea l wor l d en vi r on m en t. Th i s ca n b e l a b el ed a s ’ i n ter a cti ve m a s s com m u n i ca ti on ’ , wh i ch d en otes a n ew i n ter a cti on ty p e owi n g to th e d i ffu s i on of th e web . B efor e, i t wa s j u s t p a r t of th e d efi n i ti on of m a s s m ed i a , th a t i n ter a cti on b etween s en d er a n d r ecei ver wa s i n h i b i ted b y i n ter p os ed tech n ol og y (L u h m a n n 2000). As n ew m ea n s of com m u n i ca ti on a n d i n ter a cti on i n d u ce n ew a n d u n ex p ected for m s of b eh a vi or , we fu r th er m or e ex p ect to ob s er ve em er g en t s tr u ctu r es i n th i s ’ h y b r i d s oci ety ’ wh i ch m a y b e a l s o of i n ter es t for b a s i c r es ea r ch p r ob l em s l i k e th e s o-ca l l ed ’ m i cr o-m a cr o-l i n k ’ . Th i s p r ob l em i s of cr u ci a l i m p or ta n ce for b oth s oci ol og y a n d com p u ter s ci en ce6 a n d i s es p eci a l l y focu s s ed i n th e r ecen tl y es ta b l i s h ed r es ea r ch fi el d ’ s oci on i cs ’ (Mu el l er et. a l 1998 , Ma l s ch 1998 ). Th e a g en t’ s a b i l i ty to l ea r n wi l l b e fi n a l l y u s ed for th e Kn owl ed g e Ma n a g em en t Sy s tem (KMS). As d es cr i b ed a b ove th e ca teg or i s a ti on en g i n e wi l l en a b l e a g en ts to s ea r ch for ‘ s i m i l a r tex ts ’ b a s ed on com p a r i s on of ex tr a cted con cep ts . In p a r ti cu l a r , th i s a l l ows a g en ts to fi n d d ocu m en ts , even i f th ey d o n ot con ta i n a d es i r ed k ey wor d . Th e a g en ts ca n th en b e u s ed to r ep r es en t a p a r ti cu l a r s et of d ocu m en ts cover i n g a cer ta i n s u b j ect m a tter . P r ovi d i n g th e p a r ti ci p a n ts wi th a cou p l e of i n i ti a l l y tr a i n ed a g en ts , th e u s er s ca n fu r th er m od i fy th ei r p er s on a l cop i es b y r etr a i n i n g . F u r th er m or e, th e a g en ts ca n b e s h a r ed a m on g th e u s er s , s o th a t p a r ti ci p a n ts wi l l n ot h a ve to s ta r t th ei r own r es ea r ch fr om s cr a tch , b u t ca n r etr a i n a n ex i s ti n g a g en t a n d s o r eu s e th e ex p er ti s e of oth er s 7. Wi th th e a n on y m ou s ex ch a n g e of a g en ts b ou n d to a cer ta i n top i c, even u s er s wi th con tr a d i ctor y th eor i es or op i n i on s ca n m u tu a l l y b en efi t fr om th ei r r es p ecti ve r es ea r ch b y u s i n g for ei g n a g en ts . E ven i f th e a g en ts a r e n ot p er fectl y tr a i n ed wi th r es p ect to th e i n for m a ti on n eed s of p a r ti cu l a r u s er s , i t m a y a t l ea s t s et th em on a n ew tr a ck . Th e m a i n i d ea i s to en a b l e “ com m u n i ca ti on th r ou g h s h a r ed k n owl ed g e” (e.g . ex ch a n g e a g en ts ), wh i ch wa s on e of th e i n i ti a l i d ea s of Ti m B er n er s -L ee (1997) wh en d evel op i n g th e wor l d wi d e web .

6 7

E .g . i n th e fi el d of ‘ D i s tr i b u ted Ar ti fi ci a l In tel l i g en ce’ (G a s s er 1991) Th i s con cep t h a s i n i ti a l l y b een d evel op ed i n th e p r oj ect www.es ton i a -s i n k i n g .or g (fu n d ed b y th e Med i a II p r og r a m of th e E C), wh er e u s er s a n d g r ou p s wi th d i ffer en t (even con tr a d i ctor y ) i n ter es ts a n d p r i or k n owl ed g e ca n con d u ct th ei r r es ea r ch a b ou t th e r ea s on s for th e s i n k i n g of th e fer r y E s ton i a .

74

R . L u eh r s , T. Ma l s ch , a n d K. Vos s

R e fe r e n c e s 9.1 B en h a b i b , S., 1996. Towa r d a D el i b er a ti ve Mod el of D em ocr a ti c L eg i ti m a cy . In : S. B en h a b i b , ed . D em ocr a cy a n d d i ffer en ce: con tes ti n g th e b ou n d a r i es of th e p ol i ti ca l ; P r i n ceton , N J: P r i n ceton Un i v. P r es s , 67-94 . 9.2 B er n er s -L ee, T., 1997. R ea l i s i n g th e F u l l P oten ti a l of th e Web . P u b l i s h ed a t h ttp : //www.w3 .or g /1998 /02/P oten ti a l .h tm l , vi s i ted 2001-07-24 9.3 F l or i a n , M. et a l ., 1999. Th e F u tu r e of Secu r i ty a n d R i s k s i n Tel ecom m u n i ca ti on – E x p ecta ti on s of E x p er ts (Tel eD el p h i ). In : G . Mu el l er a n d K. R a n n en b er g , ed s . Mu l ti l a ter a l Secu r i ty i n Com m u n i ca ti on s . Vol . 3 . Mu en ch en ; R ea d i n g , Ma s s a ch u s etts : Ad d i s on -Wes l ey -L on g m a n , 4 65-4 8 0. 9.4 G a s s er , L . a n d M. N . Hu h n s , ed s . 198 9. D i s tr i b u ted Ar ti fi ci a l In tel l i g en ce, Vol u m e II, L on d on : P i tm a n , Sa n Ma teo, Ca .: Mor g a n . 9.5 G or d on , T. et a l ., 2001. Z en o: G r ou p wa r e for D i s cou r s es on th e In ter n et. To a p p ea r i n KI - Kü n s tl i ch e In tel l i g en z , Vol . 15, 2001. 9.6 Ha b er m a s , J., 1996. Th r ee N or m a ti ve Mod el s of D em ocr a cy . In : S. B en h a b i b , ed . D em ocr a cy a n d d i ffer en ce: con tes ti n g th e b ou n d a r i es of th e p ol i ti ca l ; P r i n ceton , N J: P r i n ceton Un i v. P r es s , 21-3 0. 9.7 L u h m a n n , N ., 2000. Th e R ea l i ty of th e Ma s s Med i a , P a l o Al to: Sta n for d Un i ver s i ty P r es s . 9.8 Ma l s ch , T. ed . 1998 . Soz i on i k : s oz i ol og i s ch e An s i ch ten ü b er k ü n s tl i ch e Soz i a l i tä t. B er l i n : E d i ti on Si g m a . 9.9 Ma er k er , O. a n d B . Sch m i d t-B el z , 2000. On l i n e Med i a ti on for Ur b a n a n d R eg i on a l P l a n n i n g . In : A. B . Cr em er s a n d K. G r eve, ed s . Com p u ter Sci en ce for E n vi r on m en ta l P r otecti on , Ma r b u r g (G er m a n y ): Metr op ol i s , 158 -172. 9.10 Mi l l er , M., 1992. D i s cou r s e a n d Mor a l i ty – two ca s e s tu d i es of s oci a l con fl i cts i n a s eg m en ta r y a n d a fu n cti on a l l y d i ffer en ti a ted s oci ety . Ar ch i ves E u r op é en n es d e Soci ol og i e, 3 – 3 8 . 9.11 Mu el l er , H. J. et a l . 1998 . SOCION ICS: In tr od u cti on a n d P oten ti a l . Jou r n a l of Ar ti fi ci a l Soci eti es a n d Soci a l Si m u l a ti on , Vol .1, n o.3 , h ttp : //www.s oc.s u r r ey .a c.u k /JASSS/1/3 /5.h tm l , vi s i ted 2001-07-24 9.12 Sa s s i , S., 1997. Th e In ter n et a n d th e Ar t of Con d u cti n g P ol i ti cs : Con s i d er a ti on s of Th eor y a n d Acti on . Com m u n i ca ti on 22, 4 , 4 51-4 69. 9.13 Su s s k i n d , L . a n d J. Cr u i k s h a n k , 198 9.B r ea k i n g th e Im p a s s e. Con s en s u a l Ap p r oa ch es to R es ol vi n g P u b l i c D i s p u tes . N ew Yor k : Th e P er s eu s B ook s G r ou p .

10. How to Evaluate Social Intelligence Design Nobuhiko Fujihara Naruto University of Education / Synsophy Project, CRL Takashima, Naruto, 772-8502, JAPAN email: [email protected]

In this paper, it is discussed how to estimate computer network tools which support communications among community members. So far, standard methods do not seem to be developed enough to evaluate tools appropriately. How we should evaluate network communication tools designed to support social intelligence and to facilitate knowledge creation in a community? I’ll propose some important points which should be taken into account to estimate tools, and discuss some methods of evaluations through the introduction of our trials to estimate the effect of Public Opinion Channel (POC) on knowledge creation[10.1, 10.2, 10.7].

10.1 Computer Networked Community as Social Intelligence First of all, in order to discuss social intelligence design, I propose a viewpoint that considers societies and communities (especially computer networked communities) as having a kind of intellectual existence. The viewpoint would enable us to apply some useful research interests, theories and methodologies from studies of human intelligence to the discussion. It allows us to define the terms social intelligence and social intelligence design as follows1 . Social intelligence (SI) is defined as an ability which communities have to solve various problems. Social intelligence design is defined as the design of mechanisms of communities which are related to intellectual activities by the communities and their members. For instance, a design of SI means to arrange channels of information to facilitate knowledge creation by communities and their members. The viewpoint mentioned above generate new research interests on SI as follows: – Do SI develop? Does the development of SI relate to the development of communities? – What type of network communication systems do SI support? 1

Some researchers may define the term social intelligence as an ability to get along with others, or as the objects which have such kind of an ability[10.6]. Of course, it is very important to discuss how design this kind of objects. But in this paper, I don’t use the term SI in this manner.

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 75− 8 2, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

76

N. Fujihara

– Can SI be divided into subcategories? Can we apply the distinctions used in psychology, for example, fluid intelligence and crystallized intelligence? – Can SI quotient (SIQ) be measured? Can we create SIQ measurement tests? It is worth dealing with each of these issues, and there are further research issues also to be considered. In this paper, I focus on just one of these issues. That is, I’ll discuss how to evaluate a social intelligence design — in concrete, how the development of SI is measured when a community adopts a new network communication tool. In following sections, I emphasize here these three points. First, to evaluate tools, a baseline, “control” condition should be set up appropriately. The effects of tools can be measured by comparing the case in which tools are used and a baseline condition. Second, to evaluate tools, some different types of methods should be used together. Especially, researchers do never evaluate tools only based on users’ subjective judgments obtained by questionnaires, estimations, and introspections. Third, I’ll discuss the possibility to apply the network analysis to investigate how community members interact to each other and how knowledge creation is facilitated.

10.2 The Importance of Control Condition in Evaluating Social Intelligence Design Various types of network communication tools have been proposed. Some of these tools aim to support knowledge creation, and some aim to support communication among community members. If tools achieve their goals, their mechanisms would apply to developments of new tools. On the other hand, if the tools don’t achieve their goals, they should be improved. To estimate whether tools attain their functions, the differences should compare performance of a community or community members between the case in which tools with the function are used and when tools without the function are used. The case in which members use tools without the function is called as control condition. When control condition is biased, the effect of the function devised on tools cannot be estimated exactly. Thus, it is very important for the estimation of the tool’s effectiveness that control conditions are set up appropriately. Furthermore, when control condition is not set up, it could not be denied the possibility that a community and community members achieve performances even if without the tools. But in some researches, tools seem to be estimated without setting a control condition. It’s not enough to decide whether the tools really support activities of a community and community members. How should a control condition be set up? One of the appropriate methods is that tools are designed as a composition of a basic part and some additional parts. The case in which people use a tool constructed with only a

10. How to Evaluate Social Intelligence Design

77

basic part may constitute a control condition, and the cases where people use tools constructed with a basic part and some additional parts may constitute experimental conditions. The effect of a tool would be observed as the difference between the control condition and the experimental conditions. For example, Public Opinion Channel (POC) which is developed and researched by my colleagues and me, is designed in such a way[10.1, 10.2, 10.7]. POC is an interactive community broadcasting system. POC collects information from community members, edits and summarized information, and broadcasts it as a story. Community members listen to a story, and respond to it. Repeating the cycle, POC creates continuous information circulation in a community. To estimate the functions of POC, the case where the system is used which has only basic functions, that is, collecting messages and broadcasting them, is set up as a control condition. Research issues on POC are “How should information be summarized to facilitate knowledge creation in a community?”, “Are anonymous communication systems effective to inhibit troubles in communication like flames?”, and so on. The cases can be used as experimental conditions where POC with additional functions reflecting these issues is used. One of possible experimental conditions would be a POC with a summarization function. The effect of the summarization function could be observed when comparing the differences of some measurements, for example, quantities of message circulation, and the increasing rate of users, between the control condition and the experimental condition. In a similar way, some modules which aim to implement the same function can be compared. Tools are not always designed with modules. As another way to set up a control condition, typical situations can be used where people use ordinal network communication systems like mailing lists, bulletin board systems, and chats. For the purpose, it is useful to define typical situations and to standardize procedures to collect data and to analyze data. Fujihara and Miura observed search engine users who query information from WWW and analyzed their behavior[10.4, 10.5]. In the research, they proposed categories to describe information query behaviors from WWW with search engines. Such research would reveal our common activities in network communities. It will give us a baseline to estimate novel network communication tools.

10.3 How to Evaluate POC Methodologies to estimate whether network communication tools facilitate knowledge creation could be classified into following three categories: – analyses of users’ subjective estimations and introspection collected through questionnaire – log analyses of users’ behavior in natural conditions – experimental methods

78

N. Fujihara

Analysis of users’ subjective estimations and introspection is a very effective method because it is easy to operate and it gives us rich information on users’ thoughts directly. On the other hand, the data can be easily biased by subjection of users and experimenters. Some researchers reported that users do not always recognize their own behavior exactly and their subjective judgment and behavior sometimes are divided[10.8, 10.9]. Log analyses and experimental methods supplement such a methodological problem because they give us information users’ behaviors. But, of course, these methods have some problems. It is difficult to operate, and show us only a small part of facts on usage of tools. In order to estimate network communication tools, it is necessary to use these three methods together. Now, my colleagues and I estimates POC with these three methods. Among these estimations, I’ll focus on the result of log analysis. It is because we are trying to develop the method for analyzing network communication tools and knowledge creation generating on network communications, that is, the application of the method called network analysis[10.10]. Network analysis is the method to analyze relationships among community members and relationships among companies. It is mainly used in the field of sociology. It describes networks as graph structure (Figure 10.1). Each

Fig. 10.1. Graph structure of network analysis.

node described as circle means a person or a company, and each link means the relation between people or companies. It is used to investigate the structures of networks, the effect of network structures on community members, and its mechanisms. Some methods for quantification are proposed. One of the representative quantification methods is degree. Degree means the numbers of links each node has. Especially, links which come into each node are called in-degree, and links which go out from each node are called out-degree. In this case in-degree is 3 and out-degree is 2. With considering each message sent to POC as node, I described a network. According to the ways how to link nodes, there are some possibilities to describe a network. For example, one message and a message replied to it can be linked to describe a network, and messages sharing same topics can be linked. Here, I adopted the latter way, that is, messages sharing two or more content words (almost of which were nouns) were linked. Among all messages (about 1530 messages), first 100 messages were used to make a graph structure (Figure 10.2). In the usual network analysis, each node repre-

10. How to Evaluate Social Intelligence Design

79

Fig. 10.2. Progress of number of logs collected into POC

sents each person or each company. But in this analysis, each node represents each message. This may be characteristic of our analysis, that is, analysis of knowledge creation2 . Figure 10.3 shows the network structure based on 100 of POC messages. Each square represents each message, and the numbers written in squares represent ID numbers of messages. Smaller the ID number is, earlier the corresponding message were sent to POC. Twenty nodes had no links to other messages (e.g., nodes 2, 7, 20), and some constructed very simple links (e.g., links of nodes 69 → 70, 33 → 40 → 44). Sixty-five messages constructed highly complex network. It is found that some nodes have many links and play cores, central roles in the network (e.g., nodes 53, 82, 97). Other nodes have only a few links and play peripheral roles in the network (e.g., nodes 5, 10, 99). The centrality of nodes can be quantified by degrees. Figure 10.4 shows degrees, in-degrees, and out-degrees for nodes. The average of degrees was about 9. On POC, members would communicate on multiple topics in a time. Probably this would lead smaller size of the average of degree. Other media, like BBS, people tend to debate one fied theme. It is expected that massages have a tendency to share more words and the average of degrees is larger. Out-degree decreased as the function of ID number, and in-degree increased. It was probabilistically reasonable. But some messages had larger degrees than this trend. Probably, we could regard such shifted messages as an index of the centrality. Five messages had in-degrees lager than the average plus 2 standard deviations and six messages had out-degrees larger than the average plus 2 standard deviations. These were larger than probabilistically calculated values (2.3) if messages were linked according to normal distribu2

There are only a few of researches which use a network analysis to describe knowledge representation. For example, Ferstl and Kintch described knowledge representations which people made when reading texts[10.3].

80

N. Fujihara

Fig. 10.3. Network analysis of POC

tion. These central nodes have opportunities to connect with themes which are originally unrelated to each other. It could be considered the number of such nodes reflects how tools facilitate knowledge creation in a network community. If so, POC would be regarded as an effective tool to support intelligence.

10.4 Future Works The analyses mentioned above were just a first step of our trials. So, we have a lot of issues to discuss as future works. It is necessary to compare the result with results of network analysis of other media like BBS. As POC

10. How to Evaluate Social Intelligence Design

81

Fig. 10.4. Degrees, in-degrees, and out-degrees for nodes.

is developing, results of network analysis of POC with additional functions should be compared with the results mentioned above, that is, POC with only basic function. Based on the network, there are other possible ways to investigate whether the system facilitate our knowledge creation. For example, based on the degrees messages could be classified into some clusters. If there were links which connected messages from different clusters, that may indicate the system facilitate our knowledge creation. Also, the numbers of links which connected chronologically separated messages may be one index of knowledge creation. But the network analysis would give us an interesting viewpoint to evaluate network communication tools. We have to elaborate to the method of applying the network analysis. It is expected that the analysis is an effective way to evaluate network communication tools and to investigate our knowledge creation.

References 10.1 Azechi, S., Fujihara, N., Sumi, K., Hirata, T., Yano, H., Nishida, T. (2000): Public Opinion Channel, Journal of Japanese Society for Artificial Intelligence, 15, 69–73.

82

N. Fujihara

10.2 Azechi, S., Fukuhara, T., Fujihara, N., Sumi, K., Matsumura, K., Hirata, T., Yano, H., Nishida, T. (2001): Public Opinion Channel — Toward KnowledgeCreating Community, Journal of Japanese Society for Artificial Intelligence, 16, 130-138. 10.3 Ferstl, E. C., Kintsch, W. (1999): Learning from text: Strucural knowledge assessment in the study on discourse comprehension. In: H. v. Oostendrop, S. R. Goldman (eds.) The construction of mental representations during reading, Lawrence Erlbaum Associates, N.J., Chapter 10, pp. 247-277. 10.4 Fujihara, N., Miura, A. (2000): The effect of the nature of a task on the strategy to search information from Internet. XXVII International Congress of Psychology, Stockholm, Sweden, July 23–28. (Abstract was published in International Journal of Psychology, 35 (3/4), 84) 10.5 Miura, A., Fujihara, N. (2000): Experimental study of searching strategy on World Wide Web. XXVII International Congress of Psychology, Stockholm, Sweden, July 23–28. (Abstract was published in International Journal of Psychology, 35 (3/4), 84) 10.6 Goleman, D. (1995): Emotional Intelligence. 10.7 Nishida, T. (ed.) (2000): Dynamic Knowledge Interaction, CRC Press LLC. 10.8 Watson, A., Sasse, M. A. (1998): Measuring percceived quality of speech and video in multimedia conferencing applications, Proceedings of ACM Multimedia ’98, 12-16 September ’98, Bristol, England, pp. 55-60. 10.9 Wilson, G. M., Sasse, M. A. (2000): The head or heart? Measuring the impact of media quality, CHI 2000 Extended Abstracts, pp. 117-118. 10.10 Yasuda, Y. (1997): Network analysis. Shin-yo Sha.

11. Overview Akira Namatame

AESCS-2001 The first international workshop on Agent-based Approaches in Economic and Social Complex Systems (AESCS) was initiated as a result of the growing recognition of the importance of the computational approaches to study complex economic and social phenomena. The fundamental objective of AESCS 2001 was to foster the formation of an active multi-disciplinary community on multi-agents,computational economics, social dynamics, and complex systems. The aim of AESCS 2001 was also to bring together researchers and practitioners from diverse fields, such as computer science, economics, physics,sociology, psychology, and complex theory for understanding emergent phenomena or collective behavior in economic and social systems. We also discussed on effectiveness and limitations of computational models and methods in social sciences. This workshop also intended to increase the awareness of researchers in many fields with sharing the common view that many problems economic and social systems will require collective informationprocessing with a large collection of autonomous and heterogeneous agents. The technical issues to be investigated include the follwoings: 1. Formal Theories on Agent-based Approaches – agent-based computational foundations – theories on rationality, intention, emotion, social action, social interaction – heterogeneity and diversity of agents 2. Computational Economics and Organization – agent-based economics – market-based computing – artificial markets – agents in financial engineering – econophysics – computational organization theory 3. Formal Theories of Social Dynamics – methodologies of modeling social behaviors – chaotic and fractal dynamics – dynamics of populations 4. Collective Intelligence – collective decision and behaviors – emergent intelligence T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 8 5− 8 7, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

86

A. Namatame

– social intelligence 5. Related Areas – evolutionary economics – complex theory – evolutionary computation – evolutionary games We could solicit many high quality papers which reflect the result of the growing recognition of the importance of the areas. All papers have receive a careful and supportive review, and we selected 13 papers out of 27 for the proceedings. We hope that as a result of reading the proceedings you will share with us the intellectual excitement and interest in this emerging discipline. Finally, we would like to acknowledge the support and encouragement of many peoples in helping us getting this new conference started.

General Chair Akira Namatame National Defence Academy, Japan ([email protected])

Program Committee: Yuji Aruka Chuo University, Japan ([email protected]) Kathleen Carley Carnegie Mellon University, USA ([email protected]) Shu-Heng Chen National Chengchi University, Taiwan ([email protected]) Sung-Bae Cho Yonsei University, Korea ([email protected]) Hiroshi Deguchi Kyoto University ([email protected])

11. Overview

Hitoshi Iba University of Tokyo, Japan ([email protected]) Kiyoshi Izumi Electrotechnical Laboratory, Japan ([email protected]) Masao Kubo National Defence Academy, Japan ([email protected]) Koichi Kurumatani Electrotechnical Laboratory, Japan ([email protected]) Thomas Lux Kiel University, Germany ([email protected]) Hideyuki Mizuta IBM Tokyo Research Laboratory, Japan ([email protected]) Azuma Ohuchi Hokkaido University, Japan ([email protected]) Hiroshi Sato National Defence Academy, Japan ([email protected]) Keiji Suzuki Future University-Hakodate ([email protected]) Keiki Takadama ATR, Japan ([email protected]) Hideki Takayasu Sony Computer Science Laboratory, Japan ([email protected]) Takao Terano University of Tsukuba, Japan ([email protected]) David Wing Kay Yeung Hong Kong Baptist University, China ([email protected])

87

12. Analyzing Norm Emergence in Communal Sharing via Agent-Based Simulation Setsuya Kurahashi and Takao Terano University of Tsukuba, Otsuka 3-29-1, Bunkyo-ku, Tokyo 113-0012, Japan

This paper describes an agent-based simulation study on the emergence of norms on information communal sharing. To carry out the study, we utilize our simulator TRURL, which (1) contains software agents with decision making and communication functions, and (2) has the capability to evolve artificial societies with specific characteristics defined by a given objective function to be optimized by genetic algorithms. Unlike the literature in social psychology research, which mainly applies evolutionary game theory to homogeneous agents for the simulation, TRURL focuses on the decision making behaviors of heterogeneous agents. Our experimental results have suggested that, contrary to the results of social psychology study so far, for information oriented properties, free riders in the society will not collapse the norm of communal sharing of the properties.

12.1 Introduction A norm in a society generally means expected behaviors of the members, decision criteria of the members, and/or the evaluation criteria that the society expects. Norm constitutes social pressures to conform people in a group. There are various levels and forms among public and private norms. Examples of such norms are (1) customs resulting from daily repeated behaviors, (2) morality as criteria of right and wrong, and (3) the law as public forces. In this paper, we will focus on a communal sharing norm By the communal sharing norm, we means that people share their resources together. Such sharing of resources plays an important role as a reciprocal norm in human behaviors. Communal sharing encourages us to maintain human relations and closeness [12.1]. The resources for communal sharing include money, physical properties, services, love, social approval, and information [12.2]. Recent rapid development of the Internet has widely changed our society characterized by information networks. Based on the viewpoint, this paper analyzes the birth, growth, and stability of communal sharing of information resources in a society. To carry out the study, we adopt an agent-based simulation model. Agentbased models can usually find macro phenomena from the interactions among agents. Although a model designer knows functions and natures of agents, (s)he doesn’t know what phenomena would happen as a whole during the simulation. Contrary, in the following aspects, our agent-based model is different from conventional macro models to analyze social phenomena. Our T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 8 8 − 98 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

12. Analyzing Norm Emergence in Communal Sharing

89

approach is characterized by the facts that: (1) the simulation model consists of heterogeneous agents, which have functions of decision-making and communication; (2) we observe emergence of social phenomena as a result of optimization of a social macro index by genetic algorithms; and (3) we analyze the emergent phenomena and characteristics of each agent. This paper is organized as follows: We first discuss several existing norm studies so far. Then, we briefly describe our simulator TRURL and apply it to the analysis on the communal sharing norm. Finally, we state the effectiveness of our agent-based simulation model.

12.2 Related Work on Studies of Norms Norms include personal norms and group norms. They can prevent someone from doing deviant behaviors through rewards and punishments in order to reduce tensions in a group. Norms urge people to conform to common judgments and behavioral patterns. Norms are predominant means to control a society and/or firms. We classify studies of norms into the following areas. Economic institution analysis. Economic institution analysis usually utilizes evolutional game theory. Researchers on the area have discussed the emergence and stability of diverse economic institutions [12.3, 12.4, 12.5]. Their basic technique, evolutional game theory analyzes economic institutions based on the concept of Evolutional Stable Strategy (ESS). Using the concepts, they have described the stability of economic institutions, the path dependency, and the complementarities of institutions. Aoki[12.6] has found two institutions of corporation systems as equilibrium points of the evolutional game. Their approach is applicable to analyze the emergence and stability of economic institutes about norms, however, they do not consider dynamic interactions among agents nor mutual understanding about agents’ inside models. Social network. In social network research, graph theory is often used. A center of an organization and a hidden relation among members are discussed using graph theoretic mathematical models [12.7]. In a network structure and a protocol analysis of electronic communities, socio-metric measures such as a degree of leadership existence have been proposed. They show birth, growth and maturity of norms in electronic communities. Social psychology and cognitive science. In social psychology, norms of human behaviors have been investigated with various data of psychological experiments. Processes to form norms have been analyzed experimentally in terms of leadership [12.8], the effects of group pressure [12.9], and the influence of a consistent minority [12.10]. Intolerant members who show an attitude of refusing resource sharing are critical barriers for the free riders. The experimental results have shown that

90

S. Kurahashi and T. Terano

those intolerant members are able to inhibit the emergence of the free riders. In summary, they have reported that (1) Intolerance is a stability condition of a norm, and (2) As a result that people choose adaptive behavior at a micro level, a communal sharing norm emerges at a macro level. These researches [12.11] [12.12] show us new viewpoints about group human behaviors that there is an evolutional process to adapt to an environment for a basis of forming a norm.

12.3 Artificial Society Model TRURL The roles of computer simulations in organization theory have been reevaluated in social science literature. However, many of the approaches seem to report too artificial results. To overcome such problems, we have developed a novel multi-agent-based simulation environment TRURL for social interaction analysis. – The agents in the model have detailed characteristics with enough parameters to simulate real world decision making problems. – Instead of manually changing the parameters of the agents, we evolve the multi-agent worlds using GA-based techniques. – Each agent exchanges knowledge and solves its own multi-attribute decision problems by interacting with the other agents 12.3.1 Agent Architecture Roughly, an agent in TRURL has event-action rules. Each agent exchanges knowledge and solves its own multi-attribute decision problems by interacting with the other agents. Predetermined parameters define the agents’ congenital characteristics. The parameters are not changed during one simulation, but are tuned by GA operations when the world evolves. pp = (cp , ps , pr , pa , pc , n, α, β, γ, δ, μ), where, pp is gene sequences, cp is physical coordinates, ps is probability of message sending, pr is probability of message reading, pa is probability of replying attitudes for pros-and-cons, pc is probability of replying attitudes for comment adding, δ is metabolic rate, μ is mutation rate of knowledge attribute values, α, β, and γ are parameters , and n is the number of knowledge attributes the agent has. These parameters represent characters of agents. The agent usually has some subset of knowledge only which the agent can use for decision-making. The knowledge the agent has is a set of knowledge attributes, which is defined as: Kd = {N, W, E, C}, where N is name of the knowledge attribute, W is importance weight of the attribute, E is evaluation value of the attribute; and C is credibility weight of the attributes.

12. Analyzing Norm Emergence in Communal Sharing

91

12.3.2 Communication and Action Energy A communication process can be considered as a decision-making process on the basis of conformity behavior. In this model, we define some parameters of knowledge attributes, which change when an agent receives a message. We show those definitions of parameters, which are weight w, evaluation e, and credibility c. As the result, knowledge of a high credible agent may affect a low credible agent. When both agents have the same tendency about some knowledge, their credibility increases each other. i = Δwkd



j i α(wkd − wkd ) · max(0, cjkd − cikd )

j∈S

Δeikd

=



β(ejkd − eikd ) · max(0, cjkd − cikd )

j∈S

Δcikd

=



γ((1 − 2 · |ejkd − eikd |) · max(0, cjkd − cikd ))

j∈S i wkd , eikd , cikd are weight, evaluation, credibility of agenti ’s knowledge attributes kd, α,β,γ are transfer ratio, S is a set of agents who send messages to agenti in period t. Action energy m, which is an acquired parameter increases in proportion to the amount of information that the agent has gotten. m is initialized in a random order by normal probability distribution. It decreases by metabolism δ when the agent send information to the other agent. On the other hand, if the agent receives valuable information from the other agent, it increases. It regularly decreases while it does not communicate others.

12.3.3 Inverse Simulation In a regular simulation method, we get results successively while the parameters are adjusted. The inverse simulation of TRURL gives an objective function at the beginning, then searches for parameters evolutionarily. We don’t adjust them intentionally. Accordingly we can know what nature or character of agents creates the organizational structure of the society after communication. Artificial society TRURL generates many societies with genetic algorithms, then it can recreate a similar society in terms of a social macro index. Each society is represented as genes of predetermined parameters of agents who constitute those societies. Those societies are evaluated with a social macro index after interactions among agents. Selection, crossover, mutation and reproduction are repeatedly carried out. The social architecture is gradually organized by a social index as an objective function.

92

S. Kurahashi and T. Terano

Social network researches have shown that the process of communication and opinion formation in a community can be measured with a socio-metric. If this socio-metric is the objective function of the artificial society, we can recreate the same phenomenon as a real society[12.16].

12.4 Experiments In this section, we describe experiments whether the sharing norm of information properties is stable or not. We constitute three kinds of society, and experiment about an amount of information, a free rider, an intolerant agent and information gap. 12.4.1 An Amount of Information in Each Society We design the following three artificial societies: 1. Face-to-Face communication oriented society (FFS) The communication among the agents is constrained by both the physical and mental coordinates. They interact with physical and mental neighborhoods. The ratio is parameterized. 2. E-Mail oriented society (EMS) The communication among the agents is constrained by the mental coordinates. In this society, agents interact each other one by one at each step. 3. Net-News oriented society (NNS) NNS is an extension of EMS. It has a virtual whiteboard at the center of The world. Agents in the world send messages to the whiteboard, and the whiteboard distributes the messages to all the agents. The credibility value of the messages is the same as the one of the senders. n i participate in We set one agent with a lot of information kd=1 eikd wkd each society. Figure 12.1 shows the change of information amounts in each society after 300 periods of communication. Y-axis is an average of all agents’ information amount and X-axis is the communication amount. The initial values of predetermined parameters are set to random. While the information amount of FFS changes slowly, the amount of EMS changes rapidly at some parts. The amount of NNS changes rapidly at a part, and then it is saturated at a burst. We consider that the cause is the restriction of information. In EMS, a credibility distance decides a receiver. If the society forms a crowded group temporarily, the agents will communicate rapidly in the group and vice versa. In NNS, if an agent sends a worthy message, credibility of a forum where the agent participates will increase and the agents will communicate rapidly.

12. Analyzing Norm Emergence in Communal Sharing

93

Fig. 12.1. The change of information quantity (300terms, 30agents/society)

12.4.2 Emergence and Collapse of a Norm A society with the common sharing norm is advantageous. Such an advantage is observed in the rapid increasing of information in NNS. One agent tends to send messages to Netnews, because the agent can get more worthy information in NNS than FFS. Netnews is thought of as equipment that shares information resource in network society. It appears as phenomena that agents approach to Netnews. Figure 12.2 shows the experiments in NNS.

Fig. 12.2. Occurrence and collapse of norm in Netnews society(left: after 10terms, right: after 50terms, 30agents/society)

The center rectangle represents Netnews. In the early stage of communication and interaction among agents, posting and acquiring information via Netnews increases rapidly, then the agents concentrate to the center (the left figure). This indicates that maintenance of a communal sharing norm is the advantage for each agent. In the second stage that agents communicate frequently, however, the agents leave from Netnews (the right figure). The cause is likely the uniformity of knowledge. It is difficult for agents to get new information in this stage.

94

S. Kurahashi and T. Terano

It shows that if free riders that pay no cost for posting messages increase, Netnews will lose its worth. 12.4.3 Emergence and Control of Free Riders Figure 12.3(the left figure) shows the change of average send-gene of all agents in FFS. It is one of predetermined parameters. Y-axis is average sending probability. X-axis is the amount of communication. Sending probability decreases slowly. It demonstrates free riders emergent in FFS. In FFS simulation results, we have also found the same phenomenon in EMS. An agent loses the energy for communication gradually, while it gets the energy for worthy information. The free rider can live forever because it doesn’t send and only gets information. As the result, the amount of sending messages decreases, and agents who have the sharing norm lose the energy. Because they can’t get worthy information, though they expect rewards as sharing for sending messages. Eventually all of them would go away.

Fig. 12.3. The change of send-gene and the effect of tolerant agents. Left is increasing of free riders. Right is controling of free riders (1000 terms, Average of 30 agents/society)

We extended the model not to send a message if the agent is a free rider. The result of the experiment is shown in Figure 12.3(the right figure) . Free riders can’t get more information, and they lose their energy. As the result, decrease of send-gene is controlled in the society. So existence of intolerance agents can control free riders without an explicit punishment. It demonstrates the reason why implicit norms exist except explicit norms such as the law.

12. Analyzing Norm Emergence in Communal Sharing

95

12.4.4 Information Gap The information gap among agents and efficiency of information acquisition can be examined using Inverse simulation of TRURL, We can know the nature of information rich agents. The information gap can be measured with Gini index. Gini index is a sample statistic in economic categories and represents an income gap. The larger Gini index values means the more income gaps, that is, there are the more difference of incomes among the rich and the poor.

Fig. 12.4. Max Gini factor of FFS (100 terms, 300 generations, 20 societies , 30 agents/society)

 ZGini = 1 −



Ek − Ei )Ai ) (Etot Atot )

i=1..N ((2

k=1..i

sort data: Ei/Ai > Ei−1 /Ai−1 , Ai : ”people” (population  in groupi ), Ei: ), A = ”wealth” (the amount of information in group i tot i=1..N Ai , Etot =  E i=1..N i As shown in Figure 12.4.4, we simulated 20 societies with Gini factor as an objective function The results are that maximum Gini factors are 63% in FFS, 54% in EMS and 48% in NNS. It shows a relation as F toF > Email > N etnews(the upper part of Table 12.1). Netnews society has less information gap than the other societies. We have observed the genes of the information rich agents in each society. In FFS, the rich agent is to send many messages (Probably 0.75) and to read them frequently (Probably 1.0). In NNS, the rich agent is to send few messages (Probably 0.20) and to read them occasionally (Probably 0.92). These results suggest the following hypotheses: In FFS, the active agent, which gathers information by itself sends the information and listens to other agents frequently, can become the information rich. In NNS, the Net surfer

96

S. Kurahashi and T. Terano

Table 12.1. The gap and nature of information rich persons / Difference of energies in each society Max Gini index Sending probability of the rich Receiving probability of the rich Max energy Max energy ratio

FFS

EMS

NNS

63% 0.75 1.0 54 1.0

54% 0.63 0.94 77 1.4

48% 0.20 0.92 165 3.1

agent, which sends few messages, can become the information rich. It reads information on the Net instead of gathering information spending costs, EMS is seated at the midpoint. n In addition, we used the following objective function for the simulation. i=1 mi : mi isactionenergyof agenti This maximizes the amount of action energy, which is the difference between information value and gathering cost. It indicates that the agents communicate their information efficiently and represents the efficiency of the society to gather information. The result is shown in the under part of Table 12.1. NNS has the ability to gather information 3.1 times as much as FFS. Although the information rich agents exist in NNS, the information gap is less than other societies. 12.4.5 Discussion When we observe the change of information in three societies, the following results are suggested: NNS has big communication ability, a free rider occurs in any society, an intolerance agent can control free riders, and the information gap in NNS is the smallest. From the viewpoint of the communal sharing norm, the experimental results have implies the following items: Information property has a different nature from physical resources in terms of sharing. Sharing and distribution of information don’t mean to reduce their property values. Netnews, which is an equipment to share information, can control free riders and reduce the information gap in NNS. On the contrary, agents, which don’t participate to the Netnews society, might expand the information gap. Digital Divide might be one of these phenomena. So the results may persuade to change a definition of a free rider in NNS. Before the experiments, our hypothesis was that the communal sharing norm would easily collapse and increase free riders in NNS such as an advanced information society, because the society wouldn’t have severe morality like punishments and intolerance. However, the simulation results have refused the hypothesis. The information gap didn’t expand more than our prediction in NNS. Although free riders emerged in NNS, they didn’t collapse the

12. Analyzing Norm Emergence in Communal Sharing

97

norms. Then we assumed that a manager of venture type easily gets richer than a manager of traditional type. However, the former types have become richer than the latter types.

12.5 Conclusion This paper has described agent-based simulation and their experiments about a communal sharing norm. We have simulated several evolutional artificial societies with multiagents. We have observed the emergence, collapse and control of norms in FFS, EMS and NNS. Using TRURL, we could analyze the nature of social interactions in the artificial world. Wwe have also demonstrated that the technique of Agent-based Simulation could contribute to resolve organizations and social phenomena.

References 12.1 D. T. Regan. Effects of favor and linking on compliance. Journal of Experimental Social Psychology, Vol.7, pp. 627-639, 1971. 12.2 U. G. Foa. Interpersonal and economic resources. Science, Vol.171, pp. 345351, 1971. 12.3 M. Aoki. Information Incentives and Bargaining in Japanese Economy. Cambridge University Press, 1988. 12.4 H. Young. The Evolution of Conventions. Econometrica, Vol.61, pp. 57-84, 1997. 12.5 M. Kandori, G. Mailath, and R. Rob. Learning, Mutation, and Long Run Equilibria in Games. Econometrica, Vol.61, pp. 29-56, 1997. 12.6 M. Aoki and M. Okuno, Editors Comparative Institutional Analysis: A New Approach to Economic Systems. University of Tokyo Press, 1996. 12.7 Granvetter, M. S., The Strength of Weak Ties. American Journal of Sociology, Vol.78, pp1360-1380, 1973. 12.8 F. E. Fiedler. Personality and Situational Determinants of Leadership Effectiveness. Group Dynamics, 3rd ed, Harper & Row, 1967. 12.9 Asch, S. E., Effects of Group Pressure upon the Modification and Distortion of Judgments, Group, Leadership and Men, Carnegie Press, 1951. 12.10 Moscovici, S., Naffrechoux, M., “Influence of a Consistent Minority on the Responses of a Majority in a Color Perception Task”, Sociometry, Vol.32, 1969. 12.11 Axelrod, R., The Evolution of Strategies in the Iterated Prisoner’s Dilemma, The Complexity of Cooperation, Princeton University Press, 1997. 12.12 Takezawa, M., Kameda, T., “Ownership and Sharing: Exploring Social Foundations of Communal Sharing Norm by Evolutionary Game Analysis”, Cognitive Studies, Vol.6, No.2, pp. 191-205, 1999. 12.13 Carley, K. M., Gasser, L., Computational Organization Theory, Weiss,G(Eds), Multiagent Systems, The MIT Press, 1998.

98

S. Kurahashi and T. Terano

12.14 Mitchell, Melanie., An Introduction to Genetic Algorithms, The MIT Press, 1996. 12.15 Terano, T., Kurahashi, S., and Minami, U., “Trurl: Artificial world for social interaction studies”, Artificial Life,pp.326-335, 1998. 12.16 Kurahashi, S., Minami, U., Terano, T., “Inverse Simulation for Analyzing Emergent Behavior in Artificial Societies”, Transactions of the Society of Instrument and Control Engineers, Vol.36,No.2,pp.1454-1461, 1999.

13. Toward Cumulative Progress in Agent-Based Simulation Keiki Takadama and Katsunori Shimohara ATR International, Information Sciences Division 2-2-2 Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0288 Japan {keiki,katsu}@isd.atr.co.jp Abstract. This paper stresses the importance of focusing on modeling processes in order to make cumulative progress in agent-based approaches. In this paper, we introduce our approach to analyzing modeling processes and investigate its possibilities toward cumulative progress. The capabilities of our approach can be summarized as follows: (1) our approach has great potential to promote cumulative progress in agent-based approaches; and (2) the elements found by our approach have high possibilities of affecting the real world, being utilized as tool-kits, and supporting the KISS principle. Keyword: agent-based approach, computational simulation, cumulative progress, modeling process

13.1 Introduction An agent-based approach can provide techniques and tools for analyzing complex organizations and social phenomena. This approach explicitly examines organizing processes and social dynamics and builds theories by clarifying vague, intuitive, or under-specified issues in conventional approaches. Although research on agent-based approaches has recently attracted much attention, the approaches actually have a long history. Major examples originally included garbage can model [13.10], iterated prisoner’s dilemma (IPD) [13.2], multiagent soar [13.6], and Virtual Design Team (VDT) [13.17]. Following these models, several others are proposed such as sugarscape [13.13], ORGAHEAD [13.8], PCANS 1 [13.15], simulating society [13.14], and agent-based computational economics (ACE) [13.24]. These several models and methods contributed to our understanding of complex organizations and social phenomena. However, in Cohen’s phrase, “disciplines or fields of study do not get much progress due to a lack of cumulative progress in agent-based approaches [13.12].” This indicates that agent-based approaches do not compel new investigators to build on the accomplishments of older works, even though these previous works provided a ∗ 1

Paper submitted to Exploring New Frontiers on Artificial Intelligence, Series on Advanced Information Processing (AIP), Springer This model is currently extended to PCANSS.

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 99− 109, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

100

K. Takadama and K. Shimohara

lot of useful results and showed high potential to understand other important issues in organizational and social science. So, what is a main cause of this problem? How do we overcome this problem? Unfortunately, these questions are left behind the colorful and powerful simulations. Since agent-based approaches cannot avoid tackling these questions, this paper aims to summarize the factors that prevent cumulative progress in agent-based approaches and shows that our approach offers possibilities of assisting cumulative progress in agent-based approaches. This paper is organized as follows. Section 13.2 starts by describing the factors that prevent cumulative progress in agent-based approaches, and Section 13.3 explains our approaches toward cumulative progress. The potential and capabilities of our approach to promoting cumulative progress are discussed in Section 13.4, and our conclusions are finally made in Section 13.5.

13.2 Can We Assist Cumulative Progress? 13.2.1 Problems in Agent-Based Approaches In the previous section, we pointed out that the main problem of agent-based approaches is the lack of cumulative progress. So, what has caused this lack of cumulative progress? There are many reasons. According to Cohen, “A lack of mathematical tools is a part of the problem. But, there are many other problems, including the way we train our students and evaluate research projects, with too little emphasis on building on what is known, and too much emphasis on novelty and on the promise of more powerful computation [13.12].” 13.2.2 Points for Cumulative Progress Toward overcoming the above difficulties, it is useful to enumerate the points that promote cumulative progress in agent-based approaches. Considering Cohen’s claim, the following solutions offer significant toward cumulative progress. – (a) Common test-beds: First, sharing common test-beds is a promising approach for cumulative progress. The reasons are summarized as follows: (1) common test-beds enable researchers to narrow an argument down to concrete and detailed issues, which help to providing a fruitful and productive discussion; and (2) common test-beds encourage researchers to share results, which leads to progress in the field by comparing results or competing with other researchers. – (b) Standard computational models: Next, standard computational models are necessary for cumulative progress. This is because (1) researchers do not need to design computational models, which contribute to

13. Toward Cumulative Progress

101

bringing several researchers together toward a progress in the field; and (2) common parts of various research efforts become clear through the development of libraries of computational models, which provides the essential parts of agent-based simulations. – (c) Validation and advance of older works: Third, it is important to validate older results and advance older works for cumulative progress. In this case, the replication of older models is essential to validate and advance older works. To promote this, researchers should share and understand what were done and what are not in agent-based approaches. – (d) Standard evaluation criteria: Finally, standard evaluation criteria for results (including papers and projects) are indispensable for cumulative progress. Although it is difficult to evaluate results appropriately, it is important to apply the same evaluation criteria. For instance, a benchmark in evaluation criteria would be useful for cumulative progress. In addition to the above points, the following points are also important to promote cumulative progress, though they are not restricted to agent-based approaches: (1) regular meetings that enable researchers to constantly share results; and (2) appropriate teaching of students. 13.2.3 Cumulative Progress in Current Projects Based on the above four points, this subsection analyzes how current agentbased research can achieve these goals. For common test-beds, the U-Mart project [13.19], for instance, is recognized as a common test-bed for a virtual stock market in the economic field. Although this project began a few years ago, it has promoted cumulative progress by narrowing arguments down to a concrete and detailed stock market and by sharing the results among researchers. For standard computational models, Axelrod and his colleagues developed standard computational models by employing their existing models (such as garbage can model and iterated prisoner’s dilemma (IPD)) [13.3]. Kurumatani is also developing libraries of standard parts in World Trade League [13.16], which aims to provide a multiagent-based universal environment for analyzing economic and financial systems. Although these efforts have promoted cumulative progress, conventional agent-based approaches have not so far fully address the four points.

13.3 Exploring Key Elements Since it is not easy to promote cumulative progress in agent-based approaches as described in the previous section, this paper starts by investigating how our approach [13.21] can promote such cumulative progress as the first stage of our research.

102

K. Takadama and K. Shimohara

13.3.1 Interpretation by Implementation Outline. The Interpretation by Implementation (IbI) approach is a trial and error method for seeking underlying elements of organizations or societies through a process of continuing the implementation and interpretation phases in turn. The concrete algorithm of the IbI approach proceeds as follows. 1. First, the IbI approach implements a model (i.e., model A in Figure 13.1) while focusing on a modeling process. In this stage of the IbI approach, the following three processes are employed: (a) concept breakdown, (b) assumptions/premises modification, and (c) investigating layers change (all three detailed process will be described later). 2. Next, the IbI approach interprets results to investigate the underlying elements that determine the characteristics of multiagent organizations or societies. 3. If the essential elements are found, then this process is finished. If not, new models (i.e., models B, C, · · · in Figure 13.1) are implemented to investigate other elements; then, goto 2.

B r e a k d o w n /A s s u m p tio n /L a y e r

In te r p r e ta tio n

In te r p r e ta tio n Im p le m e n ta tio n

M o d e l A M o d e l B M o d e l C

M o d e l ? Fig. 13.1. Interpretation by Implementation

What is important to note here is that the IbI approach focuses on the influence of elements embedded in a modeling process on results. Since these elements have a big influence on results, we must consider such an influence when employing agent-based approaches. However, it is difficult to visualize these elements, and thus the IbI approach employs a trial and error method that explores essential elements by changing them. Elements Embedded in a Modeling Process. From the previous section, the important point is to decide what kinds of modeling processes we should focus on. In this stage of the IbI approach, the following three processes are employed for the following reasons. Note that we never claim that the following three processes are sufficient for finding the underlying elements in

13. Toward Cumulative Progress

103

organizations or societies. Other viewpoints can be considered in addition to the following processes. – Concept breakdown: When implementing a concept in a computational model, we must clarify abstract parts by breaking the concept down into detailed and operationalized parts from the computational viewpoint. Since characteristics of multiagent organizations or societies change depending on such a breakdown process, key elements are likely to be embedded in this modeling process. – Assumptions/premises modification: We tend to implement computational models under assumptions or premises that are generally set unconsciously. This tendency increases as we concentrate on investigating issues. However, such assumptions or premises have a high possibility of being key elements because the results drastically change by varying assumptions or premises. – Investigating layers change: When investigating characteristics of organizations or societies, some of the characteristics are found in a certain layer while others may be found in another layer. This indicates that a change in the layer for an investigation has the potential of finding new key elements that affect the characteristics of organizations or societies. 13.3.2 Applications of IbI Approach This section briefly describes three applications of the IbI Approach. Concept Breakdown. Organizational learning(OL)[13.1, 13.11] is roughly characterized as organizational activities that solve problems that cannot be solved at an individual level, and it has a large influence on the characteristics of organizations. However, the concept of OL can be implemented (broken down) in many ways from a computational viewpoint. Focusing on this feature, we found that the following three elements affect the characteristics of multiagent organizations through breaking the concept of OL down in a certain way [13.20]: (1) the independence of learning mechanisms; (2) the execution order of learning mechanisms; and (3) the combination of exploration at an individual level and exploitation at an organizational level. These implications can be revealed through the implementation of a concept breakdown and the interpretation of simulation results. Assumptions/premises modification. As shown in the typical example of the prisoner’s dilemma [13.4], agents are roughly divided into the following two categories: (1) the selfish or competitive type and (2) the altruistic or cooperative type. This classification effectively distinguishes goals of agents at individual levels from those at organizational levels. However, we found that an evaluation of agents affected the characteristics of multiagent organizations more than the goals of agents [13.22]. This implication cannot be revealed from a goal-related perspective but through the implementation of

104

K. Takadama and K. Shimohara

varying premises by adding an evaluation perspective and the interpretation of simulation results. Investigating layers change. One of the important problems in an organization is solving the trade-off between exploration and exploitation [13.18]. To address this issue, we focused on the fact that the trade-off between exploration and exploitation is not embedded in one layer but found in several layers. Then, we found that a certain problem-specific trade-off could contribute to solving the fundamental trade-off between solutions (related to exploration) and costs (related to exploitation) [13.23]. This implication cannot be revealed by only considering fundamental trade-offs but through the implementation of a framework that provides an investigation of other tradeoffs and the interpretation of simulation results.

13.4 Discussion 13.4.1 Cumulative Progress First, we discuss how our approach has the potential to promote cumulative progress in agent-based approaches. As mentioned in Section 13.2, the following four points are important for cumulative progress: (a) common test-beds; (b) standard computational models; (c) validation and advance of older works; and (d) standard evaluation criteria. Although conventional agent-based approaches do not encourage researchers to fully address the four points, our approach tackles them as follows. – (a) Common test-beds: Since factors and assumptions embedded in a modeling process are mostly general, researchers not only can share results but also utilize factors. This indicates that our approach does not require common test-beds to share and utilize results. This advantage does not force researchers to adjust their ideas or methods to common test-beds. – (b) Standard computational models: Factors and assumptions embedded in a modeling process are kinds of common parts in simulation models. Therefore, standard computational models can be developed by combining several kinds of factors and assumptions. – (c) Validation and advance of older works: Factors and assumptions embedded in a modeling process are simple because they can be divided into each element. From this feature, it is easy to replicate older models if the factors and assumptions of older models are analyzed in advance, and such replication encourages researchers to validate older results. Furthermore, researchers have the chance to advance older works by simply adding and removing factors and assumptions. – (d) Standard evaluation criteria: Since factors and assumptions embedded in a modeling process are independent from addressed issues, researchers can concentrate on evaluating the essential degree of these elements. For instance, we measure such factors and assumptions in terms

13. Toward Cumulative Progress

105

of degrees of influence on results, simplicity of implementation, and so on. These degrees in evaluation criteria can be considered as bench-marks in agent-based approaches. From the above analysis, our approach has great potential to promote cumulative progress in agent-based approaches. However, we should not neglect the following point to effectively receive the advantages of our approach: it is important to (1) store a lot of factors and assumptions embedded in a modeling process and (2) systematize these elements in advance for easy utilization. If the above points are achieved, our approach enables us to understand what has been done and what remains to be done in agent-based approaches by simply investigating the repository of underling factors and assumptions. 13.4.2 Potential of Our Approach Next, we discuss the potential of our approach in terms of the following viewpoints: (1) linkage to real world; (2) tool-kits; and (3) the KISS principle. Linkage to Real World. Linkage to the real would is one of the major problems in agent-based simulations. Even though many useful implications can be found in computational simulations, we cannot guarantee that these implications are valid in the real world. For this problem, Axelrod answered in his book as follows: “Although agent-based modeling employs simulation, it does not aim to provide an accurate representation of a particular empirical application. Instead, the goal of agent-based modeling is to enrich our understand of fundamental process that may appear in a variety of application. [13.4].” Carley, who proposed the concept of computational organization theory (COT) [13.7], responded as follows: “Human organizations can be viewed as inherently computational because many of their activities transform information from one form to another, and because organizational activity is frequently information driven [13.9]”. This assertion supports the effectiveness of computational analysis. Concerning our approach which focuses on factors and assumptions embedded in a modeling process, these elements offer potential power to affect the real world. This is because the elements have a large influence on results even when they slightly change. Although simulation results do not follow the real world because the real world includes several kinds of unexpected factors and the observed phenomena only show one aspect of the real world, our approach can identify essential keys that affect the real world. Tool-kits. Recently, a lot of agent-based simulators, including Swarm,2 have been proposed and these have contributed to understanding complex organizations and social phenomena. However, the following important problems still remain: (1) agent-based simulators are mostly useful for visualization 2

See http://www.swarm.org for details.

106

K. Takadama and K. Shimohara

tools, not for computational simulation tools. This is because we have to design essential parts of simulations such as internal models of agents. (2) Agent-based simulators are often built for specific issues. Researchers also build their own tools instead of using tools built by others, and thus it is difficult to share the same agent-based simulators. These two problems clearly prevent cumulative progress in agent-based approaches. To overcome these problems, Axelrod devoted himself to developing general tools for agent-based approaches [13.3], but he finally gave it up, because most tools for social and organizational simulations have to be designed for specific tasks, and thus few parts can be shared or applied to other models [13.5]. In comparison with the above conventional agent-based simulators, our approach has the capability of extracting common parts of simulations by exploring factors and assumptions embedded in a modeling process. Since these common parts are mostly related to fundamental parts of an agent design and are not specific to addressed issues, they can be used as tool-kits. This indicates that our approach provides general tool-kits that are difficult to find by developing domain-specific tool-kits. KISS Principle. The KISS principle3 proposed by Axelrod claims that simple models should be implemented to understand the fundamental processes in organizational or social phenomena [13.4].4 This suggestion implies that one can be confident of understanding results by knowing everything that went into the model. Note that the KISS principle does not merely claim to make everything simple but also to leave essential parts by removing nonessential ones. Based on this claim, one important question remains: how do we figure out the essential parts? According to Axelrod, one method is to conversely derive the essential parts by investigating results and facts [13.5]. However, this derivation requires good sense, and it is neither easy nor an application of the scientific method. In comparison with this situation, the factors and assumptions found by our approach have high possibilities of being essential parts because these elements change the characteristics of multiagent organizations or societies. Of course, all elements are not required to implement models, but it is significant to consider such elements as candidates before implementing models. From this advantage, our approach offers great potential to support the KISS principle in terms of finding essential parts. 3 4

This principle stands for the army slogan keep it simple, stupid. Strictly, he pointed out that assumptions underlying the agent-based model should be simple and also claimed that the complexity of agent-based modeling should be in the simulated results, not in the assumptions of the model.

13. Toward Cumulative Progress

107

13.5 Conclusions This paper stressed the importance of focusing on modeling processes toward achieving cumulative progress in agent-based approaches. In particular, this paper suggested that the analysis of modeling processes can help to find elements that directly affect the characteristics of multiagent organizations or societies. Furthermore, we also showed that these elements were useful for an alternative understanding of complex organizations or social phenomena. By investigating the capabilities of our approach, we found the following two implications. First, our approach has great potential to promote cumulative progress in agent-based approaches in terms of (a) common test-beds, (b) standard computational models, (c) validation and advance of older works, and (d) standard evaluation criteria. Second, the elements found by our approach offer the high possibilities of affecting the real world, being utilized as tool-kits, and supporting the KISS principle. However, this paper only discussed the high potential of our approach for cumulative progress and did not prove them in the real would. Furthermore, this paper did not specify the range in which our approach showed its effectiveness. These should be addressed in the near future. We also have to investigate when and what elements found by our approach should be considered for particular situations. Acknowledgements. The authors wish to thank Prof. Cohen and Prof. Axelrod, both from the University of Michigan, for their useful comments via private e-mails, and Prof. Carley from Carnegie Mellon University for helpful discussions.

References 13.1 Argyris, C. and Schon, D. A. (1978): Organizational Learning, AddisonWesley. 13.2 Axelrod, R. M. (1984): The Evolution of Cooperation, BasicBooks. 13.3 Axelrod, R. M. (1997a): “Advancing the Art of Simulation in the Social Sciences”, in R. Conte, R. Hegselmann, and P. Terna (Eds.), Simulating Social Phenomena, Springer, pp. 21–40. 13.4 Axelrod, R. M. (1997b): The Complexity of Cooperation: Agent-Based Models of Competition and Collaboration, Princeton University Press. 13.5 Axelrod, R. M. (2000): Private communication. 13.6 Carley, K. M., Kjaer-Hansen, J., Prietula, M., and Newell, A. (1992): “A prolegomenon to Artificial Agents and Organizational Behavior”, in M. Masuch and M. Warglien (Eds.), Distributed Intelligence: Applications in Human Organization, pp. 87–118, Elsevier Science Publications. 13.7 Carley, K. M. (1995): “Computational and Mathematical Organization Theory: Perspective and Directions”, Computational and Mathematical Organization Theory, Vol. 1, No. 1, pp. 39–56.

108

K. Takadama and K. Shimohara

13.8 Carley, K. M. and Svoboda, D. M. (1996): “Modeling Organizational Adaptation as a Simulated Annealing Process”, Sociological Methods and Research, Vol. 25, No. 1. pp. 138–168. 13.9 Carley, K. M. and Gasser, L. (1999): “Computational and Organization Theory”, in Multiagent Systems – Modern Approach to Distributed Artificial Intelligence –, G. Weiss (Ed), The MIT Press, pp. 299–330. 13.10 Cohen, M. D., March, J. G., and Olsen J. P. (1972): “A Garbage Can Model of Organizational Choice”, Administrative Science Quarterly, Vol. 17, pp 1– 25. 13.11 Cohen, M. D. and Sproull, L. S. (1995): Organizational Learning, SAGE Publications. 13.12 Cohen, M. D. (2000): Private communication. 13.13 Epstein, J. M. and Axtell, R. (1996): Growing Artificial Societies: Social Science form the Bottom Up, Brooking Institution Press. 13.14 Gayload, R. and D’Andira, L. J. (1998): Simulating Society: A Mathematica Toolkit for Modeling Socioeconomic Behavior, Springer-Verlag. 13.15 Krackhardt, D. and Carley, K. M. (1997): “PCANS Model of Structure in Organizations”, The 1997 International Symposium on Command and Control Research and Technology. 13.16 Kurumatani, K. and Ohuchi, A. (2001): “World Trade League: Standard Problems for Multi-agent Economics (1) – Concept and Implementation of X-Economy System”, Meeting of SIG-ICS (Special Interest Group on Intelligence and Complex System) of IPSJ (Information Processing Society of Japan), 2001–ICS–123, pp. 55–60, (in Japanese). 13.17 Levitt, R. E., Cohen, G. P., Kunz, J. C., Nass, C. I., Chirstiansen, T., and Jin, Y. (1994): “The Virtual Design Team: Simulating How Organization Structure and Information Processing Tools Affect Team Performance”, in Carley, K. M., and Prietula, J. (Eds.): Computational Organization Theory, Lawlence-Erlbaum Assoc., pp. 1–18. 13.18 March, J. G. (1991): “Exploration and Exploitation in Organizational Learning”, Organizational Science, Vol. 2, No. 1, pp. 71–87. 13.19 Shiozawa, Y. (1999): “Virtual Market as a Common Test Bed – for the construction of a ”robo-cup” in economics –” The 1999 JAFEE (Japan Association for Evolutionary Economics) Annual Meeting, pp. 253–256, (in Japanese).5 13.20 Takadama, K., Terano, T., Shimohara, K., Hori K., and Nakasuka, S. (1999): “Making Organizational Learning Operational: Implication from Learning Classifier System”, Computational and Mathematical Organization Theory (CMOT), Kluwer Academic Publishers, Vol. 5, No. 3, pp. 229–252. 13.21 Takadama, K., Terano, T., and Shimohara, K. (2000): “Interpretation by Implementation for Understanding Multiagent Organization”, The CASOS (Computational Analysis of Social and Organizational System) Conference 2000, pp. 157–160.

5

The name V-Mart changed to U-Mart.

13. Toward Cumulative Progress

109

13.22 Takadama, K. and Shimohara, K. (2001a): “What Kinds of Properties Determine Characteristics of Multiple Learning Agents? ∼ Implications from goal and evaluation in agents ∼,” The International Workshop on Autonomy Oriented Computation (AOC’01) at the 5th International Conference on Autonomous Agents (Agents’01), pp. 21–30. 13.23 Takadama, K. and Shimohara, K. (2001b): “Exploration and Exploitation Trade-off in Multiagent Learning,” The 4th International Conference on Computational Intelligence and Multimedia Applications (ICCIMA’01), to appear. 13.24 Tesfatsion, L. (2001): “Introduction to the Computational Economics, Special Issue on Agent-Based Computational Economics”, Computational Economics, to appear.

14. Complexity of Agents and Complexity of Markets Kiyoshi Izumi Cyber Assist Research Center, AIST and PREST, JST 2-41-6 Aomi, Koto-ku, Tokyo 135-0064, Japan, [email protected], http://www.carc.aist.go.jp/∼kiyoshi/index.html

In this study we rethought efficient market hypothesis from a viewpoint of complexity of market participants’ prediction methods and market price’s dynamics, and examined the hypothesis using simulation results of our artificial market model. As a result, we found the two differences from the hypothesis. (a) Complexity of markets was not fixed, but changed with complexity of agents. (b) When agents increased the complexity of their prediction methods, structure of dynamic patterns of market price didn’t disappear, but it can’t be described by equation of any dimensions.

14.1 Introduction Are you surprised if the performance of financial specialists’ forecasts is the same as that of randomly generated forecasts? In the field of economics, the theory of financial markets called the efficient market hypothesis was proposed in the 70s, and it has caused many arguments till today. By this hypothesis, the movement of the price of financial markets is a random walk, and cannot be predicted. Therefore, the performance of all the forecasts is the same. Theories of financial engineering, which developed greatly today, are based on this hypothesis, and they assume financial prices as the stochastic process. Although many statistical verification of the hypothesis was performed using actual data, since the hypothesis included a market participant’s expectation formation, it has not been verified directly. In recent years, however, the artificial market approach which builds a virtual market model and performs a simulation into a computer appeared, and researches in this approach try to verify the hypothesis directly[14.1, 14.2, 14.3]. This study rethinks the efficient market hypothesis from the new viewpoint of the relation between the complexity of market participants’ prediction formulas and the complexity of the movement of a market price. And this study examines the hypothesis from the simulation result using the artificial market model.

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 110− 120, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

14. Complexity of Agents and Complexity of Markets

111

14.2 The Efficient Market Hypothesis Seen from Complexity The main points of the efficient market hypothesis are summarized as follows. – Each market participant of a financial market takes in very quickly and exactly all the information related to the movement of a market price, and uses it for price expectation. – The market price that determined by the dealings between such market participants is reflecting properly all the relevant information that is available at present. – Therefore, there is no room for a certain person to find out the new relation between a market price and the available information, and to become advantageous from other persons. That is, the movement of a market price becomes a random walk driven only by new information, and nobody can predicte it. When the above-mentioned main points are recaught from the viewpoint of complexity, the efficient market hypothesis contains the following things implicitly. – In order to take in suitable information, each market participant is going to complicate his prediction formula by learning, and is going to hold the structure of the determination formula of the market price. – The structure of a price determination formula is fixed and independent of the learning of market participants. Finally the market participants detect the structure, and it will disappear. That is, the efficient market hypothesis needs the two premises: (a) the independence of the complexity of the movement of a market price from the complexity of each market participant’s prediction formula and (b) the existence of motivation of leaning by each market participant. On the other hand, by the artificial market simulation, de la Maza[14.4] found that when the dimension of market participants’ prediction formula went up from 0 to 1, the movement of a market price also changes from a random walk to linearity. That is, he showed the possibility the complexity of market participants and the complexity of a market are not independent. Then, what is the motivation to which each market participant complicates his prediction formula? Joshi et.al.[14.5] think that it is because the situation similar to the prisoner’s dilemma game has occurred. In their artificial market model, taking in the technique of the moving average of a technical analysis to a prediction method, and raising the dimension of a prediction formula from 0 to 1 corresponds to the default strategy of the prisoner’s dilemma game. On the other hand, not using a technical analysis for prediction corresponds to the cooperation strategy. From the simulation result, the two following conditions for becoming a prisoner’s dilemma situation were seen.

112

K. Izumi

Condition 1. If one raises his prediction dimension, his prediction becomes more accurate and the profit of his dealings result increases. Thus, the motivation of the default strategy exists. Condition 2. However, when everybody raised the dimension, the movement of the market price became more complicated, and the prediction accuracy has fallen rather than the time of everybody’s not using the technical analysis. Thus, since everybody raised the dimension of his prediction formula in pursuit of profits, the prediction accuracy becomes worse than before. In the following sections, by the artificial market simulation, we analyze the complexity of a market and the prisoner’s dilemma situation when a prediction dimension becomes larger.

14.3 Artificial Market Model The artificial market is a virtual financial market with 50 virtual dealers (agents) in a computer. One financial capital and one non-risk capital exist in this artificial market. Each agent expects the movement of the financial price, and he changes the position of the financial and non-risk capital so that the utility of his expected profit may become the maximum. In the artificial market, one term consists of four step of expectation, an order, price determination, and learning, and time progresses discretely by repeating these four steps. 14.3.1 Expectation Each agent expects the change value of the financial price of this term using the weighted sum of the change value of past financial price. That is, in this study, since fundamentals information does not exist in a market, the agents expect the change value of the financial price only by the technical analysis. The expectation formula of each agent is auto the regressive integral moving average model ARIMA(n, 1, 0), where n means the number of the terms of the price changes used for expectation. The larger n is, the larger the dimension of an expectation formula is. Thus in this study, n is regarded as the complexity of each agent’s expectation. The expectation formula is as follows, when Pt is the financial price of this term which is not yet determined and y˜t is the expectation the change of financial price (Pt − Pt−1 ).

y˜t = =

n 

bi yt−i + et

i=1 xt bt

+ et

(14.1)

14. Complexity of Agents and Complexity of Markets

113

Here, et is the normal distribution whose average is 0 and standard deviation is 0.1, bt is a vector with the coefficient of the prediction formula1 , (b1 , · · · , bn ) , and xt is a vector of the explanation variables of the prediction formula, i.e., the past price changes2 , (yt−1 , · · · , yt−n ) . 14.3.2 Order It is assumed that each agent has the utility function of expected profit with risk avoidance. Then the optimum quantity of the position of the financial capital with the maximum utility, qt∗ , is proportional to the expected change value yt of the formula (14.1). qt∗ = ayt ,

(14.2)

where a is a coefficient. Each agent’s amount of orders ot is the difference between the optimum position qt∗ and the current position qt−1 . ot = qt∗ − qt−1

(14.3)

If the market price Pt is lower (higher) than his expected price (Pt−1 + yt ), each agent order to buy (sell). The amount of order is ot .  If ot > 0  If ot < 0

Buy ot (Pt ≤ Pt−1 + yt ) No action (Pt > Pt−1 + yt )

No action (Pt < Pt−1 + yt ) (Pt ≥ Pt−1 + yt ) Sell ot

14.3.3 Price Determination All the orders of 50 agents in the market are accumulated, and the market price of this term is determined as the value where the demand and supply are balanced. Dealings are transacted between the buyer who gave the price higher than a market price, and the seller of a lower price. 14.3.4 Learning Each agent updates the coefficients bt of the prediction formula (14.1) using the successive least-squares method with the information on the change 1 2

The initial value of the coefficients b0 is given with the uniform random numbers from -1 to 1. At the start, the initial values of price x0 are generated by the normal distribution whose average is 0 and standard deviation is 1.

114

K. Izumi

yt of the newly determined market price3 . The least-squares method is as follows[14.6]. −1

bt+1 = bt +

(Xt Xt )

xt (yt − xt bt ) , ft

(14.4)

where Xt is a learning matrix which starts by X0 = 100 × I (I is a unit matrix), and is updated by the following formula. −1  −1 (14.5) (Xt Xt ) = Xt−1 Xt−1   −1   −1 X Xt−1 xt xt Xt−1 Xt−1 − t−1 ft  −1 ft = 1 + xt Xt−1 Xt−1 xt

(14.6)

14.4 Simulation Result In the next section, we examine the complexity of the market and the prisoner’s dilemma-situation when the prediction dimension became large using the artificial market model. 14.4.1 Merit of Complicating a Prediction Formula We investigated the merit of complicating the prediction formula. The dimensions of 25 agents’ prediction formulas was set to n, and the dimension of the prediction formula of the other 25 agents was n + 1. Each simulation had 4000 terms which consisted of the four steps in section 14.3. The averages of forecast errors were calculated both about the agent group with n dimensions and about the group of n + 1 dimensions. The forecast errors were the difference between each agent’s prediction value and a market price. The initial value of random numbers was changed and 100 simulations was carried out4 . Figure 14.1 shows the difference between the forecasts errors of the group with n + 1 dimensions and those of the group of n dimensions. While the number of dimensions in the prediction formula is small, the merit of complicating prediction formulas is large. The agent who can predict correctly can increase his profit. Thus, when the number of dimensions is small, the conditions 1 of the prisoner’s dilemma situation in the section 14.2 are hold. However, when the number of dimensions becomes large, the merit of complicating prediction formulas disappears. 3 4

When n = 0, the prediction value is a random number and learning is not performed. Since the calculation of averages were impossible when the market price had diverged, we carried out simulations until we could get 100 simulations whose paths did not diverge.

D i ffer en ce of a ccu m u l a ted for eca s ts er r or s

14. Complexity of Agents and Complexity of Markets

115

% 70 60 50 4 0 3 0 20 10 0 -10 0 vs .1 1 vs . 2 2 vs . 3 3 vs . 4 4 vs . 5 5 vs . 6 6 vs . 7 7 vs . 8 8 vs . 9 9 vs . 10 D i m en s i on of a g en ts ’ for ea ca s t eq u a ti on .

Fig. 14.1. Comparison of forecast errors: Y-axis is a difference of forecast errors (forecast errors of the group of n dimensions are 100). Positive (negative) values mean that forecast errors of the group of n + 1 dimensions are small (large).

14.4.2 The Demerit in the Whole Market We examined whether the prediction of prices becomes harder in the whole market as increase of the dimension of prediction formulas. In this simulation, 50 prediction formulas of all agents were the same n dimension. We carried out the simulation with 4000 terms 100 times5 . After having accumulated the forecasts errors in 4000 terms and taking an average of 50 agents in 100 simulations. (Fig.14.2). As a result, when the number of dimensions in the prediction formula was small, the forecast error became large, as the number of dimensions increased. That is, the conditions 2 of the prisoner’s dilemma situation in the section 14.2 were hold. However, it has converged to the fixed value when the number of dimensions was lager than three. 14.4.3 Development of the Complexity of a Market In order to examine the independence of the complexity of the movement of a market price from the complexity of each market participant’s prediction formula, we carried out the correlation dimension analysis6 . All 50 agents have the prediction formulas of the same n dimension. We carried out the simulation with 4000 terms 100 times. Changed the embedding dimensions, 5 6

The path to diverge was not seen when all agents’ prediction formula was the same dimension. The procedure of the correlation dimension analysis was described in [14.7, 14.8].

116

K. Izumi Aver a g e of a ccu m u l a ted for eca s t er r or s

1200 1000 8 00 600 4 00 200 0

0

2 4 6 8 D i m en s i on of a g en ts ’ for ea ca s t eq u a ti on .

10

Fig. 14.2. Forecast errors

the correlation dimensions was calculated using the price data of 3885 terms at the second half while learning were stabilized to some extent (Fig.14.3). As a result, when a prediction dimension was 0, the correlation dimension curve was convex downward like the theoretical value of a random walk (fig. 14.3a). That is, there is no structure in the dynamics of the market price. However, when the prediction dimension increase a little, the correlation dimension curve was convex upward and saturated (fig. 14.3b). Thus, the structure that could be described by an equation of a finite dimension appeared in the dynamics of the market price. Furthermore, when the prediction dimension was raised, the correlation dimension curve became a straight line (fig. 14.3c). Thus, the correlation dimension curve was neither convex downward like a random walk nor saturated. That is, there was a structure in the dynamics of the market price, but it could not be described by an equation of any finite dimension. According to Nakajima [14.7, 14.8], as a result of analyzing Tokyo Stock Exchange Stock Price Index data, the logarithm of a correlation dimension went up linearly like this simulation result in fig. 14.3c. That is, when each agent’s prediction dimension increases, like the price data in the real-world, the dynamics of the price in the artificial market can be described roughly by an equation of some dimensions. And the more precise description is also attained by increasing the number of dimension. However, the movement of price data cannot be described completely by an equation of any finite dimensions. That is, the number of the variables related to the movement cannot be specified completely.

14. Complexity of Agents and Complexity of Markets

117

a) The agents’ prediction dimension is 0 4 5

cor r el a ti on d i m en s i on

4 0 3 5 3 0 25 20 15 10 5 0

10 em b ed d i n g d i m en s i on

b) The agents’ prediction dimension is 1 4 5

cor r el a ti on d i m en s i on

4 0 3 5 3 0 25 20 15 10 5 0

10 em b ed d i n g d i m en s i on

c) The agents’ prediction dimension is 10 4 5

cor r el a ti on d i m en s i on

4 0 3 5 3 0 25 20 15 10 5 0

10 em b ed d i n g d i m en s i on

Fig. 14.3. Correlation dimensions : X-axis is the logarithm of embedding dimensions. A solid line is an average of the correlation dimension of 100 paths. A dotted line is the theoretical value of a random walk.

118

K. Izumi

14.5 New Efficient Market Hypothesis The simulation results are summarized as follows. – When each market participant’s prediction dimension is 0, the movement of a market price resembles a random walk. If the prediction dimension increases, the structure that can be described by an equation of a finite dimension appears in the movement of price. – Therefore, if each agent increases his prediction dimension, since the prediction dimension approaches to the dimension of the price determination formula and his prediction becomes more accurate. Thus, the merit of complicating prediction formulas exists. However, if everybody increases his or her prediction dimension, prediction accuracy becomes smaller than before. That is, it will become the prisoner’s dilemma situation. – If everybody continues to increase the prediction dimension in the prisoner’s dilemma situation, the movement of a market price come to have the structure that can not be described completely by an equation of any finite dimensions. The structure of the movement of a market price changed as market participants changed their prediction formulas. That is, the complexity of market participants and the complexity of a market are not independent unlike the efficient market hypothesis. The simulation results also suggest that the structure of the dynamics of price data did not disappear when market participants continue to complicate their prediction formulas. In the final state, however each market participant increases his prediction dimension, he cannot predict the market price completely. In such the state where there is no “correct answer” of learning, it is thought that a procedure of learning by each market participant becomes the key factor to the movement of a market price in addition to a result of learning. As Kichiji[14.9] said, the efficiency of learning by a market participant, the difference in the cognitive framework, the interaction between market participants, and the method of informational choice, etc. become important. Another key point is the mechanism of market price determination. In this study we assumed that the market price were determined discretely as an equilibrium price. Alternatively we can assume that the market price is determined continuously as transaction prices of dealings. The mechanism of market price determination is the mechanism how to accumulate the individual complexity on the complexity of a market. Therefore, it has large influence on the relation between the complexity of market participants’ prediction formulas and the complexity of the movement of a market price. It is interesting to examine whether the same simulation can be acquired when the mechanism of market price determination changes.

14. Complexity of Agents and Complexity of Markets

119

14.6 Conclusion This study examined an efficient market hypothesis using artificial market approach. As a result, the following two points different from an efficient market hypothesis were found. – While the prediction dimension of agents is small, the structure which can be described to the movement of a market price exists, and the motivation of increasing the prediction dimension exists. – Even if the market participant increases the prediction dimension, the structure of the movement of a market price does not disappear. Finally, however each market participant increases his prediction dimension, he cannot predict the market price completely. As future works, we want to investigate the influence of (a) the procedure of learning by a market participant and (b) the mechanism of the price determination on the relation between between the complexity of market participants’ prediction formulas and the complexity of the movement of a market price. Acknowledgement. I want to be deeply thankful to Prof. Yoshihiro Nakajima who did offer useful comments in execution of this research.

References 14.1 Chen, S.H., Yeh, C.H. (1996): Genetic programming and the efficient market hypothesis. In Koza, J., Goldberg, D., Fogel, D., eds.: Genetic Programming: Proceedings of the 1st Annual Conference. the MIT Press, 45–53 14.2 Chen, S.H., Yeh, C.H., Liao, C.C. (1999): Testing the rational expectations hypothesis with the agent-based model of stock markets. In Proceedings of Internatinal Conference on Artificial Iintelligence 1999. Computer Science Research, Education, and Application Press, 381–387 14.3 Chen, S.H., Yeh, C.H., Liao, C.C. (2000): Testing the rational expectations hypothesis with the agent-based model of stock markets. In Papers of the Fourth Annual Conference of The Japan Association for Evolutionary Economics. The Japan Association for Evolutionary Economics, 142–145 14.4 de la Maza, M. (1999): Qualitative properties of an agent-based financial market simulation. In: Proceedings of ICAI99. CSREA, 367–373 14.5 Joshi, S., Parket, J., Bedau, M.A. (2000): Technical trading creates a prisoner’s dilemma: Results from an agent-based model. In Abu-Mostafa, Y.S., LeBaron, B., Lo, A.W., Weigend, A.S., eds.: Computational Finance 1999, MIT Press, 465–479 14.6 Harley, A.C. (1981): Time Series Models. Philip Allan Publishers 14.7 Nakajima, Y. (1999): An equivocal property of deterministic, and stochastic processes observed in the economic phenomena. Information Processing Society of Japan, Transaction on Mathematical Modeling and Its Applications 40, (in Japanese).

120

K. Izumi

14.8 Nakajima, Y. (2000): Keizai no yuragi to fractal. In Shiozawa, Y., ed.: Houhou to shiteno shinnka. Springer Verlag Tokyo, 207–235, (in Japanese). 14.9 Kichiji, N. (2000): Fukajitusei ka deno kitaikeisei to kasetu no shinnka. In Shiozawa, Y., ed.: Houhou to shiteno shinnka. Springer Verlag Tokyo, 173– 206, (in Japanese).

15. U-Mart Project: Learning Economic Principles from the Bottom by Both Human and Software Agents Hiroshi Sato1 , Hiroyuki Matsui2 , Isao Ono3 , Hajime Kita4 , Takao Terano5 , Hiroshi Deguchi2 , and Yoshinori Shiozawa6 1 2 3 4 5 6

Dept. of Computer Science, National Defence Academy, Hashirimizu 1– 10–20, Yokosuka, Kanagawa JAPAN, e-mail: [email protected] Graduate School of Economics, Kyoto University, JAPAN Faculty of Engineering, University of Tokushima, JAPAN National Institution for Academic Degrees, JAPAN Graduate School of Systems Management, University of Tsukuba, JAPAN Faculty of Economics, Osaka City University, JAPAN

U-Mart is an interdisciplinary research program of agent-based artificial market. U-Mart proposes an open-type test bed to study trading strategies of agents, behavior of the market and their relationship. An experiment open to public (Pre U-Mart 2000) using the proposed system is held in August 2000. More than 40 software agents (computer programs for trading) from 11 teams participated in this experiment. This paper reports the outline of the experiment, the trading strategies of the participated agents and the results of the experiment. While Pre U-Mart 2000 treated only software agents, the U-Mart system is designed considering participation of the human players as well as the software agents. A gaming simulation by human using the U-Mart system held in Kyoto University is also introduced briefly.

15.1 Introduction Complex behavior of market economy, typically observed in financial markets, is not fully explained by conventional economic theories. A new approach to this problem is an artificial market which enables computational experiments on virtual markets using agent simulation[15.1]. Studies on artificial markets have achieved a variety of interesting results. However, they also clarified the difficulties peculiar to this agent simulation approach, such as that: – researchers from different fields need to cooperate due to the interdisciplinary nature of this approach, – it is not easy to design a model which combines complexity (to imitate real markets) and simplicity (to enable computational experiments), and – researchers need to share common understanding on experimental configurations and results which are more complicated than theoretical models. T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 121− 13 1, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

122

H. Sato et al.

U-Mart1 [15.2, 15.3, 15.4] is a research program to address these problems of artificial market studies. We have developed an artificial market simulation system, called U-Mart system, to provide a test bed for researchers from economics and information science to carry out experiments with common understanding. We are promoting diversified researches on markets by opening this system to public. We have conducted an open experiment, Pre U-Mart 2000, on this system, inviting more than 40 software agents from public. This paper reports the result of the experiments, along with the strategies of the participated agents. The U-Mart system is designed to allow human players to participate in market experiments. This paper briefly introduces the human gaming simulation conducted at Kyoto University as well.

15.2 Outlines of U-Mart System In the U-Mart system, ‘futures’ of real stock index are traded in a virtual market. This allows the market simulation environment to reflects the complexity of real markets, and at the same time, enables independent price formation. The U-Mart system is implemented as a client-sever system, which exchanges information, such as buying and selling, via the Internet using a dedicated protocol implemented on TCP/IP. A sever, which imitates an ‘exchange’, accepts orders from clients, determines prices, matches buying and selling orders, and manages clients’ accounts. Each client obtains the information, such as market performance, from the sever and places order under its own decision. In the U-Mart system, human agents, as well as software agents, are allowed to participate in market experiments. Details of the U-Mart system are provided in [15.4]. U - M a r t C lie n ts

F in a l S e ttle m e n t w ith S p o t P r ic e

U -M a rt S e rv e r

P ro to c o l fo r U -M a rt H u m a n A g e n t

F u tu r e s P r ic e in U - M a r t O rd e r M a n a g e m e n t

P r ic e

In te rn e t

H u m a n A g e n t

P r ic in g V o lu m e A c c o u n tin g

S p o t P r ic e F in a l S e ttle m e n t D a y

S e llin g O r d e r C o n tra c te d V o lu m e

B u y in g O r d e r

T im e C o n tra c te d P r ic e

Fig. 15.1. U-Mart Artificial Market System

1

originally called V-Mart

P r ic e

S o ftw a re A g e n t S o ftw a re A g e n t

15. U-Mart Project

123

15.3 Outline of Open Experiment, Pre U-Mart 2000 15.3.1 Open Experiment and Its Objectives We conducted an open experiment, Pre U-Mart 2000, on August 19th. 2000 as a part of 6th. Emergent System Symposium of The Society of Instrument and Control Engineers in Japan. The objectives of this experiment are: to investigate variations of trading strategies and development methods for software agents, and to verify the actual behavior of market simulation among independently developed agents. Since it is the first open experiment for us, we limit the entry only to software agents. This is the reason that we name it “Pre U-Mart 2000”, which targets only a part of U-Mart conception. The participants have received an agent development package of U-Mart system in advance. This package contains templates of simple software agents and track record of J30 stock indices (used as spot data). 15.3.2 Experimental System At the occasion of the experiment, Pre U-Mart 2000 committee set up a server machine, and the participants run agent programs on their note PCs connected to the server via Ethernet. The participants and the audience can watch the progress of the experiment through a video projector. We tested the operation of the system on the first day of the symposium (August 18th.), and conducted the experiment in the afternoon of August 19th. 15.3.3 Configuration of Experiment The price determination and contract algorithms are described in [15.4]. Table 15.1 shows the parameters for the market. We use Dow Jones Industrial Average (scaled to J30 equivalent) to prevent participants from estimating the spot market data from distributed J30 data. The exchange (server) settles the accounts of agents at the end of one virtual day. When cash balance of an agent is less than zero after the settlement, the exchange automatically loan the agent up to its loan limit. The loan costs interest of 10% per annum and the exchange collect it at the settlement of the next virtual day. An agent goes into bankruptcy if the cash balance is still less than zero after obtaining the maximum loan, then the agent is not allowed to make any more deal.

15.4 Participated Agents and Their Strategies Eleven teams participated in the experiment, seven from engineering and four from economics. Each team was assigned a quota of five agents.

124

H. Sato et al.

Table 15.1. Parameters of Pre U-Mart 2000 Item

Setting

Underlying Indices Dow Jones Industrial Average, scaled to J30 equivalent Period 60 virtual days Order Methods market order/limit order Pricing Method ITAYOSE∗ Pricing Interval 15 seconds (real time) Number of Pricing 4 times/virtual day Trade Unit 1000-fold of contracted indices Bid and Asked indices in increments of one point Price Range no restriction Margin Money 300,000 YEN/Trade Unit Settlement System mark-to-market at closing price of the day Membership Fee none Cash on Hand 1 billion YEN/agent Loan Limit 30 million YEN * A pricing method that accumulates orders for a certain period, and decides a price so as to achieve the maximum contracted volume for the accumulated orders.

The basic strategies of participated agents are mainly based on time-series analysis (technical analysis) or the price difference between spot and futures markets2 . Some agents have been manually programmed and the others use learning/adaptation methods such as GAs and neural networks. There are other interesting agents such as: the one refers to buying and selling behaviors of other agents, the one implements explicit risk management, and the one learns in real time basis. The followings describe the strategies of each team. 1. University of Tokushima team (Engineering): #1 - #5 – Authors: Takao, I.Ono, N.Ono – Strategy: Some of their agents have learned neural networks (input: timeseries of price differences, output: buying/selling) using GA. The other agents implement technical analysis methods, such as moving average, oscillator[15.5], and psychological line. 2. Kyoto University team (Economics): #6 - #10 – Authors: Koyama, Zaima, Matsui, Deguchi – Strategy: Some of their agents place orders based on the deviation between short-term and very short-term moving averages. The other agents implement the improved version of psychological line. Contrivances have been made on number and amount of orders (for example, to make larger buying in the morning). 3. Tokyo Institute of Technology - Fukumoto team (Engineering): #11 - #15 2

Actual futures markets allows a strategy called “arbitrage”, which gains profit margin from the price difference by combining futures deals and spot deals. Since U-Mart only allows futures deals, the pure “arbitrage” strategy can not be implemented.

15. U-Mart Project

4.

5.

6.

7.

8.

9.

10.

11.

125

– Author: Fukumoto – Strategy: Their agents predict market trend with regression equation, and place orders based on deviation between current spot price and futures price. The parameters are learned with GA. They manage positions and implement bullish/bearish. Tokyo Institute of Technology - Yamamura Lab. team (Engineering): #16 #20 – Authors: Yamashige, Kira, Ishii – Strategy: Some of their agents have learned neural networks (input: deviation between gradient of moving average and closing price, and deviation between lowest and highest prices in the past, output: expected price) using a hybrid algorithm. The other agents are: the one sells/buys at crests and troughs of price movement, and the one places orders after comparing its position with price difference between spot and futures markets. Univ. of Tsukuba and Yamatake Industrial team (Engineering): #21 - #25 – Author: Murakami – Strategy: Their agents implement real-time learning of futures price prediction using classifier system, F-OCS. The agents have learned heavy rises and falls of markets and have incorporated the skills to cope with them. Osaka Pref. University team (Engineering): #26 - #30 – Author: Mori – Strategy: The parasitic agents which do not use price information. They depend only on ordering information of other agents and place the same orders with majority. Osaka Sangyo Univ. team (Economics): #31 - #35 – Authors: Taniguchi, Ozaki – Strategy: Some of their agents place orders according to the trend and against the trend. The other agents react to the gradient of price movement sensitively. National Defense Academy - Sato team (Engineering): #36 - #40 – Author: Sato – Strategy: Their agents implement basic day-trading. They place selling order with few percent higher and purchase orders with few percent lower than the closing price of previous virtual day and aim at the profit from the difference between them. Kyoto Sangyo Univ. team (Economics): #41 - #45 – Author: Nakashima – Strategy: Some of their agents place buying orders only or selling orders only base on dollar cost averaging method. The other agents place orders based on the ‘ren-gyo-soku’ method, a method of technical analysis. National Defense Academy - Ishinishi team (Engineering): #46 - #50 – Author: Ishinishi – Strategy: Their agents place buying order when spot price is higher than futures price, and place selling order when spot price is lower than futures price. Osaka City University team (Economics): #51 - #55 – Author: Shiozawa – Strategy: Basic technical analysis.

126

H. Sato et al.

15.5 Experimental Result We have conducted the experiments twice with different spot price series. The numbers of attended agents are 47 for the first round and 43 for the second round. Not every team uses its full quota of five agents. 2 7 0 0

2 5 0 0

3 5 0 0

2 6 5 0

2 5 0 0

3 4 0 0

F u tu r e s P r ic e 2 6 0 0

2 0 0 0

2 0 0 0 3 3 0 0

2 4 0 0

1 0 0 0

3 2 0 0

1 5 0 0

T r a d e d V o lu m e

1 5 0 0

2 4 5 0

S p o t P r ic e

P r ic e

P r ic e

2 5 0 0

T r a d e d V o lu m e

S p o t P r ic e

2 5 5 0

3 1 0 0 1 0 0 0

3 0 0 0

2 3 5 0 2 9 0 0 2 3 0 0

F u tu r e s P r ic e

5 0 0

5 0 0

2 8 0 0

2 2 5 0 2 2 0 0

0 0

1 0

2 0

3 0

4 0

5 0

6 0

V ir tu a l D a y

2 7 0 0

0 0

1 0

2 0

3 0

4 0

5 0

6 0

V ir tu a l D a y

Fig. 15.2. Prices and Traded Volumes for 1st. Round (left) and 2nd. Round (right)

Table 15.2. Top 10 Performance of Agents for 1st. and 2nd. Round Agent

1st. Round Profit∗1 Team∗2

Agent

2nd. Round Profit∗1 Team∗2

#41 3,960,884 9 #12 3,005,755 3 #27 582,474 6 #13 1,792,902 3 #26 380,437 6 #18 1,686,144 4 #7 317,955 2 #19 820,168 4 #5 310,538 1 #43 710,379 9 #33 307,773 7 #44 388,575 9 #21 266,145 5 #27 285,245 6 #28 258,410 6 #7 254,108 2 #16 225,309 4 #16 206,260 4 #30 204,743 6 #9 197,120 2 *1: price unit: 1,000 YEN, *2: team is represented by their entry number

15.5.1 First Round The spot price series for the first round repeats up and down several times and ends at the beginning price. Figure 15.2 (left) shows the transitions of price and trade volume. Table 15.2 and 15.3 show the performance of each agent and each team at the end of the game. The heavy rises and falls are repeated at the beginning because of excessive limit order and market order combinations. Five agents go into bankruptcy during 11th. and 14th. virtual days. No agent goes into bankruptcy

15. U-Mart Project

127

Table 15.3. Performance of Teams at Pre U-Mart 2000 Team

1st. Round

2nd. Round

Kyoto Sangyo Univ. 2,717,039 –1,059,526 Osaka Pref. Univ. 1,512,561 –4,309,662 Univ. of Tokushima 661,096 –1,393,736 Kyoto Univ. 635,519 –1,175,857 Sato (NDA) 622,257 111,153 Osaka Sangyo Univ. 501,101 –1,504,747 Yamamura Lab. (TIT) 358,853 2,751,064 Univ. of Tsukuba and Yamatake 332,358 192,780 Osaka City Univ. 156,941 –53,780 Fukumoto (TIT) –232,420 4,079,164 Ishinishi (NDA) –4,711,406 –99,237 descending order of 1st. round profit (unit: 1,000 YEN)

before 11th. because rises and falls do not occur at the closing price, which directly affect to the end of the day settlement (c.f. 2nd. round). After the five agents go into bankruptcy, the market calms down and the deals are made around the spot price. The trade volume increases at the rapid price movements because of the huge volume of market orders. 15.5.2 Second Round The spot price series for the second round shows long-term downtrend. Figure 15.2 (right) shows the transitions of price and trade volume. Table 15.2 and 15.3 show the performance of each agent and each team at the end of the game. The second round shows only a few times of rapid price movements. This is because the price movement at the first day is too big (the futures price is 19,332 YEN, while the spot price is 3,178 YEN) and the market closes at this price. Three agents go into bankruptcy and the other agents are damaged seriously as well. Consequently, the trade volume decreases after the second day. Two more agents go into bankruptcy on 12th. day because of the huge price movement at the closing. Total of five agents go into bankruptcy on second round. The trade volume increases at the rapid price movements because of the huge volume of market orders. 15.5.3 Variety of Agents Eleven teams participated in these experiments and the variety of the agents exceeded our expectations. When agents show similar behavior, deals tend to fail because their decisions are similar. In such a case, to achieve deals, agents which place random

128

H. Sato et al.

orders need to be introduced on the market. In our experiments, the prices have been formed between the varied agents without random agents. Although several teams use the same analysis methods (moving average and psychological line), the final asset of these teams differs remarkably. This means that these teams interpreted the indices differently in implementation of the methods as software agents. Technical analysis indicates “the time to buy (or sell)”, but it does not recommend “the amount to buy (or sell) in which price”. We expect that this point is clarified with larger number of experiments. It is interesting that the agents #41-45 (selling only/buying only) and #26-30 (do not use price data) have made good results especially on the first round. It does not mean that these strategies are always effective. However, they are obviously against the common practice that winners need to predict the future based on price data and to manage their position appropriately. Their successful performance contribute to the variety of agents. In the future, more agents will implement the position management (implemented only on #11-#15) or the online learning for real-time modification of strategy (implemented only on #21-#25). 15.5.4 Reason of Heavy Rises and Falls

V o lu m e

V o lu m e

The heavy rises and falls occur at the beginning of both rounds. At these experiments, we have not restricted the price range and the agents are allowed to place orders at unrealistic price. Although these unrealistic orders normally do not affect price determination, they may be contracted when huge volume of market orders are placed. In the price determination algorithm of U-Mart system, selling market orders are considered as “limit orders lower than the lowest limit order” and buying market orders are considered as “limit orders higher than the highest limit order”. This makes the price formation vulnerable to huge volume of market orders (See Figure 15.3).

D e m a n d

S u p p ly

D e m a n d

S u p p ly

M a r k e t S e llin g O r d e r

M a r k e t S e llin g O r d e r L o w

P r ic e M a r k e t B u y in g O r d e r

H ig h

L o w

P r ic e

H ig h

M a r k e t B u y in g O r d e r

Fig. 15.3. Price Determination by ITAYOSE: When limit orders are dominant (left); When market orders are dominant(right).

There are two types of agents which place excessive orders. One type gives “very low buying limit and very high selling limit” (i.e. #38) and another type gives “very low selling limit and very high buying limit” (i.e. #35). We had assumed that they do not affect the market because the former

15. U-Mart Project

129

type has difficulty in making deal and the latter type goes into bankruptcy immediately. However, they have hazardous nature to rattle the market in relation with market orders. We may need to restrict the price range or to reconsider the price determination method.

15.6 Experiments with Human Agents Heavy rises and falls have resulted at the beginning of the experiments with software agents. What happens if more sophisticated human agents deal in this virtual market? The U-Mart system can answer this question since it is designed to allow human agents to participate in market experiments. As an example of the behavior of virtual markets constructed by human agents, this section introduces the experiments conducted at Kyoto University as a part of a lecture on gaming simulation3 . The experiments with human agents have been conducted three times under the similar conditions as Pre U-Mart 2000, using different spot data for each time. In these experiments, small number of software agents are introduced on the market. They place limit orders at the prices determined by random numbers which comply with normal distribution around the spot price. Initially, the students made deals without strategy. It was natural because they were not familiar with the client software and they did not know much about futures markets or futures trade mechanisms. However, they started to understand these mechanisms by accumulating experience and became more strategic. The result of third experiment (conducted on November 16th.) is shown in Figure 15.4. It shows the transition of the spot price, the virtual market price (U-Mart Price), and the asset position of each agent. In this experiment, a software agent has made the best profit among one software agent and seven human agent (including one faculty), and three students go into bankruptcy. According to the students’ reports after the experiments, the bankrupt students predict down-trend of spot price in long-term. They focuses on buying initially and continues selling after that, then go into bankruptcy along with the up-trend of spot price. On the other hand, the profited students respond to short-term price movements. They make small profits with a general strategy, that is to sell when price increases and to buy when price decreases. They maintain the stable position. The experimental results show remarkable differences on behavior of human agents and the present software agents. Human agents not only make 3

“Economics System Gaming” (Dr. Deguchi) given at School of Economics, Kyoto University. This is a two class period on end (180 min.) biweekly lecture geared to undergraduate and graduate students.

130

H. Sato et al.

§ÉƽÀË

§ÉÀº¼ ‹ƒ‡‡‡ мÅ

‰ƒŒ‡‡Ïˆ‡µ мÅ

ŠƒŒ‡‡ мŠ‰ƒ‡‡‡Ïˆ‡µ мŠŠƒ‡‡‡ мŠ‰ƒŒ‡‡ мÅ

ˆƒŒ‡‡Ïˆ‡µ мÅ

‰ƒ‡‡‡ мÅ

ļĹ¼Éˆ ļĹ¼É‰ ļĹ¼ÉŠ ļĹ¼É‹ ļĹ¼ÉŒ ļĹ¼É ļĹ¼É ĸº¿Àż ¬„¤¸ÉË §ÉÀº¼ ª§¦« §ÉÀº¼

ˆƒ‡‡‡Ïˆ‡µ мÅ

Œ‡‡Ïˆ‡µ мÅ

ψ‡µ мÅ

ˆ Š Œ Ž  ˆˆ ˆŠ ˆŒ ˆŽ ˆ ‰ˆ ‰Š ‰Œ ‰Ž ‰ Šˆ ŠŠ ŠŒ ŠŽ Š ‹ˆ ‹Š ‹Œ ‹Ž ‹ Œˆ ŒŠ ŒŒ ŒŽ Œ ˆ ›¸Ð

ˆƒŒ‡‡ мŠˆƒ‡‡‡ мŠŒ‡‡ мŠ‡ мÅ

Fig. 15.4. Experimental Result with Human Agents

technical analysis of short-term price movement, but they predict long-term market trend and conceive a strategy based on impression. Although the software agent has made the best profit in this experiment, it highly depends on contingency in connection with the used spot data and the strategies of human agents. From now on, more experimental cases need to be accumulated to analyze U-Mart as a market and to examine differences between human and software agents. We will also look into the availability of this system as an educational tool.

15.7 Conclusion and Acknowledgements In this paper, we have reported on the experiments of open-type artificial market, U-Mart, conducted with software agents and/or human agents. The results of experiments have shown the possibility to construct a variety of software agents and clarified the strategic differences between human and software agents. We will carry this study program forward by integrating the knowledge obtained from both type of agent simulations. It is also interesting that the results indicated the usefulness of the U-Mart system as an educational tool for both economics and information science. At the last, we are grateful to the participants of Pre U-Mart 2000 and everyone concerned with 6th. Emergent System Symposium. Also, we would like to thank Dr. Deguchi, Graduate School of Economics, Kyoto University, who provides the opportunity of human agents simulation using U-Mart system, and the students participated in the experiments.

References 15.1 (2000): Special Issue ‘Artificial Markets’, J. of Japanese Society for Artificial Intelligence, Vol. 15, No. 6 (in Japanese)

15. U-Mart Project

131

15.2 Koichi Kurumatani, Yuhsuke Koyama, Takao Terano, Hajime Kita, Akira Namatame, Hiroshi Deguchi, Yoshinori Shiozawa and Hitoshi Matsubara (2000): U-Mart: A Virtual Stock Market as a Forum for Market Structure Analysis and Engineering, in Proc. 5th Joint Conference on Information Science (JCIS’00), 1st Int’l Workshop on Computational Intelligence in Economics and Finance, Vol. 2, pp. 957-960 15.3 http://www.u-mart.econ.kyoto-u.ac.jp/ 15.4 H. Satoh, M. Kubo, R. Fukumoto, Y. Hirooka and A. Namatame (2000): System Structure of an Artificial Market, J. of Japanese Society for Artificial Intelligence, Vol. 15, No. 6, pp. 974-981 (in Japanese) 15.5 John J. Murphy (1999): Technical Analysis of the Financial Markets: A Comprehensive Guide to Trading Methods and Applications, Prentice Hall Press

16. A Multi-objective Genetic Algorithm Approach to Construction of Trading Agents for Artificial Market Study Rikiya Fukumoto1 and Hajime Kita2 1

2

Interdisciplinary Graduate School of Science and Engineering, Tokyo Institute of Technology, Yokohama 226-8502, JAPAN National Institution for Academic Degrees, 3-29-1 Otsuka, Bunkyo, Tokyo 112-0012, JAPAN [email protected]

To construct agents that have trading strategies with adequate rationality and variety is an intrinsic requirement for artificial market study. Difference of preference to return and risk among agents will be one candidate reason of variety of the trading strategies. It can be treated as a multi-objective optimization problem taking both criteria as objective functions. This paper proposes a multi-objective genetic algorithm(MOGA) approach to construction of trading agents for an artificial market. The U-Mart system, an artificial market simulator, is used for a test bed. Agents are evaluated in the U-Mart with other agents having simple strategies, and evolved with the MOGA. Computer simulation shows that various agents having non-dominated trading strategies can be obtained with this approach.

16.1 Introduction Recognizing complex behaviors of the prices in the real markets and limitation of conventional theories in economics, analysis of markets using agent based simulation, called artificial markets, attracts attention[16.1, 16.2, 16.4]. In some of simulation models for artificial market, rather simple agents are employed so as to establish clear relationship between microscopic behavior of the agents and macroscopic behavior of the market. On the other hand, some of the models use more complex agents to study adaptation, learning, and evolution of the agents in the market. For the artificial market study, it is required that the agents should have trading strategies with adequate rationality as a model of microscopic economic behavior on the one hand, and on the other hand, their strategies should have variety to form price in the market. To construct agent meets such requirements is, therefore, a one of key issues in artificial market study. In this paper, considering difference of preference to ‘return’ and ‘risk’ among agents as one of the important reasons of variety of the trading strategies, problem of designing agents is studied as a multi-objective optimization T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 13 2− 14 1, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

16. A Multi-objective Genetic Algorithm Approach

133

problem taking the both criteria as objective functions. A multi-objective genetic algorithm(MOGA) is taken as an approach to construction of trading agents. For this study, the U-Mart system, an artificial market simulator, developed as a common test bed in this field is used. This paper is organized as follows: succeeding to introduction in this section, the U-Mart System and the Multi-Objective Genetic Algorithms are briefly explained in Sections 2 and 3, respectively. In Section 4, the structures of the agents and implementation of the MOGA for this study is described. Section 5 shows the results of numerical experiments. Section 6 concludes this study.

16.2 The U-Mart System For study of variety of trading strategies, learning and evolution of them, emerging behaviors of the market with them, and indirect control of the market through institutional design, artificial market systems with adequate complexity are required. Inspired by the RoboCup[16.6], the ‘U-Mart’ research program have organized and the U-Mart system has been developed[16.4]. The U-Mart system has following characteristics: – In the U-Mart, futures of an existing stock index is traded. Thus, complexity of the real world is introduced keeping ability of autonomous price forming in the artificial market. – The U-Mart system can be used for experiments with program agents, human traders, and their mixture. Thus it makes various research plans both in the communities of economics and computer science possible. – The U-Mart takes server(futures market)-client(trading agent) structure over TCP/IP. Communication between the server and client is regulated by a readable text-base protocol called the Simple Virtual Market Protocol(SVMP). It makes development of servers and trading agents in parallel on various platforms, and experiments over the Internet possible. – The U-Mart server is implemented in Java considering experiments on various platforms. In August 2000, the first open trading contest limited to program agents (Pre U-Mart 2000) was held. More than 10 teams both from economics and computer science fields participated. This experiment shows that feasibility of the research program. Further, the U-Mart system has also been used for education both in the computer science and economics[16.5].

16.3 Multi-objective Genetic Algorithms (MOGA) Multi-objective Optimization Problem (MOP) is a problem of optimizing multiple objectives simultaneously. In general, there exists trade-off among

134

R. Fukumoto and H. Kita

objectives, and therefore usually no single solution can be the optimum. As rational solutions of the MOP, non-dominated solution (or the Pareto optimal solution) is considered. The non-dominated solution is a solution that has at least one objective function whose objective value is better than that of any other solutions. Hence, the goal of the solver for the MOP is to obtain the set of the non-dominated solutions called ‘the Pareto optimal set’. Genetic Algorithms (GA) are optimization technique inspired by the natural selection theory of evolution[16.7, 16.8]. In the GAs, population of candidate solutions are evolved by repetitive application of genetic operators such as selection/generation alternation, crossover/recombination, and mutation. Multi-objective Genetic Algorithms(MOGA) are GAs that try to obtain various Pareto optimal solutions of a MOP simultaneously making use of the population-based search of the GAs[16.9]. MOGA is constructed by extending a single objective GA by introducing – Mechanisms of selecting non-dominated solution among population as survivors to make population evolve closer to the Pareto optimal set. – Mechanisms of maintaining diversity of the population to make the population cover the whole Pareto optimal set well. Details of implementation of the MOGA in this study is discussed in the next section.

16.4 Construction of Trading Agents with a MOGA 16.4.1 Structure of Trading Agents The U-Mart system carries out simulation with discrete time steps t = 1, · · · , tend . In period t, each trading agent can observe – S(t) = {s(1), · · · , s(t−1)}: spot prices of the stock index up to the previous period t − 1, – F (t) = {f (1), · · · , f (t−1)}: futures prices in the U-Mart up to the previous period t − 1, – position(t − 1): the position of the agent, – cash(t − 1): amount of cash possessed by itself, and – rest(t): remaining time up to the final period. Observing these variables, each agent must decide his action consisting of – p(t): limit price of the order, – sb(t): type of order, i.e., sell or buy, – q(t): quantity of the order, for each period t. Hence, the strategy of the agent can be formalized as the following function F :

16. A Multi-objective Genetic Algorithm Approach

(p(t), sb(t), q(t)) = Strategy(S(t), F (t), position(t − 1), cash(t − 1), rest(t − 1))

135

(16.1)

We have constructed agents having the following two structures. Model 1. This model is a sort of agent having strategies based on technical analysis, i.e., time series prediction of the prices. The agent consists of the following three parts: Risk Management Part : In the U-Mart, to maintain adequate position to avoid bankruptcy is a basic requirement for program agents. In this model, the agent memorizes the maximum price change max d for the past n periods, max d =

max

τ =t−n,··· ,t−1

|f (τ ) − f (τ − 1)|

(16.2)

and obtain a pessimistic estimate of its asset based on the history of past n periods when it keeps the current position as follows cash − (margin + max d × unit) × position

(16.3)

where margin is the margin for the contracted orders deposited in the market unit is trading unit. If it gets negative, it means the bankruptcy. Hence, the maximal possible position, say position ∗ , can be estimated as a solution of cash − (margin + max d × unit) × position ∗ = 0.

(16.4)

Trend Prediction Part : With linear regression analysis, both the spot prices s(t) and the futures prices f (t) for the past n periods are fitted as linear functions of period t: s˜(t) = as t + bs ,

f˜(t) = af t + bf

Further, we assume that the futures prices f (t) can be explained by linear combination of them fˆ(t) = y(t)s(t) + (1 − y(t))f (t) = at + b

(16.5)

where y(t) ∈ [0, 1] is the combination weight function depending on period. In the beginning, the futures price f will be explained better by f˜(t) than s˜(t), hence small y(t) will be preferred. Closing to the end, to use s(t) will be better, and hence large y(t) will be preferred. In this model, y(t) is represented by a piece-wise linear function of period as shown in Fig. 16.1, and 9 control points are taken as parameters to be decided. Order Making Part : Order is made based on two plans. The first plan is based on the trend of the futures price:

136

R. Fukumoto and H. Kita

Fig. 16.1. Piece-wise linear representation of y(t).

– If a > 0, take sell position. – If a < 0, take buy position. – Otherwise, take no position. However, its trend |a| is too strong, it may indicate some instability of the market. Then, to have large position will be dangerous. Considering this risk, the position (normalized by position  ) g(a, v) to be made is represented as a non-monotonic function of a as follows ⎧ 0 (a/v ≤ −1) ⎪ ⎪ ⎪ ⎪ 2a/v + 2 (−1 < a/v ≤ −0.5) ⎪ ⎪ ⎨ −2a/v + 2 (0.5 < a/v ≤ 0) g(a, v) = (16.6) 2a/v (0 < a/v ≤ 0.5) ⎪ ⎪ ⎪ ⎪ −2a/v + 2 (0.5 < a/v ≤ 1) ⎪ ⎪ ⎩ 0 (1 < a/v) where v is a parameter. Let position obtained from Eq. (16.6) be po 1 = position ∗ × g(a, v).

(16.7)

The other plan uses difference between estimated price fˆ(t − 1) and the actual futures price f (t − 1). Position to be taken po 2 is given by po 2 = position ∗ × d ×

fˆ(t − 1) − f (t − 1) maxτ =1,···t−1 |fˆ(τ ) − f (τ )|

(16.8)

where d is a parameter. These two positions are combined through a weight parameter w1 . Multiplying a parameter w2 representing ‘aggressiveness’ of the agent to it, we obtain the position po  to be taken as follows: po  = w2 (w1 po 1 + (1 − w1 )po 2 ).

(16.9)

The difference between po  and the current position q = po  − position

(16.10)

is taken as the amount of order to achieve po  . The limit price p is decided by extrapolating the estimated price fˆ(t) to the n step future.

16. A Multi-objective Genetic Algorithm Approach

137

Genetic Representation : the above strategy is represented by a chromosome consisting of the following 14 parameters: – Number of steps n ∈ [2, 60] used for linear regression analysis. – Nine parameters to decide function y(t). – Parameter v ∈ [0, 100] used in the function g. – Parameter d ∈ [−2, 2] used for deciding po 2 . – Weight parameters w1 and w2 ∈ [0, 1]. Model 2. This model takes a strategy based on arbitrage. Risk Management Part : As same as the Model 1, maximum possible position, position ∗ is calculated. Evaluation of Arbitrage Opportunity : This model decides position to be taken based on the difference between the prices of the spot and the futures: po = position ∗ × y(t) ×

s(t − 1) − f (t − 1) maxτ =t−m,··· ,t−1 |s(τ ) − f (τ )|

where 0 < y(t) < 1 is a weight function, and m is the size of the window to evaluate the arbitrage opportunity. As same as Model 1, y(t) is represented by a piece-wise linear function consisting of 8 segments. Order Making Part : Amount of the order is decided as follows: q = po  − position The limit price is taken as same as the latest spot price. Genetic Representation : the above strategy is represented by a chromosome consisting of the following 11 parameters: – Number of steps n ∈ [2, 60] used for risk management. – Parameter m used for assessment of arbitrage opportunity. – Nine parameters to decide function y(t). 16.4.2 Implementation of MOGA Objective Functions. Performance of a strategy taken by an agent is measured by ProfitRatio ≡

FinalPropety − InitialProperty InitialProperty

(16.11)

As for the objective functions representing return and risk, the mean and the variance of the ProfitRatio in 30 simulation runs with different spot price series are used. The number of the simulation runs for evaluation of an individual is decided considering the trade-off between stability of the results and computation time through preliminary experiments.

138

R. Fukumoto and H. Kita

Market Configurations. Each individual in the population is evaluated independently. That is, each individual is put into a separate market with prescribed agents, and its performance in the market is evaluated. Concerning composition of the market, we used two configurations: Configuration 1 : Market consists of the agent to be evaluated and 20 other agents having rather simple strategies as follows: Type r :5 agents that generate orders with random prices around the previous futures price. Type s : 5 agents that generate orders with random prices around the previous spot price. Type t :5 agent that buy futures if the previous price is higher than before, and sell otherwise following the trend of the market. Type a : 5 agent that buy futures if the previous price is lower than before, and sell otherwise. That is, they are anti-trend traders. Configuration 2 : The following 9 agents developed in the educational program held in Tokyo Institute of Technology using the U-Mart system are added to Configuration 1: Agent 1 : An agent that utilizes moving averages of the spot prices with large and small windows. Agent 2 : An agent that utilizes large and medium window moving averages of both the spot prices and the futures prices. Agent 3 : An agent that utilizes moving average of the futures prices. Agent 4 : An agent that utilizes current futures price, the moving averages and their variances. Agent 5 : An agent that utilizes the differences of the spot price and futures price, and variation of the futures prices. Agent 6 : An arbitrager that decides position based on the difference between the prices of spot and futures. Agent 7 : An agent that decides order based on the difference of the futures price and average price of its contracted orders. Agent 8 : An agent that utilizes quadratic approximation of the price curve and tries to capture the peak and bottom of the prices so as to decide its order. Agent 9 : An agent that makes orders using the strategy of ‘Type t’ if the property is larger than the initial value, and ‘Type a’ otherwise. Algorithm An algorithm of the MOGA based on the PESA[16.10] is used. Outline of the algorithm is as follows: 1. Generate initial N individuals randomly and evaluate them. Let generation counter g = 0. 2. Increment g. If g = G, terminate the algorithm. Otherwise choose two parents randomly from the population.

16. A Multi-objective Genetic Algorithm Approach

139

3. 4. 5. 6.

Let the counter of generated children m = 0. Increment m. If m = M go to Step 2. Generate a child with the UNDX[16.11], and evaluate its objective values. If the child is dominated by one of the individual in the current population, go to Step 4. 7. If all the individuals in the current population is non-dominated, go to Step. 9. 8. Replace a dominated individual with the child, and go to Step. 4. 9. Replace one of the two nearest individuals to the children in Euclidean distance with the children, and go to Step. 4. Considering available computation time and reliability of the solution, we set N = 30, M = 1, and the maximum generation G = 10000. Suppression of Non-active Agents. Preliminary experiment with a single objective GA that considers only ‘return’ shows that 1. Initial individual generated randomly usually yields negative returns. 2. Evolution path of strategies shows that return of the agent is gradually improved keeping rather larger risk, and finally positive return is achieved. In the multi-objective GA, strategy of ‘do nothing’, which yields no return with no risk dominates most of the initial population. Hence, in runs of the naively implemented MOGA, we observed a tendency that population converges to such useless strategies. To avoid this phenomenon, we evaluate each strategies giving a certain initial position. That is, in 10 runs among 30, the agent starts trade with initial position of 300 unit sell, in 10 runs with 300 unit buy, and in the remaining with no position. The amount of initial position is decided by trialand-error in the preliminary experiments. Thus, even non-active agents face risk due to the initial position, and therefore it has more selection pressure than in naive implementation.

16.5 Results of Experiments Results of experiments are shown in Figs. 16.2 (a) ∼ (d). These figures show distribution of the objective values of the agents in the market. Good solutions have large values in the return, and small values in the risk, and therefore located in the right lower area of the figures. In these figures, a curve of y = x2 is also plotted. If the profit ratio follows a normal distribution, strategies under this curve yield positive return more than 84% in probability. As for the Fig. 16.2 (d), a curve of y = x2 /4, that corresponding 98% positive return is also plotted. As for Model 1, the MOGA finds good solutions that dominate other simple strategies in Configuration 1. However, in Configuration 2, i.e., in

140

R. Fukumoto and H. Kita

the market having more sophisticated agents, solutions found by the MOGA located relatively small risk area, and dominated by some of them such as Type s, and Agent 6, which are a sort of arbitrager. It is interesting that the performances on the simple agents of Type r, s, t and a change largely in Configuration 1 and 2. As for Model 2, the solution by the MOGA achieves better results. Even in Configuration 2, the obtained strategies dominates most of the other agents, and performances under the 98% curve are achieved. It shows advantage of the arbitrage-based strategies in the futures market. It should be noted that agents in the population evolve based on evaluation in the separate markets. Evolution of agents in the same market, i.e., co-evolution of strategies is a subject of the future study. 0 .5

T y p e T y p e T y p e T y p e M o d e l

0 .4

0 .5 r

T y p e T y p e T y p e T y p e M o d e l

s t a

0 .4 1

t 2

R is k

0 .3

R is k

0 .3

r s a

0 .2

0 .2

0 .1

0 .1

0

0 -0 .4

-0 .3

-0 .2

-0 .1

0

0 .1

0 .2

0 .3

0 .4

-0 .4

-0 .3

-0 .2

-0 .1

R e tu rn

0

0 .1

0 .2

0 .3

(a) Model 1 in Configuration 1

(c) Model 2 in Configuration 1

1

1

A A A A

0 .6 A

R is k

A A A

0 .4 A M

r s t a

0 .8 1

A

2

A 3

A 4

A

0 .6 5 6

A A

R is k

0 .8

T y p e T y p e T y p e T y p e g e n t g e n t g e n t g e n t g e n t g e n t g e n t g e n t g e n t o d e l

0 .4

R e tu rn

7 8

A 0 .4

A

9 1

M

A

0 .2

T y p e T y p e T y p e T y p e g e n t g e n t g e n t g e n t g e n t g e n t g e n t g e n t g e n t o d e l

r s t a 1 2 3 4 5 6 7 8 9 2

0 .2

0

0 -1

-0 .5

0

0 .5

R e tu rn

(b) Model 1 in Configuration 2

1

-1

-0 .5

0

0 .5

1

R e tu rn

(d) Model 2 in Configuration 2

Fig. 16.2. Results of Experiments

16.6 Conclusion This paper proposes a multi-objective genetic algorithm(MOGA) approach to construction of various trading agents for an artificial market. That is, return and risk are treated objective functions for designing trading agents

16. A Multi-objective Genetic Algorithm Approach

141

using the U-Mart system, an artificial market simulator, as a test bed. Several techniques are also developed to achieve efficient evolution of the agent. Computer simulation shows that various agents having non-dominated trading strategies can be obtained with this approach. Acknowledgement. The authors express their sincere gratitude to the UMart program and its members for valuable suggestions, and Prof. Shigenobu Kobayashi of Tokyo Institute of Technology for valuable comments. and distribution. They also thank to graduate students participated in the educational program using the U-Mart system in the Department of Computational Intelligence and Systems Science of Tokyo Institute of Technology, for let them utilize the trading agents developed in this program.

References 16.1 W. B. Arthur et al. (1996): Asset Pricing under Endogenous Expectation in an Artificial Stock Market, SFI Working Paper 16.2 K. Izumi and K.Ueda (1998): Emergent Phenomena in a Foreign Exchange Market: Analysis Based on an Artificial Market Approach, Artificial Life, VI, 398–402 16.3 A. Sato and H. Takayasu (1998): Dynamic numerical models of stock market price: from microscopic determinism to macroscopic randomness, Physica A, 250, 231–252 16.4 K. Kurumatani et al. (2000): U-Mart: A Virtual Stock Market as a Forum for Market Structure Analysis and Engineering, Proc. 5th Joint Conference on Information Science (JCIS’00), 1st Int’l Workshop on Computational Intelligence in Economics and Finance, 2, 957–960 16.5 H. Sato et al.(2001): U-Mart Project: Learning Economic Principles from the Bottom by both Human and Software Agents, JSAI2001 Int’l Workshop on Agent-based Approaches in Economics and Social Complex Systems, Matsue, Japan 16.6 http://www.robocup.org 16.7 J.H. Holland (1975): Adaptation in Natural and Artificial Systems, Univ. of Michigan Press 16.8 D. Goldberg (1989): Genetic Algorithms in Search, Optimization, and Machine Learning, Addison-Wesley 16.9 H. Tamaki, H. Kita and S. Kobayashi (1996): Multi-Objective Optimization by Genetic Algorithms: A Review, Proc. of the 3rd IEEE Conference on Evolutionary Computation (ICEC’96), 517-522 16.10 D.W. Corne et al. (2000): The Pareto Envelop-Based Selection Algorithm for Multiobjective Optimization, Proc. PPSN VI, 839–848 16.11 I. Ono and S. Kobayashi (1997): A Real-coded Genetic Algorithm for Function Optimization Using Unimodal Normal Distribution Crossover, Proc. 7th ICGA, 246–253

17. Agent-Based Simulation for Economic and Environmental Studies Hideyuki Mizuta1 and Yoshiki Yamagata2 1

2

IBM Tokyo Research Laboratory Shimotsuruma 1623-14, Yamato, Kanagawa 242-8502, Japan [email protected] Center for Global Environmental Research National Institute for Environmental Studies Onogawa 16-2, Tsukuba, Ibaraki 305-0053, Japan [email protected]

The need for new theoretical and experimental approaches to understand dynamic and heterogeneous behavior in complex economic and social systems is increasing. Computational simulation with dynamically interacting heterogeneous agents is expected to be able to reproduce complex phenomena in economics, and helps us to experiment with various controlling methods, to evaluate systematic designs, and to extract the fundamental elements which produce the interesting phenomena in depth analysis. To implement various applications of the agent-based simulation effectively, we have developed a simple framework. We also consider a new application of agent-based simulation for an environmental study and implement a preliminary simulation model of the international greenhouse gas (GHG) emissions trading.

17.1 Introduction In real economic situations, the dynamic behavior and interactions between people are very complicated and may often seem irrational. Further complicating the situation, the recent progress and popularity of network communication technologies greatly widens the diversity of participants and affects the market mechanism itself, and increases the dynamic fluctuations of economic systems. In the past, traditional economic theories have only considered idealized representative participants in equilibrium states. It is very difficult to analyze dynamically changing situations involving heterogeneous subjects using such static and homogeneous methods. In the last decade, many researchers, including physicists and computer scientists, are starting to apply new approaches to investigate such complex dynamics in their studies of economics. One of these approaches is the agent-based simulation approach. The term “agent” is often used with different meanings by different researchers (see Fig. 17.1). For example, the word agent may refer to an autonomous graphical user interface with animation, a robot who gathers information from a network, an artificial lifeform, or a distributed application which collaborates with other components over the network. In economics, an agent usually T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 14 2− 152, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

17. Agent-Based Simulation for Economic and Environmental Studies

143

means an independent economic entity like a household or a firm. However, traditional economic theories usually consider only representative agents in equilibrium states. By using simulation technology, we can endow such economic agents with heterogeneous and dynamic properties. Thus, when we refer to an agent-based simulation, we assume a simulation study of an economic system composed of heterogeneous and dynamic economic entities.

User Interface Distributed Intelligence

Artificial Life

Agents Network Infrastructure

Economic Subject

Robot Fig. 17.1. Various Concepts of Agents.

Large-scale agent-based simulations have become possible only relatively recently, with the advent of fast, cheap, and readily available computers. The approach has been championed by physicists using the paradigm of computational statistical physics. De Oliveira et al. [17.1] review several papers from the past few years that exemplify the methodology, especially the work of Levy, Levy, and Solomon [17.2]. This opens the door to the study of the interaction of large numbers of heterogeneous, interacting agents. In this paper, we will introduce a simple framework for agent-based simulation and three applications: a commodities market, a dynamic online auction, and international greenhouse gas emissions trading.

17.2 Agent-Based Simulation Framework: ASIA For effective implementations of the agent-based economic and social simulations, we developed a simple framework, Artificial Society with Interacting Agents (ASIA), using Java. This framework provides only very simple and fundamental functionality for social simulations. Recently, a lot of researchers have begun to investigate agent-based simulations or artificial markets. Also a number of agent systems or frameworks have been proposed to systematically implement models. Many of these frameworks aim at constructing unified structures with object-oriented design methods (For example, [17.3]) and some of them also possess an intelligent collaboration mechanism using the network. On the other hand, our framework mainly determines the dynamic interactions and trading process as foundations, and leaves the concrete design of

144

H. Mizuta and Y. Yamagata

the agents’ hierarchy, social structure and individual strategy for the users. We believe that this difference mainly comes from differences in the agent concept as described in the introduction. We constructed our framework with a layered structure as shown in Fig. 17.2. The agent layer contains a basic agent class and the fundamental environment for the agents. The environment provides the fundamental facilities for agents and users to create agents, to dispose of agents, and to send messages through a MessageManager class.

Application Layer Social Layer Agent Layer

Trading Roles Creation Messaging

Java Virtual Machine

Fig. 17.2. Layer Structure in ASIA.

The MessageManager collects and distributes messages sequentially with its own thread according to the predetermined schedule. Agents also have their own threads to process the distributed messages. Thus, users of the upper layers can construct parallel communication among agents without worry about the message passing mechanism. The social layer describes the basic role of agents in the society and gives the example of message exchanges for trade. We implemented Central, Participant, and Watcher agents and a simple market process using RFB and BID messages. The Central agent creates, registers and initiates Participant agents and Watcher agents. Users can start, stop, and reset trading through the GUI window provided by the Central agent. One sample trade procedure can be executed as follows (see Fig. 17.3). To Info

Central Watcher RFB

RFB BID

Participant

BID

Participant

Fig. 17.3. Message Transactions in the Social Layer.

begin a trade, the Central agent sends a Request For Bid (RFB) message to each Participant. Upon receiving a RFB message, a Participant agent replies with a BID message. The Central agent collects all of the BID messages and proceeds to the trade transaction if the users have customized the descendant

17. Agent-Based Simulation for Economic and Environmental Studies

145

appropriately. Finally, each Watcher agent receives information about the trade and report it to the users in the desired format. The social layer only determines a formal procedure for trading and the users must customize the behavior of agents at the Application layer. In the following sections, we will give example applications using this framework.

17.3 Market Simulation The stability of prices in asset markets is clearly a central issue in economics. From a systems point of view markets inevitably entail the feedback of information in the form of price signals, and like all feedback systems may exhibit unstable behavior. K. Steiglitz and D. Shapiro created the price oscillation and bubbles in a simple commodity market with producer/consumer agents and two types of speculators [17.4]. H. Mizuta, K. Steiglitz and E. Lirov considered the stability in this model with various price signals and found that the antiweighted average of bid price stabilizes the market dramatically [17.5]. In this section, we reproduce the simulation model described in [17.5] with the ASIA framework. We use two commodities: food and gold. As descendant of the Central agent class, we consider a central auctioneer. There are three kinds of Participant agents. Regular agents produce food or gold and consume food; value traders and trend traders are solely speculators. One trading period is executed as follows. The auctioneer sends to each agent a Request For Bid (RFB) containing price signals. Consider first the case when the price signal is simply the previous closing price, as in [17.4, 17.6]. Based on this signal, the regular agents decide on their levels of production for that time step and speculators update their estimates. The agents then send bids to sell or buy. Finally, the market treats the submitted bids as a sealed-bid double auction and determines a single price which maximizes the total amount of food to be exchanged. In each trading period the regular agents can produce either food or gold. They make this production decision to maximize profit, but in a shortsighted way, based only on the current price and their production skills. Fig. 17.4 shows a screen shot of the system. The PriceAmount Watcher window shows two graphs showing the market clearing price and the trade volume. In our previous work we showed that the price oscillation with Regular agents is stabilized by introducing different price signals. On the basis of the simulation, we also gave analytical results on the simplified dynamical system with different signals in [17.5].

146

H. Mizuta and Y. Yamagata

Fig. 17.4. Market Simulation with Agent Framework ASIA, showing a price bubble.

17.4 Dynamic Online Auctions The use of online auctions is rising at a dramatic rate, and in general many segments of the economy are becoming granulated at a finer and finer scale. Thus, understanding behavior in auctions, and especially the interaction between the design of auctions, agent behavior, and the resulting allocations of goods and money has become increasingly important—first because we may want to design auctions that are as profitable as possible from the sellers’ point of view, but also because we may want to bid in auctions, or design computer systems that respond well to the loads that auctions generate. To investigate such dynamic interactions between heterogeneous bidders and the price formulation through successive auctions, H. Mizuta and K. Steiglitz developed an agent-based simulation of dynamic online auctions [17.7]. In this section, we re-implement the auction simulation on the ASIA framework. The model considers a single auction involvin the sale of one item by one seller to one of n bidders, who submit their bids over time in the interval [0, T ) to an auctioneer, who awards the item to the highest bidder at closing time. A bidder can submit more than one bid during the auction. We define the auctioneer as a Central agent and the bidders as Participant agents. The starting bid price is fixed at 1, and the duration of the auction is 500 time units. At the beginning of each auction, each bidder determines his first valuation of the item. At each time period 0 ≤ t < T , each bidder receives the status of the auction, can update his estimation on a fixed schedule or probabilistically, and can submit bids if the conditions for his strategy are satisfied. We consider two different types of bidders; early bidders, who can bid any time during the auction period, update their valuations continuously and compete strongly with each other, and snipers, who wait until the last moments to bid. We can briefly characterize the strategy of early bidders as watch/modify/bid, and that of snipers as wait/bid. An example auction simulated by the complete system is shown in Fig. 17.5.

17. Agent-Based Simulation for Economic and Environmental Studies

147

Fig. 17.5. Sample Auction Simulation with Agent Framework ASIA.

17.5 Greenhouse Gas Emissions Trading In this section, we consider the application of the agent-based simulation for the international greenhouse gas (GHG) emissions trading under the Kyoto Protocol (KP). To prevent global warming, 160 countries agreed to the KP on limiting GHG emissions at COP3 in 1997. KP sets targets for Annex I countries at assigned reductions below the 1990 levels, with the targets to be met during the commitment period 2008-2012. For example, Japan and the US should reduce 6% and 8% of their emissions, respectively. The KP allows international GHG emission trading, where countries who cannot reach the reduction targets can buy the emissions rights from other countries who can easily satisfy the target. Such a market mechanism is expected to reduce the worldwide cost for GHG reduction because of the large range in the marginal abatement cost curves (MACs) for reducing GHG emissions. In the previous two sections, we have applied the simulation to relatively traditional market systems, that is, a commodities market and an online auction. Now we will investigate the anticipated properties of an emerging new market through a simulation study. Such a study in advance is important to establish efficient rules, but difficult without simulation. J. Gr¨ utter [17.8] developed the CERT model which calculates the equilibrium price with various options and parameters for MACs. The CERT model treats only one trade in 2010 and each country must achieve the targets in that year. Because this model is implemented with a spreadsheet and macros, it is difficult to expand the model to treat successive trades and to assign different strategies to different countries. Now we have developed a prototype for GHG emissions trading with the ASIA framework. Because we modeled countries as agents, we can easily modify the behavior of each country and investigate the dynamic interactions between heterogeneous strategies. The structure of the simulation system is as follows. The COP agent is a descendant of the Central agent and manages the international trading. The Nation agents are descendants of the Participant agent and correspond to countries or groups. In this model, we created 12 Nations; 6 are Annex I countries and 6 are Non Annex I countries who are not assigned targets for reduction. Nations behave autonomously and independently to achieve the

148

H. Mizuta and Y. Yamagata

assigned KP targets with minimum costs or to receive maximum profits from the trades. Fig. 17.6 shows the basic trading procedure through message exchanges. We consider both a static equilibrium market with only one trade in 2010, as was discussed in [17.8], and dynamic market development through the commitment period 2008-2012. In each trading year, a COP agent sends Request for Bid (RFB) messages to all Nations which have an asking price. Upon receiving the RFB message, a Nation agent examines the asking price and his MAC to decide the amount of the domestic reduction. Then he sends back a Bid message to the COP agent which says how much he wants to buy or to sell at the asked price. After repeating this RFB-BID process, the COP model will find the equilibrium price where the demand and the supply balance, and send the Trade message to approve the trades for the year. Thus, the equilibrium price for each year is determined when the MAC functions and the assigned reductions of all of the participants are given. Price Supply

COP

RFB

Demand

RFB BID

BID

Nation1

Nation2

MAC1 P

MAC2 P

Domestic

Domestic

Fig. 17.6. Trading Procedure.

Then we considered multiple trading periods. Nation i divides up the assigned total reduction Ri for each trading period n = 0, 1, 2, . . . ,  Rin = Ri . n

As described previously, we can find the equilibrium price Pn∗ for each year using a partition of the assigned reduction Rin and a MAC function at this time. To consider the dynamics of MAC, we introduce a technology function tin (p) which gives the amount of reduction using the available technology at a given cost p for the Nation i at the year n. Then the MAC is given as the inverse function of the integral of the technology function. For each year, all countries determine the amount of the domestic reduction with which the values of MAC for all countries agree with one international value, that is, the equilibrium price, to minimize the worldwide reduction cost. Similarly, they try to minimize the total cost over the commitment period by choosing the partition Rin (n = 0, 1, 2, . . . ) for the assigned reduction

17. Agent-Based Simulation for Economic and Environmental Studies

149

which has the smallest variance in the differential coefficient of the total cost for each trading period. As a simple dynamic process for the reduction technology tin (p), we adopt reusability 0 ≤ α ≤ 1 and deflation 0 ≤ γ ≡ 1/β ≤ 1. Once the technology whose cost is lower than the price P ∗ is used, the reusability of the technology will be restricted with the coefficient α. On the other hand, technical innovations and deflation decreases the cost of each technology. With ∗ P¯in ≡ max{γin P0∗ , γin−1 P1∗ , . . . , γi Pn−1 }, we can obtain the technology function as  αi βin ti0 (βin p) p < P¯in tin (p) ≡ βin ti0 (βin p) otherwise. We set the initial technology function to be ti0 (p) with two coefficients ai and bi to reproduce the quadratic MAC function used in the CERT model, ti0 (p) ≡ 

1 b2i

+ 4ai p

.

In our simulation, we fixed the parameters {ai }, {bi } and {Ri } for the 12 countries as given in the CERT model and use randomly distributed {αi } and {βi }. Each Nation agent i determines the initial partition of the reduction {Rin } and updates the partition after the commitment period so that the variance of the marginal reduction cost decreases. Fig. 17.7 shows an example of the simulation result. Users can start, stop, and reset the trades and select the trading duration in the upper left window provided by the COP agent. This main window provides information for each Nation’s agents, and buttons to open a GUI window for each Nation. Two graphs in the lower left window show the movement of the equilibrium price and the trading amount. There are also graphs for the marginal reduction cost (upper) and the partition of the assigned reduction (lower) of two Nations representing USA (left) and Japan (right). By simulating the dynamic adjustment of the partition, we can see the worldwide cost reduction and the spontaneous selection of strategies. In this particular result, USA chose the late action strategy and Japan chose the early action strategy according to their estimation of rate of the technical innovation and other circumstances. We can observe changes of the total reduction cost for the entire world and for each country with the view shown in Fig. 17.8. In the beginning of the simulation, all countries fix their partition as the average value through the trading period, Rin = Ri /N (for all i). Then they determine the equilibrium price Pn∗ , the domestic reduction Din (Pn∗ ), and the trading amounts Tin (Pn∗ ) ≡ Rin − Din (Pn∗ ) for each trading period. Simultaneously, they calculate the marginal reduction cost

150

H. Mizuta and Y. Yamagata

Fig. 17.7. Dynamic GHG Emissions Trading over the commitment period, 2008– 2012.

 Pn + Tin (Pn∗ )/τn∗ where τn∗ ≡ j tjn (Pn∗ ). This marginal reduction cost represents the approximate effects of the partition on the total cost for each country. By adjusting the partition after all of the trades so that the marginal reduction costs becomes a constant value over the trading periods, each country expects that the total cost will be optimized. Though each country tries selfishly to decrease only its own cost, the total cost for the world can be reduced via this process as shown in Fig. 17.8.

Fig. 17.8. Changes of the total costs via adjustment of the partition.

17. Agent-Based Simulation for Economic and Environmental Studies

151

17.6 Concluding Remarks We have developed a dynamical simulation for the international GHG emissions trading with our agent-based simulation framework, ASIA. In a simulation study of the international emissions trading, we observed the price formulation for each trading year and the dynamic improvement of strategies which reduce the total cost. The implementation of various types of the agent-based simulation can be easily done with this framework, since it offers simple and fundamental facilities for agents including messaging, multi-threading, and an example of social negotiation transactions in separate layers. We designed the framework to be very simple following the well-known KISS (Keep it simple, stupid) principle, which enabled us to concentrate on the essential factor in the system and to investigate the dynamics. At this stage of development, we did not provide intelligence or the network functions for agents which most other frameworks require, because our fundamental concept of an agent does not necessarily require these facilities. However, we do think that a wide range of agent-based simulations can be constructed within this framework. However, we also consider it will be useful for some users if some of these options are available in the higher layer as components they can choose. These optional components for our framework remain for future work. Furthermore, much of the research and analysis required to evaluate GHG emissions trading are also left for the future. We believe that this preliminary work will help in the effective construction of the emerging international market and that such an agent-based approach will have more importance in the near future.

References 17.1 Oliveira, S. M. de, Oliveira, P. M. de, Stauffer, D. (1999): Evolution, Money, War, and Computers: Non-Traditional Applications of Computational Statistical Physics. B. G. Teubner, Stuttgart 17.2 Levy, M., Levy, H., Solomon, S. (1994): A Microscopic Model of the Stock Market; Cycles, Booms and Crashes. Economics Letter 45, 103–111 17.3 Iba, T., Hirokane, M., Takabe, Y., Takenaka, H., Takefuji, Y. (2000): Boxed Economy Model: Fundamental Concepts and Perspectives. The First International Workshop on Computational Intelligence in Economics and Finance (CIEF2000) 17.4 Steiglitz, K., Shapiro, D. (1998): Simulating the Madness of Crowds: Price Bubbles in an Auction-Mediated Robot Market. Computational Economics 12, 35–59 17.5 Mizuta, H., Steiglitz, K., Lirov, E. (1999): Effects of Price Signal Choices on Market Stability. The 4th Workshop on Economics with Heterogeneous Interacting Agents (WEHIA’99)

152

H. Mizuta and Y. Yamagata

17.6 Steiglitz, K., Honig, M. L., Cohen, L. M. (1996): A Computational Market Model Based on Individual Action. Chapter 1 in Market-Based Control: A Paradigm for Distributed Resource Allocation, S. Clearwater (ed.), World Scientific, Hong Kong 17.7 Mizuta, H., Steiglitz, K. (2000): Agent-based Simulation of Dynamic Online Auctions. Proceedings of the 2000 Winter Simulation Conference 17.8 Gr¨ utter, J. M. (2000): World Market for GHG Emission Reductions. Prepared for the World Bank’s National AIJ/JI/CDM Strategy Studies Program

19. Effects of Punishment into Actions in Social Agents Keji Suzuki Future University-Hakodate, Kameda-Nakano 116-2, Hakodate City, Hokkaido, 040–8655, JAPAN, E-MAIL:[email protected]

This chapter shows the agent based approach to solve the tragedy of the common. The tragedy of the common is known to treat the problem that is how to manage the limited common resource. In the agent-based approach, a meta-agent is introduced to restrict the activity of agents by charging levies. It is supposed that the meta-agent and the agents don’t know the payoff function explicitly. Under this setting, the meta-agent try to make levy plan to restrict the agent activity and the agents tries to make the prediction of payoffs for decision making. To create the levy plan and prediction of payoffs, the genetic algorithms are used in each agent. Throughout the experiments, the formation of the levy plan and the prediction of payoffs to avoid the tragedy are shown.

19.1 Introduction Agent based social behavior simulations are research field that treats complex game situations and examines artificial intelligence [19.1]. Social dilemmas are one of the complex game situations and suite to examine the intelligence of agents. In this paper, the Tragedy of the Common [19.2], which is one of the social dilemmas, is treated in the agent-based simulation. In this game, players use common limited resources to get the reward. If players behave based on the individual rationality, all players will face to tragedies loosing higher payoff. To avoid such tragedies, players have to make the relationship between other agents to prevent the selfish behaviors or change the problem structure, for example, changing the payoff functions. The proposed approach is kind of the changing problem structure. That is, the meta-agent is introduced to control the levy charging to the players [19.3]. In addition, it is assumed to all players doesn’t know the structure of payoff function explicitly. The assumption can be thought as reflecting a part of complex real situations. Under this assumption, the objective of the simulation is to show the effectiveness between the coevolved levy plan of meta-agent and payoff predictions of agents. In the next section, the problem structure of the tragedy of common is introduced. Then the proposed approach is described.

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 162− 173 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

19. Effects of Punishment into Actions in Social Agents

163

19.2 The Tragedy of the Common The tragedy of the common [19.2] is famous game problem as one of the n-persons social dilemmas [19.4]. This game enables for us to analyze the behaviors of players sharing common limited resources. Owing to the common resources are limited, higher activity of agents to get the higher payoff will become to bring lower payoff. The example of the payoff function is shown as follows; P ayof fi = ai (16 − a1 + a2 + a3 + a4 ) − 2ai

(19.1)

where, P ayof fi is payoff of agent i. ai represents the degree of activity of agent i. Here, 4 agents participate and 4 degrees of the activity, ai ∈ {0, 1, 2, 3}, is supposed. The payoff function becomes like as Table.19.1. Table 19.1. Payoff table of the Tragedy of the Common

ai

1 2 3

0 13 24 33

Total 1 12 22 30

agent activity expect agent i 2 3 4 5 6 7 11 10 9 8 7 6 20 18 16 14 12 10 27 24 21 18 15 12

8 5 8 9

9 4 6 6

Let’s consider the game in which the player decides own activity based on the individual rationality. The game assumes the activity of agents consuming the limited common resources. Therefore, the payoff becomes will decrease when total activities increase. However, the Agent i will increase the activity against any total activity of other agents, because the agent i can increase own payoff until the total activity reaching 11 in the example. Namely, the strategy of higher activity always dominates the strategy of lower activity. Thus all players will decide to increase their activities based on the individual rationality. Thus the decisions based on the rationality will cause the limited common resources being exhausted and all agents will be face to the tragedy. In the example, the tragedy arises when total activities reached 12. The characteristic of the game is known that no technical solution exists. Therefore, to solve this game, players should change the individual rationalities to other types of rationality or problem structures should be changed to payoff function. One of the objectives of the agent-based simulations is examined what kinds of rationalities and extended problem structures can avoid social dilemmas like as the tragedies. In this paper, the architecture of the proposed agent based simulation is belonging to the extension of the problem structure. Namely, the meta-agent is introduced to prevent the agents based on the individual rationality causing the tragedy. The detail of the proposed approach is described in next section.

164

K. Suzuki

19.3 Coevolving Levy Plan and Payoff Prediction 19.3.1 Approach To solve the social dilemma, especially the game type of the tragedy of the common, it is proposed that coevolution between the levy plan of the metaagent and the payoff prediction of agents. The approach is belonged to the extension of the problem structure. The charge of the levies can change the obtained payoff of players from the original payoff structure. Therefore even if the players decide their activities based on the individual rationality, suitable levy structure will prevent the activities exhausting the limited common resources. However, the issue of charging the levy approach is remained. That is how to set the suitable levy plan. The issue is connected to the planning policy of levy. In this approach, the individual rationality is employed for planning policy. Namely, the objective of the meta-agent, which will control the levy, is to maximize the incoming levy. While the individual rationality is simple and it isn’t required the meta-agent to have specific cooperative rationality, the characteristic of the meta-agent, it is afraid to increase the levy selfishly. To inhibit the selfish behavior of the meta-agent, simple payment rules of the levy from agents to meta-agent is set. The rules are that if the received reward, subtracted payoff from charged levy, become negative, the charged agent doesn’t pay the levy to meta-agent. This simple rule and other related rules could be expected to inhibit the selfish behavior of meta-agent. Related specification of the problem, it is assumed that the meta-agent and the agents doesn’t know the payoff structure. The agents are required to decide own activities without the information of other agent’s activities. The assumption will reflect the real complex situations. In real complex situations, we may not aware the similarity of the social dilemmas. Therefore, in the simulation model, meta-agent and the agents is expect to acquire the characteristic of the given dilemma structure by trial and error in the iterated games To acquire the hidden payoff structure, each individual agent tries to construct the prediction of payoff according to its activities. Because the prediction has to be constructed without information of other agent’s activities, the implicit synchronization between agents will be required. The implicit synchronization will be arisen from the charging levy by the meta-agent. The implicit synchronization of the agents means that the meta-agents can get stable incoming levies without the prevention of the charging rules. Therefore, the meta-agent also has to construct the suitable levy plan. The suitable levy plan means that the higher levy should be charged to a specific activity and lower levy should be charged to a recommended activity. According to the charged levy plan and their predictions, the agents will select their activity to maximize the rewards based on their individual rationality. Therefore, the meta-agent makes all agents to stably select activities related to higher payoff and charges to them adequate levy without loosing incomes. The relation of

19. Effects of Punishment into Actions in Social Agents

165

the plan of the levy and the prediction of the payoff is expected to avoid the tragedy situations because meta-agent can’t get enough levies from the tragedy situations. To realize the suitable levy plan and synchronized predictions of payoff, the coevolution mechanism is employed as adaptation ability of agents. Each agent and the meta-agent have independent population of chromosomes that represent the plan and predictions. Based on the evaluations function which reflect the individual rationality, the chromosomes are applied the genetic operations. Throughout the experiences of the iterated games, it is expected that the plan of meta-agent and the predictions of the agents will be fixed to avoid the tragedy situation and get higher payoff and levy. In the following subsections, the details of the proposed methods are explained. 19.3.2 Relation between Levy Plan and Payoff Prediction The meta-agent has the levy plan for acquiring the incoming levy from the agents. The Levy plan consists of the expected levy values according to the each agent’s activities. The all agents have the payoff prediction that consists of the values according to their activities. Both image of the levy plan and the payoff prediction are illustrated in Fig. 19.1. Because the payoff function and other agent’s activities are hidden, the levy plan of the meta-agent and the payoff prediction of the agent are limited material for making decision of L e v y

P a y o f f

P l a n

P r e d i c t i o n

A c t i v i t y

L e v y

p l a n

i m a g e

P a y o f f

o f

m e t a - a g e n t

A c t i v i t y

P a y o f f

P r e d i c t i o n

a 1

p r e d i c t i o n

L e v y

a 2

a 3

i m a g e

o f

a g e n t

P l a n

a 4 A c t i v i t y

Fig. 19.1. Schematic view of relation between levy plan of meta-agent and payoff prediction of agent.

166

K. Suzuki

the agents. The levy plan of the meta-agent is distributed to the agents in each game. The agents combined the accepted levy plan of meta-agent and their payoff prediction for the decision-making. The process of the decision-making is as follows: first, the agent combines the accepted levy plan and own payoff prediction. From the combined image, the agent,i, decides it’s own activity by probabilistic selection. Namely, probability of the activity aij ∈ Activity is determined from the predicted payoff at activity aij subtracting the value of the levy plan, Levyj . In this probabilistic selection of the activity, the negative probabilistic values are normalized to positive. Therefore, the activities that have higher payoff prediction and lower levy value in the image are relatively selected. 19.3.3 Reward of Agent and Incoming Levy of Meta-agent According to the decision making of the agents, total activity of all participate agents are determined. From the total activities, payoff value for each agent can be determined. The Fig. 19.2 is shown the evaluation process of reward for the agent. In this figure, the agent,i, is assumed to select activity, ai = a3 , based on the combined image of the payoff  prediction and the levy plan. If the relation between the total activities, i ai and the activity, ai become C2 in the figure, the reward for the agent is determined from the realized payoff value subtracting the levy value. In this case, the reward value becomes positive. However, if the agent will selects the activity, ai = a1 , the reward value becomes negative. When the reward value becomes negative, the requested levy value can’t pay to the meta-agent. Therefore, the reward for the agent is paid only if P ayof f (ai , T otal) > Levy(ai ) is satisfied. If the condition is satisfied, the reward value becomes in eq.19.2, otherwise the reward becomes 0. P a y off , L evy L os s of Ag en t i P a y off P r ed i cti on C1 C2

L evy P a l n

C3

R ewa r d of Ag en t i

a

l evy of Meta -Ag en t 1

a

2

a 3

a 4

Acti vi ty

Fig. 19.2. Determination process of reward for agent and incoming lavy for meta-agent

19. Effects of Punishment into Actions in Social Agents

Rewardi = P ayof f (ai , T otal) − Levy(ai )

167

(19.2)

The meta-agent can recieve the incomming levy from the agent i, Levyin (ai ), only if P ayof f (ai , T otal) > Levy(ai ). Otherwise the incomming levy becomes 0. Namely, the meta-agent can’t recieve the incoming levy if the requested levy over the realized payoff value. Therefore, the values in the levy plan will be expected to become lower values for getting incoming levies. 19.3.4 Evaluation of Game To receive enough reward values and incoming levies, suitable levy plan and payoff predictions must be constructed. To adjust the plan and the prediction, the loss values in the game are calculate as evaluation of the game. The value of loss for a game is determined as follows: Lossi = P ayof fexp (ai ) − Rewardi  (Levyin (ai ) − Levy(ai )) Lossmeta =

(19.3) (19.4)

i

where, Lossi is evaluation value of agent i with activity ai . Lossmeta is the summention of the losses related with the activity ai of the agent i. According to the received rewards, the incoming levies and the losses, the levy plan and the payoff predictions are adjusted in the coevolution process. 19.3.5 Coevolution of Plan and Predictions The whole game process is shown in Fig.19.3. Throughout the decision making, the judgment and evaluation, the rewards, incoming levies, and losses are determined. Based on the values, the evaluations of the plan, Emeta and predictions, Ei , are calculated as follows:

Emeta

Rewardi Ei = Lossi  i Levy in (a ) = i Lossmeta

(19.5) (19.6)

Using the evaluation values, each agent and the meta-agent execute the operations of GA to adjust the plan and predictions. All agents have the population of the chromosomes as the population of GA. The chromosomes represent the plan and predictions. The objective of each GA is to maximize the evaluation value. Namely, it is that maximizing the reward without the loss for the agents and maximizing the incoming levy without the loss for meta-agent. The schematic view of the coevolution process in Fig:19.4

168

K. Suzuki

P a y off P r ed i cti on Ag en t i -1

P a y off

L evy P l a n Ag en t i

P a y off

Ag en t i +1

P a y off

Meta -Ag en t

P a y off , L evy

L evy P a ln

a

a

1

a

2

a

3

a

4

Acti vi ty

a

1

a

2

a

3

a

4

Acti vi ty

a

1

a

2

a

3

4

a

Acti vi ty

a

1

a

2

a

3

4

Acti vi ty

D eci s i on Ma k i n g P a y off

a

P a y off

1

a

a

2

a

3

a

4

Acti vi ty

P a y off

a

1

a

2

a

3

a

4

Acti vi ty

a

1

a

2

a

3

4

Acti vi ty

Sel ect a cti vi ty

Ju d g m en t P a y off

E va l u a ti on C1 C2 C3

a

a

1

2

a

E i

=

E

m e ta

=

R e w a rd i L o ss i Σ i L e v y i n (a i ) L o ss

m e ta

a 4 Acti vi ty

3

T o ta l = Σ a i i P a y o f f i = f (a i , T o t a l )

Fig. 19.3. Game process including decision-making, judgment and evaluation with levy plan and payoff prediction.

Meta -Ag en t G A Ch r om os om e D ecor d i n g P a y off , L evy

L evy P a ln

a

a

1

a

2

a

3

4

Acti vi ty

G a m e

P a y off

a

1

P a y off

a

a

2

3

a

4

Acti vi ty

D ecor d i n g

a

1

P a y off

a

a

2

a

3

D ecor d i n g

Ch r om os om e

Ch r om os om e

G A

G A

Ag en t

i-1

Ag en t

4

Acti vi ty

i

a

1

a

2

a

a

3

4

Acti vi ty

D ecor d i n g

Ch r om os om e

G A Ag en t i +

1

Fig. 19.4. Coevolution between levy plan of meta-agent and predictions of agents.

19. Effects of Punishment into Actions in Social Agents

169

19.4 Simulation To confirm the effectiveness of the proposed methods for avoiding the tragedy situation in the social dilemmas, the simulation is executed. The payoff function is set as follows;

P ayof fi = ai (|A| × N −

N 

aj ) − 2ai

(19.7)

j

where A denotes the activity, A = 0, 1, 2, 3, 4. N is the number of agents. In this simulation, N is set as 4, 6 and 8 to examine the effect of the number of the agents. Each agent and meta-agent has 30 chromosomes. Each chromosome consists of 4 sections for each activity and levy. The length of the section is adjusted to represent the range of payoff function. The decoding of each section is summed up of 1’s value. According to the decoded plan and the predictions, the game, the tragedy of the common, is iterated 10 times. The averaged evaluation values are given as fitness of the chromosomes. The crossover and mutation are applied the chromosomes. The crossover rate is 1.0 and the mutation rate is 0.05. Under these parameters, coevolution of meta-agent and the agents are executed until 200 generations. 19.4.1 Game without Meta-agent To confirm the self-interesting rationality of agents, the simulations without the meta-agent are executed. The number of agents is 4 and 6 in these simulations. The results are shown in Fig.19.5 and Fig.19.6. In both figures, the acquired payoff predictions have larger value according to increasing the activity. Thus, the agents tend to select the higher activities in the game that can be seen from the histograms in the figures. Namely, the agents in both cases fail into the tragedy situations. 19.4.2 Simulations with Meta-agents To control the self-interesting agents for avoiding the tragedy situations, the meta-agent is introduced in the simulations. The size of agents are 4, 6t, and 8. One of the evolution processes of the meta-agent and 4 agents is shown in Fig.19.7. From this figure, all of the agents and meta-agent can succeed to get enough evaluation. Fig.19.8 and Fig.19.9 represent the results of the acquired payoff predictions, the levy plan and the histogram of selecting activities in the case of 4 agents and 6 agents. In both cases, the meta-agents set the levy plan of the activity 4 exceeding the payoff prediction value in the activity 4. It means that the meta-agents in both cases prohibit the agents from selecting the activity 4. The effects of the acquired levy plans can be seen in the histograms

170

K. Suzuki

Sel ecti n g Ti m es

E x p ected P a y off

9

60

8 50

7 6

40

5 30

4 3

20

2 10

1 0

0 1

2

3

1

4

2

3

Activity

Activity

(a )

(b )

4

Fig. 19.5a–b.Acquired payoff predictions of 4 agents without meta-agent (a) and histogram of selecting activities of agents (b).

E x p ected P a y off

Sel ecti n g Ti m es

18

7

16 6

14 5

12 10

4

8 3

6 2

4 1

2 0

0

1

2

3

4

1

Activity

2

3

4

Activity

(a )

(b )

Fig. 19.6a–b.Acquired payoff predictions of 6 agents without meta-agent (a) and histogram of selecting activities of agents (b).

4 50 8 0

4 00 70

3 50 60

3 00 50

250 4 0

200

3 0

150

20

100

10

50

0

0

G en er a ti on

G en er a ti on

(a )

(b )

Fig. 19.7a–b.Evolution process of meta-agent (a) and 4 agents (b)

19. Effects of Punishment into Actions in Social Agents

171

of the selecting activities. The agents didn’t select the activity 4. Therefore, the meta-agents succeed to control the agents avoiding the tragedy situations in these cases. The strategy of the meta-agents based on the self-interesting rationality evolves to get the higher levies in stable by avoiding the tragedy situation. E x p ected P a y off &

L evy P l a n

Sel ecti n g Ti m es 8

60 MetaAgent Agent1 Agent2 Agent3 Agent4

50

7 6

40

5 4

30

3

20 2

10 1

0

0

1

2

3

4

1

2

Activity

3

4

Activity

(b )

(a )

Fig. 19.8a–b.Acquired payoff predictions of 4 agents and Levy plan of meta-agent (a) and histogram of selecting activities of agents (b).

E x p ected P a y off &

Sel ecti n g Ti m es

L evy P l a n 6

60 Meta-Agent

5

Agent1

50

Agent2 Agent3 40

4

Agent4 Agent5 Agent6

30

3

20

2

10

1

0 1

2

3

Activity

(a )

4

0 1

2

3

4

Activity

(b )

Fig. 19.9a–b.Acquired payoff predictions of 6 agents and Levy plan of meta-agent (a) and histogram of selecting activities of agents (b).

In the above cases, the meta-agents succeed to control the activities of the agents. However, the situation is changed in the case of the number of agents becoming 8. The result of the 8 agents case is shown in Fig.19.10. The acquired levy plan prohibits selecting the activity 3 and some agents prohibit selecting the activity 4. Thus the almost agents can select the activity 4 and they sometimes close to the tragedy situation. That means, the strategy of the

172

K. Suzuki

meta-agent is changed in this case. If a fewer agents select the higher activity and the others select the lower activities, the agents selecting the higher activity will get large payoffs and the meta-agent will get higher incoming levy from these agents. Such situations didn’t occur in the previous cases. The meta-agent is aware of these situations in the evolution process. Thus the effective strategy of the meta-agent was changed in this case. Because the some agents prohibit selecting the highest activity, the complete tragedy situation is avoided. However, the self-interesting rationality causes to be close the tragedy situations in sometimes. E x p ected P a y off &

Sel ecti n g Ti m es

L evy P l a n

8

120

Meta-Agent A1 A2 A3 A4 A5 A6 A7 A8

100

80

7

6

5

4

60

3 40

2 20

1

0 1

2

3

Activity

(a )

4

0 1

2

3

4

Activity

(b )

Fig. 19.10a–b.Acquired payoff predictions of 8 agents and Levy plan of meta-agent (a) and histogram of selecting activities of agents (b).

19.5 Conclusion In this paper, the Tragedy of the Common, which is one of the social dilemmas, is treated in the agent-based simulation. In this game, the meta-agent prepares the levy plan base on the individual rationality. The agents make decisions based on the levy plan and their predictions of payoff. Throughout the coevolution of the plan and predictions in the simulation, the levy plan can prevent to select the activities of the agents toward to the tragedy situation in the case of the group of agents being small. However, the size of the agents becomes large, the strategy of the meta-agent is changed. The complete tragedy situation can be avoided but the agents sometimes close to the tragedy situations. This means it is remaining how to evaluate the closeness to the tragedy situation in the interaction between the meta-agent and agents.

References 19.1 Suleiman, R., Troitzsch, K. G., Gilbert, N., (Eds.) (2000): Tools and Techniques for Social Sience Simulation. Springer.

19. Effects of Punishment into Actions in Social Agents

173

19.2 Hardin, G., (1968) :The Tragedy of the Commons. Science 162:1243. 19.3 Yamashita, T., Suzuki, K., Ohuchi, A. (2001): Distributed Social Dilemma with Competitive Meta-players. Int. Trans. in Operational Research, Vol.8, No.1, 75–88 19.4 Yao, X., (1996): Evolutionary stability in the N-person prisoner’s dilemma. BioSystems, 37, 189–197.

20. Analysis of Norms Game with Mutual Choice Tomohisa Yamashita, Hidenori Kawamura, Masahito Yamamoto, and Azuma Ohuchi Hokkaido University, Kita13 Nishi 8 Kita-ku Sapporo Hokkaido, Japan {tomohisa,kawamura,masahito,ohuchi}@complex.eng.hokudai.ac.jp

In this paper, our purpose is to represent the establishment of the norm as the indirect sanction of mutual choice that individuals have the rights to refuse interaction. We introduce a mutual choice mechanism in the norms game [20.2, 20.8] instead of a direct penal regulation and then reformulate the norms and metanorms games with mutual choice. As a result, through an agent-based simulation, we confirm that the metanorm for mutual choice supports the establishment of the norm.

20.1 Introduction The aim of the norms game [20.2, 20.8] is to investigate the emergence and stability of behavioral norms in the context of a game with bounded rationality. The following definition of a norm was formulated by Axelrod: a norm exists in a given social setting to the extent that individuals usually act in a certain way and are often punished when seen not to be acting in this way. In the norms game, an individual player first decides whether to cooperate or defect. The payoff function of this alternative is similar to the N-person Prisoner’s Dilemma (N-PD) [20.3, 20.4]. If a player chooses to defect, some of the other players may observe the defection, and these observers may then choose to punish the defector based on the norm “punish those who defect.” If the defector is punished, the payoff is a very painful but the punisher has to pay an enforcement cost. The result of this game through an agent-based simulation with evolutionary approach was that the norm collapse but that, if the metanorm is introduced, the norm becomes established. The metanorm was defined as “one must punish those who do not support a norm (those who do not punish a defection).” The sanction applied in the norms game is that an individual player has the right to punish a defector, or in other words, to directly decrease the payoff of the defector. Do defectors readily agree to such enforcement of a sanction that punishes them and accept a decreased payoff without resistance? For example, a tax delinquent (defector) may not pay a penalty tax if there were no compulsory payment enforced by a centralized direct regulation mechanism. A tax delinquent may also be in arrears in his or her penalty tax. Therefore, a centralized direct regulation mechanism is necessary to compel a tax delinquent to pay the penalty tax. If there is no compelling power, defector would T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 174 − 18 4 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

20. Analysis of Norms Game with Mutual Choice

175

probably not support penal regulations against defectors. A penal regulation established by an individual would not be enforced. Therefore, it may seem strange to assume that an individual player has the right to punish a defector by directly decreasing the payoff of the defector without the backing of a centralized direct regulation mechanism. To avoid such a difficulty, we refer to the studies on partner selection in multiple IPD because the concept of partner selection in Prisoner’s Dilemma (PD) [20.3] can be considered a kind of sanction. In previous research, many partner selection mechanisms have been purposed: the ostracism option [20.9], the choice and refusal mechanism [20.1, 20.11], the mutual and unilateral choice [20.6], and the option of not game the playing [20.4, 20.6, 20.10, 20.12]. We pick up mutual choice because it does not need the right to directly decrease the payoff of other players. Although this mechanism is used for matching two players, we apply it to an N-person game in the next section. Therefore, under the situation that the payoff without game partners is lower than all payoffs with game partners, the mutual choice mechanism works as an indirect sanction because, if all players only refuse a player, the payoff of the player can be indirectly decreased although no players directly decrease. In this research, we introduce a mutual choice mechanism into the norms game instead of direct penal regulation and then reformulate the norms game with mutual choice. Furthermore, we introduce a metanorm based on mutual choice. In order to examine the influence of mutual choice, we observe the behaviors of players through an agent-based simulation.

20.2 Mutual Choice in Group Formation Although previous mutual choice [20.1, 20.7, 20.11] schemes were designed as matching mechanisms where two players play a PD game if both agree to play, we applied this mechanism to an N-person game. Here, we introduce the concept of “group formation [20.5],” which is the process of players choosing each other from within their respective groups and then interacting (playing N-IPD) with only members of the selected player’s group. A group is a subset of the overall player set, and each player can join only one group. The strategy of a player has two dimensions, boldness and vengefulness, in the same way as the original norms game. Let N = {1, .., i, .., n} be the player set, and Boldness Bi be the strategy of player i, which represents the degree of boldness to defect. Vengefulness Vi represents the degree of vengefulness to defection associated with the other players. 20.2.1 Norms Game with Mutual Choice The norm based direct sanction in the original norms game was changed to a norm based on mutual choice that instructs players to “refuse to interact (play

176

T. Yamashita et al.

the N-PD game) with defectors.” Players make decisions on group formation in random order. There are the alternatives of group formation, forming a new group, or joining an existing group. The procedure used for decision making in group formation is as follows. 1) At t-th iteration of group formation, the first player cannot join an existing group but has to form a new group. The players make decisions after the first player chooses one group k out of the group set G = {G1 , .., Gk , .., Gm }, where Gk is the set of the players that have already made a decision on group formation. Player i chooses one group based on the expected cooperation with each other player j(∈ N ), and this is denoted by πt (i|j) [20.1, 20.11]. This expected cooperation is used to determine which group is most tolerable. Given any player i, group k is tolerable for player i in iteration t, only if |G

k|

πt (i|j) |Gk |

≥ Vi .

(20.1)

We define the groups satisfying condition (20.1) as “tolerable groups.” If any groups are tolerable to player i, then player i makes an game offer to group k, whose average expected cooperation for player i is highest. 2) After the group choice of player i, the group k chosen by player i is given an opportunity to refuse or accept the game offer of player i. The players in group k decide by a majority vote whether to refuse or accept the game offer of player i. The player j in group k agrees to accept player i only if πt (j|i) ≥ Vj . If the majority of players agree to accept player i, group k accepts the game offer of player i and then player i joins group k. Player i is added to the group k as Gk ∪ {i}, which is the new group k. 3) If group k refuses player i, player i make a game offer to group l, whose average expected cooperation for player i is second highest. Player i continues making game offers until a group accepts its game offer or until all tolerable groups refuse its game offer. If player i is refused by all tolerable groups, player i forms a new group m + 1. A new group m + 1 including only player i is added to group set G, and then group set G is modified as G = {G1 , .., Gk , .., Gm , Gm+1 }. 4) After decision making for group formation, players in groups of more than two players play N-IPD with the players in the same group. In the initial iteration of group formation, prior to any interaction, all players have the same initial expected cooperation value π0 for each player. Expected cooperations are updated whenever N-PDs are played. Consider any player in group k, if player j is not in the group k that includes player i in the current iteration t, the expected cooperation value of πt (i|j) is not changed. On the other hand, if player j is in the group k that includes player i, they both play N-IPD in group k. In the N-IPD of group k, player i can observe the other player j’s decision and denote it as S(i|j). If player j cooperates at rate s in all iterations of N-IPD, player i denotes the decision making history of player j as cooperation rate S(i|j) = s (0 ≤ s ≤ 1). Player

20. Analysis of Norms Game with Mutual Choice

177

i’s expected cooperation value for player j is updated by taking the weighted average over player i’s decision making history with player j, πt+1 (i|j) = wπt (i|j) + (1 − w)S(i|j),

(20.2)

where the memory weight w controls the relative weighting of distance to recent decision making. Players can observe the decisions and update the expected cooperation values of only players in same group. 20.2.2 Metanorms Game with Mutual Choice The metanorm we adopt is “refuse to interact (play the N-PD game) with those who interact (play) with defectors.” The metanorms game with mutual choice is based on an extension of the norms game with mutual choice. When player i makes a game offer to group k, the players in group k make decisions on whether to accept or refuse player i. If the majority of players agrees acceptance of player i, the players opposing acceptance of player i consider the players agreeing acceptance of player i as players accepting a defector into the group. We define in group k the players agreeing acceptance of player i as Gagree and the players opposing acceptance of player i as k agree , where G = G ∪ Goppose and Gagree ∩ Goppose = φ. The players Goppose k k k k k k opposing acceptance of player i leave group k and form the new group Goppose k based on the metanorm. Goppose is assigned to Gm+1 , and then G is modified k ∪ as G = {G1 , .., Gk , .., Gm , Gm+1 }. Then, player i joins group k (Gk = Gagree k {i}). If group m + 1 includes only player j, player j makes a game offer to its tolerable groups based on above-described process of group formation.

20.3 Simulation Setup In this paper, because our purpose is to examine the influence of mutual choice on the norms game and the metanorms game, we concentrate on estaTable 20.1. Common parameters in the simulations of four cases. |Ck | represents the number of cooperating players in group k. Number of players Number of generations Number of mutual choices per generation Number of N-PDs per mutual choice Initial expected cooperation value π0 Memory weight w Mutation rate Payoff function of cooperators in group k Payoff function of defectors in group k Payoff for lone player Palone

50 10000 200 20 1.0 0.8 0.01 |Ck |/|Gk | 0.6 + |Ck |/|Gk | 0.01

178

T. Yamashita et al.

blishment and maintenance of a norm. We conducted simulations for both the norms game and the metanorms game under two initial conditions. The first condition is that a norm has already been established and that each player is not bold at all, that is, Vi = 1 and Bi = 0 ( ∀i ∈ N ). Under this condition, we examine whether it is possible to maintain a norm established by mutual choice. The second condition is that a norm is not established at all and each player is completely bold, that is, Vi = 0 and Bi = 1 ( ∀i ∈ N ). Under this condition, we examine whether it is possible to establish a norm by mutual choice. In our simulations, genetic algorithms are applied to evolve the player’s strategies. The two dimensions of a strategy, boldness Bi and vengefulness Vi , are each divided into 32 equal levels, from 0 to 1. Because 32 levels are represented by 5 binary bits, a player’s strategy needs a total of 10 bits, 5 bits for boldness Bi and 5 bits for vengefulness Vi . Each simulation is initialized with a population of all players. A simulation consists of a sequence of generations inter-spaced with genetic phases. Each generation consists of an iteration of the norms or metanorms games with mutual choice in which players make, refuse and accept game offers, that is, conduct the group formation and then play N-IPD. At the beginning of the genetic phase, each player’s strategy in a population is assigned a fitness equal to its average payoff given per payoff received. A partner for crossover is selected by means of a roulette wheel selection. Uniform crossover is accomplished between the strategies of a player and a partner to obtain a new strategy for one offspring. After that, the strategy of this offspring is subjected to mutation, where each bit is flipped one bit with a certain probability. It would be interesting to adopt “bandwagon effect [20.5]” using group size to the payoff function of the N-PD game, but our purpose is to examine the influence of the norm and the metanorm. Therefore, we do not adopt bandwagon effect to simplify our model and the payoff function of the N-PD game in each group depends on only the ratio of cooperating and defecting players. The important parameters and the payoff function of the N-PD game are shown in Table. 20.1.

20.4 Simulation 20.4.1 Maintenance of Norm First, we will explain the maintenance of the norm in the norms and metanorms games with mutual choice. The results of 10 runs are shown in Figs. 20.1 and 20.2. The 10 circles indicate the average boldness and vengefulness of all players after 10000 generations. The typical dynamics of the maintenance of the norm in each game is shown in Figs. 20.3 and 20.4. In all of the runs shown in Figs. 20.1 and 20.2, there was little boldness and a great deal of vengefulness. The initial condition was Vi = 1 and Bi = 0

20. Analysis of Norms Game with Mutual Choice

179

Fig. 20.1. Results of 10 runs of the maintenance of the norm in the norms game with mutual choice: average boldness and vengefulness of all players in 10000 generations under the initial condition Vi = 1 and Bi = 0 (∀i ∈ N ).

Fig. 20.2. Results of 10 runs of the maintenance of the norm in the metanorms game with mutual choice: average boldness and vengefulness of all players in 10000 generations under the initial condition Vi = 1 and Bi = 0 (∀i ∈ N ).

Fig. 20.3. Example of the maintenance of the norm in the norms game with mutual choice: transition of average boldness and vengefulness of all players under the initial condition Vi = 1 and Bi = 0 (∀i ∈ N ).

Fig. 20.4. Example of the maintenance of the norm in the metanorms game with mutual choice: transition of average boldness and vengefulness of all players under the initial condition Vi = 1 and Bi = 0 (∀i ∈ N ).

(∀i ∈ N ). Furthermore, in all runs the dynamics of average boldness and vengefulness of a population were similar to the typical dynamics shown in Figs. 20.3 and 20.4. Therefore mutual choice can maintain the norms in both the norms game and the metanorms game because little boldness and a great deal of vengefulness were kept throughout the generations. In the following explanation we represent a player with a high level of boldness as having Bhigh and a player with a low level of boldness as having Blow . In the same way, we represent players as having Vhigh and Vlow . The reason for the maintenance of the norm is as follows. The mutation of player strategies increases boldness or decreases vengefulness because initial condition there was little boldness and a great deal of vengefulness. The player with boldness increased by mutation, that is, the player with Bhigh , does not join the groups and then acquires a lower payoff because other players

180

T. Yamashita et al.

Fig. 20.5. Result of 10 runs of the maintenance of the norm in the norms game with mutual choice: average boldness and vengefulness of all players in 10000 generations under the initial condition Vi = 1 and Bi = 0 (∀i ∈ N ).

Fig. 20.6. Result of 10 runs of the maintenance of the norm in the metanorms game with mutual choice: average boldness and vengefulness of all players in 10000 generations under the initial condition Vi = 1 and Bi = 0 (∀i ∈ N ).

with Vhigh refuse the game offers of this player. Consequently, the player with an increased boldness acquires a lower payoff. The player with Bhigh cannot have a freeride on the player with vengefulness deceased by mutation, that is, the player with Blow and Vlow . The reason for this is that, if the player with Blow and Vlow join a group, other players with Vhigh in the same group would refuse the game offer of the player with Bhigh . Even if the player with Bhigh tries to have a freeride on the players with Blow and Vlow , they acquire lower payoffs. As a result, they are not selected in GA and perish. Although in this generation the number of players with Bhigh increases, in the next generation the players with Bhigh cannot have a freeride to acquire more payoffs than the players with Vhigh who cooperate with each other in the group. This is because the players with Blow and Vlow have perished. As a result, the number of players with Bhigh does not increase. Therefore, in the norms and metanorms games with mutual choice, the norm does not collapse and can be maintained. 20.4.2 Establishment of Norm Next, we explain the establishment of the norm in the norms and metanorms games with mutual choice. The results of 10 runs are shown in Figs. 20.5 and 20.6. The 10 circles indicate the average boldness and vengefulness of all players after 10000 generations. The typical dynamics of the establishment of the norms in each game is shown in Figs. 20.7 and 20.8. The norms game. In nine of the runs shown in Fig. 20.5, we can observe there a great deal of boldness but little vengefulness. Mutation of player stra-

20. Analysis of Norms Game with Mutual Choice

Fig. 20.7. Example of the maintenance of the norm in the norms game with mutual choice: transition of average boldness and vengefulness of all players under the initial condition Vi = 1 and Bi = 0 (∀i ∈ N ).

181

Fig. 20.8. Example of the maintenance of the norm in the metanorms game with mutual choice: transition of average boldness and vengefulness of all players under the initial condition Vi = 1 and Bi = 0 (∀i ∈ N ).

tegies decreases boldness or increases vengefulness because the initial condition is Vi = 0 and Bi = 1 (∀i ∈ N ). At first, we assumed that there was only one player with Blow and Vhigh . A player with boldness decreased by mutation, that is, a player with Blow , cannot acquire a higher payoff than the players with Bhigh because these players with Bhigh have a freeride on the players with Blow . Accordingly, the players with Blow do not increase in the next generation. A player with vengefulness increased by mutation, that is, the player with Vhigh , cannot acquire a higher payoff than the players with Bhigh because players with Vhigh do not join a group consisting of players with Bhigh . Therefore, the one player with Blow and Vhigh by mutation cannot acquire a higher payoff than the players with Bhigh and Vlow Consequently, this player is not selected in GA and perishes. Next, we assumed that there were plural players with Blow and Vhigh . If a group consists of only players with Blow and Vhigh , the group refuses the game offers of players with Bhigh . If a group consists of both players with Blow and Vhigh and players with Blow and Vlow , it is possible that a player with Bhigh would join this group and have a freeride. The player with Bhigh can join the group because while the players with Blow and Vhigh oppose acceptance of its game offer, the players with Blow and Vlow agree it. If the players with Blow and Vlow win the majority vote over the players with Blow and Vhigh , the player with Bhigh can join the group. The players with Blow cannot acquire higher payoffs than the free-rider. Consequently, they are not selected in GA and perish. Although there are plural players with Blow and Vhigh , the players with Bhigh and the players with Bhigh prevent the norm from establishing. The players with Bhigh directly prevent the norm’s establishment because they have a freeride on the players with Blow and Vhigh . The players with Blow and Vlow indirectly prevent the norm’s establishment because they accept game offers from the players with Bhigh who have a freeride on the players with Blow . Therefore, in the norms game the norm collapses and does not become established.

182

T. Yamashita et al.

In the remaining one run of Fig. 20.6, there was little boldness and a great deal of vengefulness. The reason for the failure to establish the norm was that the player with Bhigh can join the group consisting of both players with Blow and Vhigh and players with Blow and Vhigh . If there are players with Blow and Vhigh but no players with Blow and Vlow , the player with Bhigh cannot join the group and then they defect from each other. As a result, the player with Bhigh acquires a lower payoff than the players with Blow and Vhigh who cooperate each other. If the number of players with Blow and Vhigh increases and they predominate in the population for few generations before the number of players with Blow and Vlow increases by crossover or mutation, the norm becomes established. Therefore, since the simulation results (Fig. 20.5) show that the norm was established in only one out of ten runs, it is not impossible but difficult to establish a norm in the norm game with mutual choice. The metanorms game. In all runs shown in Fig. 20.6, there was little boldness and a great deal of vengefulness. In the norms game the establishment of the norm fails because the players with Blow and Vlow accept the game offer of the players with Bhigh . In the metanorms game, if the players with Blow and Vlow agree to accept the game offer of a player with Bhigh and the group as a whole also accepts it, the players with Blow and Vhigh leave the group based on the metanorm; they refuse to play the N-PD game with those who play with defectors. The metanorm prevents the player with Bhigh from having a freeride on the players with Blow and Vhigh . This is because, if the player with Bhigh joins the group, the players with Blow and Vhigh leave the group. As a result, if there are some players with Blow and Vhigh , they can form a group without the player with Bhigh . The players with Blow and Vhigh can acquire higher payoffs because they cooperate with each other. Throughout this process, the number of players with Blow and Vhigh increases and they predominate in the population. Therefore, the norm becomes established. In the norms or metanorms games with mutual choice, mutual choice can maintain the norm once the norm becomes established just as punishment does in the original games. The results of simulation also indicate the possibility of the maintaining the norm by mutual choice. In the norms game with mutual choice, the non-vengeful cooperators who cooperate with anyone and accept any game offers indirectly prevent from the establishment of the norm because the non-vengeful cooperators allow defectors to join the group. As a result, the norm collapses and does not become established. Therefore, it is not impossible but difficult to establish the norm by mutual choice the norm collapses and does not become established. In the metanorms game with mutual choice, although the non-vengeful cooperators accept the game offers of defectors and win the majority vote, the vengeful cooperators who play with neither the defectors nor the non-vengeful cooperators leave the group. Because the vengeful cooperators acquire higher

20. Analysis of Norms Game with Mutual Choice

183

payoffs more stably than the non-vengeful cooperators and the defectors and also because the number of vengeful cooperators increases in the genetic phase, the norm becomes established. Therefore, the metanorm concerning mutual choice supports the establishment of the norm just as in the original metanorms game.

20.5 Conclusion In this paper, rather than a direct sanction, we introduced mutual choice as an indirect sanction to the original norms and metanorms games. We proposed a norms game and a metanorms game with mutual choice by changing the original norm and metanorm based on mutual choice. In order to examine the influence of mutual choice, we picked up the maintenance and establishment of the norm. We conducted agent-based simulations under two initial conditions to study the possibility of maintaining and establishing the norm in the norms game and the metanorms game with mutual choice. As a result, we confirmed that mutual choice, as an alternative to the punishment of the original games, can maintain the norm once the norm becomes established. In the norms game with mutual choice it is not impossible but difficult to establish the norm by mutual choice. In the metanorms game with mutual choice the metanorm on mutual choice supports the establishment of the norm just as in the original metanorms game.

References 20.1 Ashlock D, Smucker S and Stanley A, Tesfatsion L (1996) Preferential Partner Selection in an Evolutionary Study of the Prisoner’s Dilemma. BioSystems 37(1-2), 99–125 20.2 Axelrod R (1986) An Evolutionary Approach to Norms. American Political Science Review 80 1095–1111 20.3 Axelrod R (1997) The Complexity of Cooperation. Princeton University Press 20.4 Dawes R M (1981) Social Dilemmas. Annual Review of Psychology 31 169– 193 20.5 Axtell R (1999) The Emergence of Firms in a Population of Agents: Local Increasing Returns, Unstable Nash Equilibria, and Power Law Size Distributions. The Brookings Institution, CSED Working Paper 3 20.6 Hauk E (1999) Multiple Prisoner’s Dilemma Games with(out) an Outside Option: An Experimental Study. Universitat Pompeu Fabra, Economics Working Papers 20.7 Hauk E and Nagel R (2000) Choice of Partners in Multiple Prisoner’s Two-person Prisoner’s Dilemma Games: An Experimental Study. Universitat Pompeu Fabra, Economics Working Papers

184

T. Yamashita et al.

20.8 Deguchi H (2000) Norm Game and Indirect Regulation of Multi Agent Society. Proceedings of Computational Social and Organizational Science Conference 2000 20.9 Hirshleifer D and Rasmusen E (1989) Cooperation in a Repeated Prisoners’ Dilemma with Ostracism. Journal of Economic Behavior and Organization 12 87–106. 20.10 Shussler R (1989) Exit Threats and Cooperation under Anonymity. Journal of Conflict Resolution 33 728–749 20.11 Tesfatsion L (1997) A Trade Network Game with Endogenous Partner Selection. In: Amman H M, Rustem B and Whinston A B (Eds) Computational Approaches to Economic Problems. Kluwer Academic Publishers 20.12 Yamagishi T and Hayashi N (1996) Selective Play: Social Embeddedness of Social Dilemmas. In: Liebrand W and Messick D (Eds) Frontiers in Social Dilemmas Research. Springer, Berlin, Heidelberg, New York, 337–362

21. Cooperative Co-evolution of Multi-agents Sung-Bae Cho Department of Computer Science, Yonsei University 134 Shinchon-dong, Sudaemoon-ku, Seoul 120-749, Korea [email protected]

In this paper, we propose a method to obtain strategy coalitions, whose confidences are adjusted by genetic algorithm to improve the generalization ability, in the process of co-evolutionary learning with a social game called Iterated Prisoner’s Dilemma (IPD) game. Experimental results show that several better strategies can be obtained through strategy coalition, and evolutionary optimization of the confidence for strategies within coalition improves the generalization ability.

21.1 Introduction Individual’s behaviors in social and economic systems are complex and often difficult to understand. Generally, individual’s action is motivated by certain stimulus, thereby the action mechanism can be a kind of dynamic system. So far, there has been much work on the complex phenomena which an individual in the dynamic systems shows from the perspective of game-theory, but it is difficult to deal with more realistic and complex models. Hence, we attempt to understand complex phenomena and systems from the view of evolution in the field of computer science. Among many economic and mathematical games, Iterated Prisoner’s Dilemma (IPD) game is simple but can deal with complex problems such as social and economic phenomena. Axelrod studied on the strategy between humans using IPD game [21.1]. Individuals in social and economic systems show adaptive behavior according to changing environment, because their behavior can be a kind of response to be able to adapt to the stimulus. Especially, immune system in biological systems is representative that shows the stimulus-response well. The immune system can defeat external invaders by gating his opponents to optimal antibody among many antibodies. In the field of co-evolutionary learning, there are many attempts to get better strategies by incorporating this property, and among them fitness sharing is one of the most well-known approaches [21.6]. In this paper, we propose a method to obtain better strategies to adapt to unknown environments, especially which can perform well against the unknown opponents in the IPD game. In order to deal with the problem, we introduce the strategy coalitions, which can be easily recognized in social and economic systems, and obtain them in the process of evolution of strategies. Here, a strategy coalition consists of better strategies extracted from T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 18 5− 194 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

186

S.-B. Cho

population. Each strategy in a coalition has the confidence that identifies the proportion of participation in determining the next move of the coalition. In order for the strategies in a coalition to behave adaptively to the changing opponent strategies, we make the confidences for strategies to be changed with his opponent using another evolutionary learning. Section 2 introduces the IPD game and evolutionary approach to model the game. Section 3 illustrates the evolution of confidences and gating of strategies in coalition to improve the generalization ability, and experimental results are shown in Section 5.

21.2 Evolutionary Approach to IPD Game One of the most well known games for modeling complex social, economical, and biological systems is the IPD game [21.2]. In the 2-player IPD game, each player can choose one of the two choices, defection (D) or cooperation (C). This game is non-zerosum and non-cooperative: One player’s gain may not be the same as the other player’s loss, and there is no communications between the two players. The game is repeated infinitely, and none of the players know when the game is supposed to end.

Table 21.1. Payoff matrix of the 2IPD game. T > R > P > S, 2R > T + P

Cooperate Defect

Cooperate R T

Defect S P

One of the most important issues in evolving game-playing strategies is their representation. There are two different possible representations [21.3, 21.7, 21.8], both of which are lookup tables that give an action for every possible contingency. In this paper, Axelrod [21.1] for the 2IPD game is used. In this scheme, each genotype is a lookup table that covers every possible history of the last few steps. History in such a game is represented as a binary string of 2l bits, where the first l bits represent the player’s own previous l actions (most recent to the left, oldest to the right), and the other l bits represent the previous actions of the other player. For example, during a game of 2IPD with a remembered history 2 steps, i.e., l = 2, one player might see this history: l = 2: Example history 11 01

21. Cooperative Co-evolution of Multi-agents

187

The first l bits, 11, means this player has defected (an ’1’) for both of the previous l = 2 steps, cooperated (0) on the most recent step, and defected (1) on the step before, as represented by 01. For the 2IPD game remembering l previous steps, there are 22l possible histories. The genotype therefore contains an action (cooperate “0,” or defect “1”) for each of these possible histories. Therefore, we need at least 22l bits to represent a strategy. At the beginning of the game, there are no previous l steps of play from which to look up next action, so each genotype should also contain its own extra bits that define the presumed pre-game moves. The total genotype length is therefore 22l + 2l bits. In the IPD game, each player can be regarded as an agent that has his own strategy, motivated from getting better payoff, and confidence within group. Agents can form a coalition as long as they can get more payoff than other agents or survive for long time. Properties of the agent are shown in Table 21.2 for the IPD game.

Table 21.2. Agent model to play 2IPD game. Property ID History Strategy BelongTo Confidence Rank

Role unique identifier keep previous moves information for next move information of coalition proportion of participation in move in coalition rank in coalition

21.3 Cooperative Co-evolution of Strategies 21.3.1 Forming Coalition It is very hard to find one fixed strategy that can play game adaptively against changing opponents in the IPD game. Several methods such as utilizing multiple better strategies such as gating have been widely used to improve generalization ability. Speciated strategies in the IPD game can be obtained by some sophisticated evolution like fitness sharing [21.6]. In this paper, we attempt to obtain the better strategies during the gameplaying with the idea of coalition. In social and economic systems, individuals often form a strategy coalition to get better interest than other individuals or survive. In the IPD game, multiple strategies can form coalitions as the

188

S.-B. Cho

same motivation. We can define the condition that the coalition of strategies can be formed with as follows. The two better strategies belongs to the same coalition, 1. when the game between them brings bad payoff, or 2. when combining them results in good payoff. In either cases, two strategies must be different because we do not have to duplicate the same strategies in a coalition. After that, confidence is given to each agent in proportional to his ranking. This confidence has an important role of determining the rate of participation in the move of coalition. Confidence that determines the proportion of participating to the move of coalition is given to each agent. The next move of coalition is determined by the sum of these confidences of agents belongs to it. 21.3.2 Evolving Strategy Coalition In order to evolve coalition, coalition below the average fitness of agents in the population should be removed and new coalition should be generated from crossover of coalitions in the evolutionary process. In this case, crossover exchanges the agents in coalition. The coalition maintains better agents and removes worse agents from the population. Hence, only strong agents are maintained in the population, and new agents are generated by mixing them within coalition to keep the population from being evolved by weak agents. Figure 21.1 shows the procedure to generate new agents using those in the coalition. Two agents are selected at random among agents within coalition, and their strategies are mixed as the same number of agents in the coalition.

Fig. 21.1. Generation of new agents by mixing agents within coalition to prevent the population from being evolved by weak agents.

21.3.3 Gating Strategies in Coalition Each agent has a confidence to determine the proportion to the move of coalition. The coalition of fixed confidences would disappear in the course of

21. Cooperative Co-evolution of Multi-agents

189

evolution, because the coalition would have the difficulty to adapt to the changing opponents. To solve this problem, we adjust the confidences of agents to be able to perform well against changing opponents. To improve the adaptivity of coalition, techniques such as opponent modeling and gating can be used. Opponent modeling is to model and guess the opponent’s strategy, and then change his strategy to be optimal against current opponent. Since this method has difficulty to model opponent’s strategy precisely, Darwen and Yao propose a gating method to improve the generalization ability [21.6]. In this method, the optimal strategy in the last population plays against opponent by looking for similar strategies as opponent in the last generation of population. This paper uses strategy coalition that has history table for his and opponent’s moves and use the information to change confidences of strategies according to the change of opponent’s action. This has advantage of finding optimal action in the given moves kept in the history. Figure 21.2 shows the modified IPD game structure including the evolution of confidences. The confidences in a coalition are randomly initialized as real numbers from zero to two. The confidence table contains all the confidences of agents for possible combination of history. The training set for adjusting confidences consists of several well known strategies such as TFT, Trigger, CDCD, and so on [21.4, 21.5]. In the evolution, the confidences leading to good result are selected among population of coalitions. Crossover exchanges the confidences between coalitions selected from the population, and mutation changes a specified confidence into a random real number from zero to two.

Fig. 21.2. The components of game for evolving the confidences of coalition.

190

S.-B. Cho

21.4 Experimental Results In this paper, we have conducted two experiments in 2IPD game with the conventional payoff function. The first one is to obtain strategy coalition using co-evolutionary learning and the second one is to evolve the confidences of obtained coalition through another co-evolutionary learning. 21.4.1 Evolution of Strategy Coalition To obtain strategy coalition we use the population size of 50, crossover rate of 0.6 and mutation rate of 0.001. One-point crossover with elite preserving strategy is also adopted. History size is 2 and maximum number of agents within a coalition is one third of population. The number of coalitions in the population is restricted under 10. Figure 21.3 shows the average fitness of coalitions in the evolutionary process. In the beginning of the evolution, average fitness of coalitions is higher than that of agents in the population. However, this difference decreases as time goes by. It does not mean that adaptivity of coalitions decreases, but that agents in the population do not know how to play against the opponents in the beginning of the game, because they are initialized at random. However, as time goes by, many agents learn how to deal with opponent’s move. In other words, agents in the population also gradually evolve to adapt for their environment. 4 3 .5 3

F itn e s s

2 .5 2 1 .5 1 0 .5 0 0

2 0

4 0

6 0

8 0

1 0 0

G e n e r a tio n

Fig. 21.3. Average fitness of coalitions and agents in the population. Solid lines are for coalitions and dashed lines are for agents.

21. Cooperative Co-evolution of Multi-agents

191

21.4.2 Gating Strategies For the experiment of adjusting confidences of agents in coalition, we have the population size of 50 and one-point crossover rate of 0.6. Also, mutation rate is 0.001 and μ-λ selection with elite preserving is used. History size is two and training set consists of well-known seven strategies and a random strategy. Table 21.3 explains the strategies in the training set, and the agents in coalition that have resulted from evolution of strategy coalition are listed in Table 21.4. For the test of generalization ability of evolved coalition, we have selected 30 agents that are top ranked in the population of 300 (as shown in Table 21.5), and conducted ten times runs that evolved coalition plays 2IPD games in round-robin. Table 21.3. Training set for evolving confidence of coalition. Strategy TFT Trigger

TF2T AllD CDCD CCD C10DAll Random

Characteristics initially cooperates, and then follows opponent initially cooperates, but once opponent defects continuously defect similar to TFT, but defects for opponent’s 2 defection always defects cooperates and defects in turn cooperates two times and defects cooperates before 10 rounds and then always defects moves at random

Table 21.4. An example of agents in coalition. History 0100 1111 0000

Lookup Table 0101110110111101 0101101010011111 0000101010110101

192

S.-B. Cho

Table 21.5. 30 opponent strategies that are extracted from the initial population and top ranked in the population. History Lookup Table History Lookup Table 1000 0111101111101110 1111 1101111001110011 0100 1101111111110011 1100 0101100111010001 1000 1111111000000111 0111 0011111101010010 0111 0111111110110111 1111 0111011101010111 1100 0111001101010011 1111 0101101100111100 0001 1011111001011100 0111 1001110101010110 0001 0011010111111110 1101 0001000110011010 1100 0001110110110010 0010 1101110011111101 1000 1101110011011101 1001 1001100101011000 0010 1101101111000110 1000 1101000111101010 1111 1111101100011011 0110 1001011001110110 0110 0111010011011111 1011 0101111101110010 1010 1011010111111100 0001 0011110111011000 1101 1101000101011110 1110 0111010101110001 1100 1101011110111011 1011 0101011111110100

In the experiments, the fitness of coalition increases gradually, and the coalitions show the adaptive behaviors that they cooperate against the conditional cooperators such as TFT, Trigger and TF2T, and defect against defectors. Coalitions defeat or tie with C10Dall or CDCD strategy and always defeat the random strategy. Figures 21.4 is an example result of evolving confidences. In the test of generalization ability of strategy coalition, the confidences are varied with changing opponents. Experimental results indicate that obtained coalition through evolving confidences of strategies performs better than most of the training set, except AllD and Trigger, in the 2IPD game with the top-ranked 30 opponents in the initial population as shown in Table 21.6.

21.5 Concluding Remarks We use the strategy coalition to obtain several better strategies in IPD game. Strategy coalition consists of agents and has confidence of each agent. This confidence has an important role in determining the next move of coalition. We have obtained the strategy coalition using co-evolutionary learning, and evolved the confidences to adapt well-known training set using genetic algorithm. In the simulation results, evolving coalitions show the adaptivity that they cooperate in the game with conditional cooperators such as TFT,

21. Cooperative Co-evolution of Multi-agents

193

2 .5 5 2 .5

A v e r a g e F itn e s s

2 .4 5 2 .4 2 .3 5 2 .3 2 .2 5 2 .2 0

2 0

4 0

6 0

8 0

1 0 0

1 2 0

1 4 0

G e n e r a tio n Fig. 21.4. Average fitness of strategy coalition. Table 21.6. Performance against opponent strategies. Strategy Before After TFT Trigger TF2T AllD CDCD CCD C10Dall

Wins 8.64±4.9 18.55±0.5 8 30 7 30 0 0 27

Ties Avg. 6±2.191.84±0.28 4±0.632.16±0.07 0 1.70 0 2.13 0 1.54 0 2.17 0 1.05 0 0.91 0 1.97

Opp. Avg. 1.75±0.59 0.92±0.29 1.77 0.80 2.40 0.7 2.75 3.34 1.12

Trigger and TF2T, but defect for AllD strategy. Besides, coalition defeats random strategy and defeats or ties with CDCD and C10Dall strategies. In the test of generalization ability with the evolved coalition, we can see that they play better than training strategies except AllD and Trigger in the game with top-ranked 30 strategies of the initial population. Although we have used the 2-player IPD game in this paper, some of the results we have obtained may be applicable to more complex games. For example, it is interesting to investigate how coalitions could be formed among different countries in the world, how coalitions could be formed among different parties in a country, how coalitions could be formed in the commercial market, etc.

194

S.-B. Cho

Acknowledgement. This work was supported by Korea Research Foundation Grant (KRF-2000-005-C00012).

References 21.1 R. Axelrod, “The evolution of strategies in the iterated prisoner’s dilemma,” in Genetic Algorithms and Simulated Annealing (L. Davis, ed.), ch. 3, pp. 32–41, San Mateo, CA: Morgan Kaufmann, 1987. 21.2 A. M. Colman, Game Theory and Experimental Games, Oxford, England: Pergamon Press, 1982. 21.3 X. Yao and P. Darwen, “An experimental study of N-person iterated prisoner’s dilemma games,” Informatica, vol. 18, pp. 435–450, 1994. 21.4 R. Axelrod, The Evolution of Cooperation. New York: Basic Books, 1984. 21.5 R. Axelrod and D. Dion, “The further evolution of cooperation,” Science, vol. 242, pp. 1385–1390, December 1988. 21.6 P. J. Darwen and X. Yao, “Speciation as automatic categorical modularization,” IEEE Transactions on Evolutionary Computation, vol. 1, no. 2, pp. 101– 108, 1997. 21.7 Y. G Seo, S.-B. Cho and X. Yao, “Emergence of cooperative coalition in NIPD game with localization of interaction and learning,” Proc of Congress on Evolutionary Computation (CEC’99), Piscataway, NJ USA), pp. 877–884, IEEE Press, 1999. 21.8 Y. G. Seo, S.-B. Cho and X. Yao, “The impact of payoff function and local interaction on the N-player iterated prisoner’s dilemma game,” Knowledge and Information Systems: An International Journal, vol. 2, no. 4, pp. 461– 478, 2000.

22. Social Interaction as Knowledge Trading Games Kazuyo Sato and Akira Namatame Dept. of Computer Science, National Defense Academy Yokosuka, 239-8686, JAPAN, {g40045,nama}@nda.ac.jp

In this paper, we propose knowledge transaction as basic constitutes of social interaction. Knowledge transaction among agents with heterogeneous knowledge are formulated as knowledge trading games. Each agent has idiosyncratic utility function defined over his private knowledge and common knowledge shared with the other agents. We consider two types of the utility functions, the convex and concave utility functions. The knowledge transaction are formulated as symmetric and asymmetric coordination games with the combination of the trading agents with those different types of the utility functions. Knowledge transaction in an organization are formulated as the continuous of heterogeneous games. We investigate what characteristics of an organization promote knowledge transaction or discourage sharing common knowledge.

22.1 Introduction The study of knowledge creation has begun to gain a new wave. Nonaka and his colleagues has developed a new theory of organizational knowledge creation [22.11]. They focus on both explicit knowledge and implicit knowledge. The key to knowledge creation lies in the mobilization and conversion of tacit knowledge. They emphasize knowledge creation in two dimensions, epistemological and ontological knowledge creation. A spiral emerges when the interaction between tacit an explicit knowledge is elevated dynamically from a lower ontological level to higher levels. The core of their theory lies in describing how such a spiral emerge. They present the four modes of knowledge conversion that are created when tacit and explicit knowledge interact with each other. The four modes, which they refer to as socialization, externalization, combination, and internalization, constitute the engine of the entire knowledge creation process. These modes are what the individual experience. They are also the mechanisms by which individual knowledge gets articulated and amplified into and throughout the organization. The goal of our research is to formalize an economic model of knowledge creation by focusing the quantitative aspects of the value of knowledge. We classify knowledge into two kinds, one is shared knowledge, which is common to each other. This kind of knowledge can be transmitted across agents explicitly. The other type of knowledge is private knowledge. It is personal knowledge embedded in individual experience or knowledge creation. In T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 195− 207, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

196

K. Sato and A. Namatame

this paper, we focus on common knowledge and private knowledge as basic building blocks in an complementary relationship. More importantly, the interaction between these two forms of knowledge is the key dynamics of knowledge creation in the organization of agents. Knowledge creation both at the individual and organizational level is a spiral process in which the above interaction takes places repeatedly as shown in Fig.22.1. In an organization, the individual interacts with other members through knowledge transaction. Knowledge creation takes place at two levels: the individual and the organization, and knowledge creation consists of the forms of knowledge interaction and the levels of knowledge creation. We consider an organization of agents with heterogeneous knowledge, and knowledge transaction among agents constitute the basic foundation of interactions in an organization. Each member of an organization with private knowledge desires to accumulate both private knowledge and common knowledge. Agents exchange their private knowledge and the transacted knowledge is shared as common knowledge, which also accelerate agents to accumulate their private knowledge. Both private knowledge of each agent and common knowledge in an organization can be accumulated through knowledge transaction. Agents benefit by exchanging their private knowledge if their utility will be increased. At knowledge transaction, each rational agent mutually exchanges his private knowledge so that his utility can be improved. Agents may consider sharing knowledge with others is important for cooperative and joint works, or they put the high value on hiding their private knowledge from other agents. Factors such as the value (worth) of acquiring new knowledge and the cost of sharing knowledge should be considered.

Cr ea ti on of N ew Kn owl ed g e

Sh a r i n g Com m on Kn owl ed g e

P r i va te Kn owl ed g e ( ¶ 1)

P r i va te Kn owl ed g e ( ¶ 2)

P r i va te Kn owl ed g e ( ¶n)

P r i va te Kn owl ed g e ( ¶n)

Kn owl ed g e Accu m u l a ti on

k n o w l e d ge t r a n s a c t i o n

Com m on Kn owl ed g e(K)

Fig. 22.1. The Process of Knowledge Creation through Knowledge transaction

22. Social Interaction as Knowledge Trading Games

197

22.2 Knowledge Transaction as Knowledge Trading Games As the tasks in an organization grow in complexity, the ways must be found to expand existing knowledge, which increase the opportunities of accessing other knowledge resources [22.2] [22.3] [22.5] [22.10]. Cooperative works, if it is by a team of engineers, or by a group of experts, also require coordination by sharing common knowledge. Many functions and tasks of computers are also carried out through transaction among autonomous agents [22.8] [22.12]. These agents need to have the rights of transparent access knowledge repositories. The knowledge repositories is the accumulated and common knowledge resources and that provides many users in the same organization to explore, to work with, and to discover. To support safe cooperation and sharing of knowledge, while preserving agents’ autonomy, agents should negotiate with each other on the access rights and deletion policies on knowledge or when necessary the rights are propagated. In this section, we formulate knowledge transaction as noncooperative games. We consider an organization of agents G = {Ai : 1 ≤ i ≤ N } with both private knowledge and common knowledge. They transact their valuable private knowledge with other agents, and the transacted knowledge can be shared as common knowledge. Agents may benefit by exchanging their private knowledge if their utility will be increased. Therefore in knowledge transaction, agents mutually trade their private knowledge if and only if their utilities can be improved. Each agent Ai ∈ G has the following two trading strategies: S1 : Trades a piece of his private knowledge S2 : Does not trade

(22.1)

We need to investigate the inductive reasoning process where each agent has different value judgments on trading. Factors such as the value (worth) of knowledge possessed by each agent, the loss for disclosing the knowledge to others should be considered. The associated payoffs of agent Ai when he trades a piece of knowledge are shown as the payoff matrix in Table 22.1. Depending on the payoffs, we can obtain the following four types of the optimal transaction rules agent Ai ∈ G (Case 1)

Ui1 > Ui3 ,

Ui2 > Ui4

(22.2)

In this case, the strategy S1 dominates the other strategy. The optimal strategy is then to transact his private knowledge without regarding the strategy of his trading partner. (Case 2)

Ui1 < Ui3 ,

Ui2 < Ui4

(22.3)

In this case, the strategy S2 dominates the other strategy. The optimal strategy is to not to transact without regarding the strategy of his partner.

198

K. Sato and A. Namatame

Table 22.1. The payoff matrix of agent Ai

Agent Ai

(Case 3)

S1 (transact) S2 (not transact)

Ui1 > Ui3 ,

Trading partne S1 (transact) S2 (not transact) Ui1 Ui2 3 Ui Ui4

Ui2 < Ui4

(22.4)

In this case, the optimal strategy is determined based on the strategy of his partner. If his partner transacts, the optimal strategy become to transacts, and he does not transact, the optimal strategy is not to transact. (Case 4)

Ui1 < Ui3 ,

Ui2 > Ui4

(22.5)

In this case, the optimal strategy also depends on the other agent. However, if he does transact, the optimal strategy is not to transact, and if he does not transact, the optimal strategy is to transact. In Case 3 and 4, the optimal strategy is obtained as the function of the strategy of his trading partner as follows: Let denote the possibility of the trading partner is given by p. Then the expected utility of agent Ai when he chooses S1 or S2 is given as follows: Ui (S1 ) = pUi1 + (1 − p)Ui2 Ui (S2 ) = pUi3 + (1 − p)Ui4

(22.6)

Then, agent will transact if the following inequality is satisfied: pUi1 + (1 − p)Ui2 ≥ pUi3 + (1 − p)Ui4

i = A, B,

(22.7)

By aggregating the payoffs in Table 1, we define the following parameter termed as threshold associated to each agent Ai ∈ G . θi = (Ui4 − Ui2 )/(Ui1 + Ui4 − Ui2 − Ui3 )

(22.8)

Then from the inequality in (22.7), Agent Ai will transacts his knowledge depending on the following two cases: (1)When Ui1 + Ui4 − Ui2 − Ui3 > 0 agentAi transacts if p > θi (2)When Ui1 + Ui4 − Ui2 − Ui3 < 0 agentAi transacts if p < θi

(22.9a) (22.9b)

22.3 Knowledge Trading as Symmetric and Asymmetric Coordination Games In this section, we show knowledge transaction can be formulated as symmetric or asymmetric coordination games, depending on the types of the

22. Social Interaction as Knowledge Trading Games

199

utility functions of the two agents. In symmetric coordination games, both agents gain benefit if they select the same strategy, on the other hand, they are better of if they choose different strategies in asymmetric coordination games. We define the utility function of each agent as the function both his private knowledge and the common knowledge. The utility function of agent Ai is defined as the semi-liner function both his private knowledge Ωi and the common knowledge K, such as; Ui (Ωi , K) = Ωi + vi (K), i = A, B,

(22.10)

The value X − vi (X) represents the relative value of agent Ai when he holds knowledge X as private knowledge or the common knowledge. If X −vi (X) > 0 , he puts a higher value on knowledge X as private knowledge. If vi (X) − X > 0 , he puts a higher value on knowledge X as the common knowledge. We also consider the following three types of the value functions: Definition: For a pair of knowledge X and Y , (X = Y ) (1) vi (X ∨ Y ) = vi (X) + vi (Y ) , and the value function vi (X) is linear. (2) vi (X ∨ Y ) ≥ vi (X) + vi (Y ) , and the value function vi (X) is convex. (3) vi (X ∨ Y ) ≤ vi (X) + vi (Y ) , and the value function vi (X) is concave. If the value function is convex, acquiring common knowledge satisfies the increasing returns. Increased common knowledge brings additional values: acquring more common knowledge means gaining more experinces of other agents and achieving greater understanding of how to achieve the common tasks. On the other hand, if the value function is concave, acquiring common knowledge satisfies the decreasing returns. We now consider a knowledge transaction between agent A with his private knowledge X and B with his private knowledge Y. The associated payoffs of both agents in Table 1 are given as follows : UA (S1 , S1 ) = ΩA − X + vA (X ∨ Y ) ≡ UA1 UA (S1 , S2 ) = ΩA − X + vA (X) ≡ UA2 UA (S2 , S1 ) = ΩA + vA (Y ) ≡ UA3

UA (S2 , S2 ) = ΩA ≡ UA4

(22.11)

UB (S1 , S1 ) = ΩB − Y + vB (X ∨ Y ) ≡ UB1 UB (S2 , S1 ) = ΩB − Y + vB (Y ) ≡ UB2 UB (S1 , S2 ) = ΩB + vB (X) ≡ UB3

UB (S2 , S2 ) = ΩB ≡ UB4

(22.12)

The above associated payoffs can be interpreted as follows: Once they decide to transact their private knowledge, it is disclosed to the other agent, and it becomes as common knowledge. When both agents decide to trade their private knowledge, the payoffs of both agents are defined as their values of common knowledge minus their values of private knowledge. If agent A does not transact, and agent B transacts, he receives some gain by knowing

200

K. Sato and A. Namatame

new knowledge Y. If agent A trades knowledge X and agent B does not trade, his private knowledge X becomes as common knowledge, and some value is lost. If both agents do not transact, they receive nothing. Knowledge trading have unique features which are not found in the commodity trading. With the knowledge trading, agents do not lose all the value of their traded knowledge. They also receive some gain even if they do not trade if the partner trades. Subtracting Ui3 from Ui1 , and Ui2 from Ui4 we define the following parameters: αA ≡ UA1 − UA3 = −X + vA (X ∨ Y ) − vA (Y ) βA ≡ UA4 − UA2 = X − vA (X)

(22.13)

αB ≡ UB1 − UB3 = −Y + vB (X ∨ Y ) − vB (X) βB ≡ UB4 − UB2 = Y − vB (Y )

(22.14)

Aggregating the payoffs, we define the following parameters which represent the values of integrating two independent knowledge Xand Y . αA + βA = vA (X ∨ Y ) − vA (X) − vA (Y ) αB + βB = vB (X ∨ Y ) − vB (X) − vB (Y )

(22.15)

The parameter βi , i = A, B, represents the difference of the values when they are private knowledge and common knowledge. If βi > 0, i = A, B, some value of knowledge is lost if it changes from private to common knowledge. If βi < 0, i = A, B, the value of knowledge increases if it is treated as common knowledge. The parameter αi + βi , i = A, B, represents the multiplier effect of knowledge X and Y . If the value functions vi (K),i = A, B, are convex, αi + βi , i = A, B, are positive, and if they are concave functions, they are negative. Depending the signs of the parameters αi , βi , i = A, B, the knowledge trading games can be classified into the following two types: (1) Symmetric Coordination Games: The value functions vi (K),i = A, B, are convex If the value functions vi (K),i = A, B, are convex, then we have αi +βi > 0 ,i = A, B. If both agents have the convex value functions, their value functions defined for common knowledge become to be the increasing return of the scale. In this case, the payoff matrix in Table 22.1, which satisfies the condition of (22.9a), can be transformed the payoff matrix in Table 22.2, which is known as a symmetric coordination game. The coordination game with the payoff matrix of Table 22.2 has two equilibria of the pairs of the pure strategies (S1 ,S1 ), (S2 ,S2 ), and one equilibrium of the mixed strategy [22.7] [22.8]. Absent an explanation of how agents coordinate their expectations on the multiple equilibrium, they are faced with the possibility that one agent expects one equilibrium and the other agent expects the other, and in this case, the coordination failure may occur by selecting the different strategy.

22. Social Interaction as Knowledge Trading Games

201

Table 22.2. The transformed payoff matrix of the knowledge trading (i) If αi + βi > 0, i = A, B, the value functions are convex, (ii)If αi + βi < 0, i = A, B,they are concave agentB S2 (no trade)

S1 (trade) agebtA

S1 (trade)

αB αA

S2 (no trade)

0 0

0 0

βB βA

(2) Asymmetric Coordination Games: The value function vi (K),i = A, B, are concave If the value functions vi (K),i = A, B, are concave, then we have αi + βi < 0, i = A, B. If both agents have the concave value functions, their value functions defined over the common knowledge become to be the decreasing return to the scale. In this case, the payoff matrix in Table 22.2 satisfies the condition of (22.9b), which is known as a asymmetric coordination game. The asymmetric coordination game has two equilibria of the pairs of the strategies (S1 ,S2 ), (S2 ,S1 ), and one equilibrium of the mixed strategy. Absent an explanation of how agents coordinate their expectations on the multiple equilibrium, they are faced with the possibility that one agent expects one equilibrium and the other agent expects the other, and in this case, another type of coordination failure may occur by selecting the same strategy.

22.4 Aggregation of Heterogeneous Payoff Matrices In this section, we consider the knowledge transaction in an organization of agents G = {Ai : 1 ≤ i ≤ N } . Each agent Ai has knowledge Xi to be transacted. The payoff matrix of each agent also depends on the knowledge to be transacted. In trading games where there are many agents with heterogeneous knowledge, it is possible to reason about others only in the average. Therefore we assume that each agent reasons the other agents have the knowledge of the same value. Then each agent has the payoff matrix in Table 22.2 reflecting his judgement on the knowledge trading. We introduce the following parameter, defined as threshold of agent Ai : θi = βi /(αi + βi ) ≡ {Xi − vi (Xi )}/{vi (Xi ∨ Y ) − vi (Xi ) − vi (Y )} (22.16) where Y represents knowledge held by the trading partner of agentAi . The denominator of threshold in (22.16) represents the multiplier effect of sharing knowledge, and the numerator represent the cost of the trading.

202

K. Sato and A. Namatame

From the analysis of the previous section, we can classify the knowledge trading games into the following two types. (1) The value function vi (K) is convex. In this case, agent Ai plays the symmetric coordination games. Let suppose the proportion of agents in G who choose the strategy S1 is given by p(0 < p < 1). From (22.9a) we have the following optimal transaction rule of agent Ai , which is the function of his threshold θi . (i) : He should transact if p > θi (ii) : He should not transact if p < θi

(22.17)

(2) The value function vi (K) is concave. In this case, agent Ai plays the asymmetric coordination games. From (22.9b) we have the following optimal transaction rule of agent Ai ,which the function of his threshold θi : (i) : He should transact if p < θi (ii) : He should not transact if if p > θi

(22.18)

Then, we can classify agents with convex value function into the following three types depending on his threshold θi : (a) θi ≈ 0 (αi  βi ) : Hard-core of trading From the optimal transaction rule in (22.17) or (122.18), an agent with low threshold has the strategy S1 as a dominant strategy. He is willing to disclose his private knowledge without regarding the other agent’s strategy. Therefore, we define an agent with low thresholds are a hard-core of trading. (b) θi ≈ 1 (βi  αi ) : Hard-core of no trading An agent with high threshold has the strategy S2 as a dominant strategy. He does not trade his knowledge without regarding the other agent’s strategy. We define an agent with high threshold is a hard-core of no trading. (c) 0 < θi < 1 : Opportunist In this case, the optimal strategy depends on his partner’s strategy. Therefore we define this type of an agent as an opportunist. Each agent has idiosyncratic payoff matrix reflecting his own value judgements for knowledge trading. The payoff matrix of Table 22.2 is characterized by threshold defined in (22.16). Therefore, we aggregate of the heterogeneous payoff matrices, one for each member of the organization, and represent as the distribution of threshold. As examples, we consider several threshold distributions in Fig.22.2. An organization with the threshold distribution in Fig.22.2(a) consists of many hard-core of trading with low thresholds. An organization with the threshold distribution in Fig.22.2(b) consists of many hard-core of no trading with high thresholds. An organization with the threshold distribution in Fig.22.2(c) consists of opportunists with intermediate thresholds. An organization with the threshold distribution in Fig.22.2(d) consists of both hard-core of trading and hard-core of no trading.

22. Social Interaction as Knowledge Trading Games 0 . 0 3

203

0 . 0 3

0 . 0 2

0 . 0 2

f (θ )

f (θ )

0 . 0 1

0 . 0 1

0 0

0 . 2

0

θ

0 . 4

0 . 6

0 . 8

1

0 . 0 3

0

0 . 2

0 . 4

θ

0 . 6

0 . 8

1

0

0 . 2

0 . 4

θ

0 . 6

0 . 8

1

0 . 0 3

0 . 0 2

0 . 0 2

f (θ )

f (θ )

0 . 0 1

0 . 0 1

0 0

0 . 2

0 . 4

θ

0 0 . 6

0 . 8

1

Fig. 22.2a–d.Distribution Functions of Threshold in an Organization (a) An organization of hard-core of trading (b) An organization of hard-core of no trading (c) An organization of opportunistic agents (d) An organization of both hard-core of trading and hard-core of no trading

22.5 The Collective Behavior in Knowledge Transaction In this section, we investigate the long-run collective transaction in an organization. We provide the evolutionary explanations of studying the collective behaviors motivated by the works in evolutionary games [22.4] [22.9] [22.13]. At any given moment, a small fraction of the organization is exogeneously given opportunities to observe the exact distribution in the organization, and take the best response against it.

C o m p le x c o lle c tiv e b e h a v io r s

L oca l i n ter a cti on R a ti on a l a g en t

Fig. 22.3. Emergent Collective Behavior in Knowledge transaction

The heterogeneity of the organization G can be represented as the distribution function of their threshold. We denote the number of agents with the

204

K. Sato and A. Namatame

same threshold θ by n(θ) in G , which is approximated by the continuous function f (θ), defined as the density function of threshold of G . The proportion of agents whose threshold are less than θ is then given by f (λ)dλ (22.19) F (θ) = λ≤θ

which is defined as the accumulative distribution function of threshold in G. We characterize the collective behaviors classify into the following two types. 1

1

0 . 8

0 . 8

0 . 6 p (t ) 0 . 4

0 . 6

p (t )

0 . 4 0 . 2

0 . 2

0 0 0

1

0

2

r e p e a t

0 0

1

1

0 . 8

0 . 8

0 . 6

p (t )

1

0

2

r e p e a t

0

0 . 6

p (t ) 0 . 4

0 . 4

0 . 2

0 . 2 0

0 0

1

0

r e p e a t

2

0

0

1

0

r e p e a t

2

0

Fig. 22.4a–d.The dynamic process of the collective knowledge transaction (a) An organization of hard-core of trading (b) An organization of hard-core of no trading (c) An organization of opportunistic agents (d) An organization of both hard-core of trading and hard-core of no trading

(1) An organization of agents with convex value functions In this case, each pair of agents play the symmetric coordination games. We denote the proportion of the trading by the t-th transaction by p(t). Since the optimal transaction rule of an agent with the convex value function is given in (22.17), agents with the threshold satisfying p(t) ≥ θi trade at the next time period. The proportion of agents who trade at the next time period t + 1 is then given by F (p(t)) . Therefore the proportion of agents who traded can be described by the following dynamics: p(t + 1) = F (p(t))

(22.20)

The dynamics is an equilibrium at p∗ = F (p∗ )

(22.21)

As specific examples, we consider the knowledge transaction in the organization G with the threshold distribution functions in Fig.22.2. (Case 1-1) The distribution function of threshold is given in Fig.22.2(a).

22. Social Interaction as Knowledge Trading Games

205

The dynamics of the knowledge transaction in this case is shown in Fig.22.4(a). The dynamics has the unique sg equilibrium p = 1, where all agents transact. (Case 1-2) The distribution function of threshold is given in Fig.22.2(b) The dynamics of the knowledge transaction in this case is shown in Fig.22.4(b). The dynamics has the unique stable equilibrium p = 0, where no agent transacts . (Case 1-3) The distribution function of threshold is given in Fig.22.2(c) The dynamics of the knowledge transaction in this case is shown in Fig.22.4(c). The dynamics has the two stable equilibria p = 0 and p = 1. If the initial proportion who transact p(0) is greater than 0.5, then the dynamics converges to p = 1, on the other hand, if it is less than 0.5, it converges to p = 0. (Case 1-4) The distribution function of threshold is given in Fig.22.2(d) The dynamics of the knowledge transaction in this case is shown in Fig.22.4(d). The dynamics has the unique stable equilibrium p = 0.5, where a half of the agents transact their knowledge. (2) An organization of agents with concave value functions In this case, each pair of agents play the asymmetric coordination games. Let denote the proportion of the agents who transact at the t-th transaction by p(t). Since the optimal transaction rule of an agent with the concave value function is given in (22.18), agents with the threshold satisfying p(t) ≤ θ . Agents with thresholds greater than p(t), which is given by 1 − F (p(t)) will be transacted at the next transaction t + 1. Then, the proportion of agents who transact at the next time period is given by the following dynamics: p(t + 1) = 1 − F (p(t))

(22.22)

(Case 2-1) The distribution function of threshold is given in Fig.22.2(a) (Case 2-2) The distribution function of threshold is given in Fig.22.2(b) (Case 2-3) The distribution function of threshold is given in Fig.22.2(c) E 1

1

3

0 . 8

p(t )

0 . 8

0 . 6

E

2

p (t )

0 . 6 0 . 4

0 . 4

0 . 2

0 . 2

E 0 0

1

0

r e p e a t

2

0

1

0 0

1

0

r e p e a t

2

0

Fig. 22.5a–b.The dynamics knowledge transaction process with cycles (a), and with convergence (b)

With the above three cases, There is no equilibrium, and starting from any initial proportion p(0), the dynamics cycles between the two external points E1 : p = 0 and E3 : p = 1. Once it reaches to one of these extreme points,

206

K. Sato and A. Namatame

it visits each of them alternatively. With this cycles occurs, we have two situations, where all agents trade and no agent trades at the next time, and they repeat this cycle for ever. This phenomenon is known as a coordination failure. (Case 2-4) The distribution function of threshold is given in Fig.22.2(d) The dynamics has the unique stable equilibrium p = 0.5, where a half of the agents transact their knowledge, which has the same property with symmetric coordination games as shown in Fig.22.5(b).

22.6 Conclusion The aim of this paper was to formalize an economic model of knowledge creation by focusing the quantitative aspects of the value of knowledge. We classified knowledge into two kinds, shared knowledge and private knowledge. We focused on common knowledge and private knowledge as basic building blocks in an complementary relationship. The knowledge transaction were formulated as non-cooperative games. Different agents necessarily have different payoff structures. We proposed a new type of strategic games, heterogeneous games. We obtained and characterized the optimal transaction rules for each type of the transaction games . Through knowledge transaction, agents, can accumulate organizational knowledge as shared and common knowledge. We characterized the dynamic behavior of knowledge behavior in the long run. We obtained the completely different collective behaviors in the knowledge transaction with an organization of agents with the convex or concave value functions.

References 22.1 Arthur, R., Increasing Returns and Path Dependence in the Economy, Michigan University 22.2 Campbell, D., Incentives, Cambridge University Press,(1995). 22.3 Carley,K, & Prietula,M., Computational Organization Theory, Lawrence Erlbaum Associates,(1994). 22.4 Carlsson, H., & Damme, E., “Global Games and Equilibrium Selection”, ´ Econometrica, 61, 989¡Y1018,(1993). 22.5 Cohendet, P. , The Economic of Networks. Springer,(1998). 22.6 Fudenberg, D, & Tivole, J. ,Game Theory. The MIT Press,(1991). 22.7 Fudenberg, D, & Levine, D. , The theory of Learning in Games. The MIT Press,(1998). 22.8 Grosz,K, & Kraus,S., “Collaborative plans for complex action”, Artificial Intelligence, Vol.86, pp195-244, (1996). 22.9 Kaniovski, Y., Kryazhimskii, A., & Young, H.,“Adaptive Dynamics in Games Played by Heterogeneous Populations”. Games and Economics Behavior, 31, 50-96, (2000).

22. Social Interaction as Knowledge Trading Games

207

22.10 McKnight, L & Bailey, J (Eds)., Internet Economics, The MIT Press, (1997). 22.11 Nonaka.I,&Takeuchi,H., The Knowledge Creating Company, Oxford Univ. Press, (1995). 22.12 Rosenschein, J .S. and Zlotkin, G., Rules of Encounter, Designing Conventions for Automated Negotiation among Computers, The MIT Press, (1994). 22.13 Young, H.P., Individual Strategy and Social Structures, Princeton Univ. Press, (1998). 22.14 Zlotkin, G,& Rosenschein, J .S.,“Mechanism design for automated negotiation”, Artificial Intelligence, Vol.86, pp195-244, (1996).

23. World Trade League as a Standard Problem for Multi-agent Economics – Concept and Background Koichi Kurumatani1 and Azuma Ohuchi2 1

2

Cyber Assist Research Center (CARC), National Institute of Advanced Industrial Science and Technology (AIST Waterfront) Aomi 2-41-6, Koto, Tokyo, 135-0064, Japan [email protected] Graduate School of Engineering, Hokkaido University N13, W8 Kita-ku, Sapporo, 060-8628, Japan [email protected]

We propose a framework called “World Trade League,” which is expected to become a standard problem in multi-agent economics. The first purpose of World Trade League is to propose a network game in the context of economic and social systems. Such a game in World Trade League is executed by several players (countries), where each of country consists of heterogeneous agents such as product makers, service suppliers, financial companies, government, and so on. A player (country) participating in the game is evaluated according to its contribution to development of the international economic system and environment protection, in addition to the development of its own country. The second purpose of World Trade League is to provide a standard problem for pure multi-agent simulations in economic context which many researchers can commonly analyze. The software to execute World Trade League is supplied by X-Economy System, where X-SS protocol is used for the common communication protocol among agents.

23.1 Introduction Multi-agent approaches have been now the focus of researchers in modeling and analysis of economic systems in contrast with conventional economic approaches such as equilibrium theory and dynamical systems. Many researchers, however, model and analyze their own economic problems and seem to lack of sharing simulation results and computational techniques. In such a context, we propose a framework called “World Trade League” [23.1], which is expected to become a standard problem in multi-agent economics. We have mainly two purposes to propose a standard problem. The first one is to propose a network game which is executed by several players connected via networks. Each player represents a country, where each country consists of heterogeneous agents such as product makers, service suppliers, households, financial companies, central bank, government, and so on. T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 208 − 217, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

23. World Trade League

209

A player (country) participating in the game is evaluated according to its contribution to development of the international economic system and environment protection, in addition to the development of its own country. This regulation is designed to make the game in cooperative and ecologyminded atmosphere rather than selfish competition in international economic systems. The second purpose is to provide a standard problem for pure multiagent simulations in economic context where many researchers can commonly analyze the problem and share the simulation results and techniques. For both of the two purposes, the software to execute World Trade League will be supplied as common library X-Economy System [23.2, 23.3], where X-SS protocol [23.4] is used for the common communication protocol among agents. We design the computational framework to include other applications, such as education, training, entertainment, economic experiment, and so on. In this paper, we describe the concept and background of World Trade League, with the basic design of network games, simulation frameworks, and computational libraries and communication protocols.

23.2 Concept of World Trade League World Trade League is a game where each player prepares a country as a multi-agent system consisting of economic agents such as agriculture, manufacturing, distribution, finance, government, and so on [23.1]. Although World Trade League can provide several types of games by changing the configurations and regulations of the system, we explain a full set of the game in the rest of this paper. An agent in a country should behave as follows. – It collects public information which is open to any agent. – It does decision-making of what to do now (or do nothing), and it selects one of the options which is available now. These two parts are essential for agent design and implementation. In addition to them, the following regulations exist. – Agents in a country should behave independently, i.e., a country has no centralized control system. In order to achieve this restriction, communication among agents even in a country is open to public. Indirect controls are possible by such communication. – An agent in a county can make international trades with agents in other countries. This is done based on mutual agreement between two agents. It is a game where heterogeneous agents in several countries collect information, manufacture goods, make international trades, exchange currency, and compete with other countries in order to achieve its own economic development and international collaboration with protecting natural environment (Fig. 23.1).

210

K. Kurumatani and A. Ohuchi

I n t e r n a t i o n a l E c o n o m i c S y s t e m F o r e i g n

P l a y e r

N a t i o n

A

N a t i o n

B

C u r r e n c y

I n t e r n a t i o n a l T r a d e

M a r k e t

N a t i o n

N a t i o n

B

C

D P l a y e r

N a t i o n

D

E

Fig. 23.1. Concept of World Trade League.

23.3 Elements of World Trade League 23.3.1 Behavior Options of Agents and Market Structure Agents which constitute a country have the following behavior options. A) Manufacturing and Service Agents To plan the amount of goods/services production and to execute it, i.e., to borrow funds from bank, issue stock and bonds, purchase of material, invest in plant and equipment, hire labor power, produce, and sell product. B) Distribution Agent To carry goods and human. C) Bank Agent To loan funds to other agents after determining how much funds it loan to a specific agent. To collect deposits from other agents. To make trades in financial markets. D) Central Bank Agent To decide interest rate and to loan funds to banks. To make trades in markets. E) Government Agent To collect tax and distribute subsidy. To issue national bonds.

23. World Trade League

211

F) Natural Resource Agent This is a special agent to represent natural resources and degree of environmental pollution in a country. It is a passive agent and it does no active decision-making. In addition to agents, markets are prepared in the system. A market provides a place where a specific goods, funds, capitals, labor powers are traded by agents in the whole system. One market is prepared in the whole system according to one specific object to be traded, i.e., food material, food, industry material, industry goods, labor power, services, national bonds, bonds, stock, and so on. 23.3.2 Game Settings and Complexity In order to control the complexity of games played in World Trade League, we can make game settings in several ways. In World Trade League, the following ways are prepared to modify the complexity of games. A) Degree of Economic Evolution Complexity of games can be modified according to the degree of historical economic evolution as follows. 1. Medieval Stage: Currency exists, but no financial system like money loan exists. 2. Modern Stage: Indirect financial system, i.e., bank loan and national bond exist. 3. Contemporary Stage: Direct financial system, i.e., stock and company bond exist. B) The Number of Agents in a Specific Agent Type Complexity depends on the number of agents in a specific agent type. At the starting point, we use only one agent in a specific agent type in a country. C) The Number of Countries The number of countries (nations) participating in the game greatly influences the complexity of the game. If the number is one, the game becomes a self-contained game, i.e., an economic simulation of a country. The more nations participate in the game, the more complicated the game becomes. At the starting point, we assume that from two to five nations simultaneously participate in the game. D) Symmetric or Asymmetric Game 1. Symmetric Game: All nations are given the same initial condition at the starting of the game. 2. Asymmetric Game: The initial conditions of nations are different. By playing games several times with changing the role of nations, the overall condition for each nation can be made equal.

212

K. Kurumatani and A. Ohuchi

23.3.3 Evaluation Function of Players Unlike games proposed in previous multi-agent researches, several types of evaluation functions are prepared in World Trade League, i.e., we can run games or simulate economic systems in several types of boundary conditions. Players are evaluated by a single or a combination of the functions, e.g., the average of functions. Each of the function represents a certain aspect of the target economic system, e.g.: A) Economic Development of the Country It represents competitiveness of the world economic system. B) Imbalance of Economic Development among Countries It represents cooperativeness of the world economic system. C) Stableness of the World Economic System It represents stableness of the whole system, in order to avoid sudden changes in a country or in the whole system. D) Degree of Environment Protection It represents ecological coexistence of a country. E) The improvement of living standards The degree of living improvement of people in a nation. These evaluation functions are composed by 1) GNP 2) The amount of produced goods 3) Pollution degree of environment 4) Degree of distribution of produced goods, and so on.

23.4 Implementation 23.4.1 System Architecture In order to run games of World Trade League, we are now implementing a server, client class structure, and sample clients programs. The architecture of the system is shown in Fig. 23.2. Although only one country exists in this figure, several countries are connected to the server in the real games or simulations. The server consists of two modules, 1) Communication Control module, which controls all message transactions among agents, and 2) Database, which stores current status and all the history of each agent at micro-level and of the whole system at macro-level. A module which collects requests from agents and acts as a mediator, such as market, is called ‘medium’ in order to distinguish it from regular agents existing in a nation. All agents and mediums are prepared in class library, and users who wish to join the game can instantiate agent and medium instances from the library.

23. World Trade League

A g e n t M e d i u m C l a s s

B o n d M a r k e t

A g e n t

&

M e d i u m

S t o c k M a r k e t

213

I n s t a n c e s

B a n k

F o r e i g n C u r r e n c y M a r k e t

P r o d u c t M a r k e t

X - S S P r o t o c o l C o m m u n i c a t i o n C o n t r o l M o d u l e

D H i s t o r C o m m u n i c a t

a t y i o

a b a s I n f o n , A

e .

( T r a d e s , g e n t A c t i o n s )

A g e n t I n s t a n c e s P r o d u c e r ( A g r i c u l t u r e , I n d u s t r y )

X - E c o n S e r v e r S e r v i c e ( d i s t r i b u t i o n , I n f o r m a t i o n )

C e n t r a l

B a n k

T r a d e r s G o v e r n m e n t

Fig. 23.2. System Architecture of World Trade League.

23.4.2 Communication Protocol X-SS In World Trade League, a series of communication protocols called X-SS (eXtensible Social System) Protocol is prepared in order to reserve the extensibility of agent communications and game regulations [23.4]. Generally speaking, we have to prepare n(n − 1)/2 types of protocol when n types of agents exist. In addition to it, we have to prepare n types of new protocols when adding a new type of agent. This way of protocol definition clearly have problems in computational complexity, extensionability, clarity, and easiness in understanding. In X-SS, the protocol definition is not based on agent types, but on objects (goods, services, currency, or information) which are traded or exchanged in the game. For example, a trade of goods can be represented as an exchange of goods and currency, and collection and transmission of information as an exchange of information and currency. If you obtain free information, the price measured in currency should be set to zero. In addition to it, because one object to be traded or exchanged is almost always currency, we have to prepare just only m types of protocol where m is the number of objects to be exchanged in the system. Objects to be traded are as follows. – Currency: Unique in the whole system, or prepared for each country. Local currency can be additionally defined. – Goods: Food material, food, industrial material, industrial goods.

214

– – – –

K. Kurumatani and A. Ohuchi

Services: Transportation, amusement, general. Labor power Financial Goods: National bond, private bond, stock. Information

Using the protocol definition, an agent can be characterized by the goods which the agent can trade, and the class hierarchy of agents can be clearly defined. We prepare several ways to define and implement the protocol and communication module, because agents and server can be implemented in several kinds of programming language and they may use several kinds of communication infrastructure: – – – –

XML representation over networks in TCP/IP, UDP CORBA representation over networks in TCP/IP, UDP JAVA class library which represents message object C++ class library which represents message object

About the latter two types of message object, the communication between an agent and the server is carried out by 1) instantiating a message object instance from a message class, and by 2) calling proxy method implemented in the communication module of the server with setting the message object instance as an argument. The detail of protocol is, therefore, clearly defined as message class, and we can keep maintenancability and extensibility of the protocol.

23.5 Requirements for Standard Problem in Multi-agent Economics To propose standard problems in economic and social system research, we think that they have to satisfy the following requirements. – Validity of the problem as a model of real systems: Unlike the simulation in natural science, it is impossible to model and simulate the whole details of the target system in economic and social science. We have to extract the essence of the structure and behaviors of the target system, and to verify whether the simulation result can explain the essence of the target system. In other words, the setting of a standard problem should fit a suitable abstraction level of the target system. – Applicability of several techniques to the problem: A standard problem should be attacked by several types of techniques in social science, computer science, and artificial intelligence, e.g., dynamical system theory, game theory, agent-based simulation, machine learning techniques, and so on. In order to satisfy the criterion, the problem should be clearly described in computational sense and it should not include unnatural constraints.

23. World Trade League

215

– Complexity: A problem should be complex enough, so that we cannot find the best strategy to easily solve the problem. Standard problems in artificial intelligence such as Chess, Shogi, and Igo have enough complexity in contrast with the simplicity and clearness of the game definition and settings. – Closed problem rather than open problem: A standard problem had better be a closed problem rather than open one. Because an open problem is influenced by information brought from outside of the target world, there is a possibility that the information gap among agents exists. In other words, the quantity and the quality of information for each player can vary, it is difficult to keep fairness among the players. Such shortcoming caused by openness becomes the essential difficulty when executing fair network games and strict simulations. The game settings of World Trade League is designed to satisfy the above requirements.

23.6 Related Work Multi-agent approach to market analysis called artificial market research is one of the active research areas, and many kinds of analysis have been carried out. Many fruitful simulation results have been already obtained in this area. It seems difficult, however, to design network games in the context of market and trading. One of such network games is U-Mart, which is designed for a stock future market [23.5, 23.6]. This kind of approach has shortcomings that it is necessary to give information to agents from the outside of the system, e.g., so-called fundamental information (interest rate, benefit and business results of companies, perspective of national and international economics, etc.) must be unnaturally given to agents. This openness of the game crucially spoils the fairness among game participants. Some research projects to model and analyze a whole economic system as multi-agent system have started. For instance, Virtual Economy Project [23.7, 23.8] provides a basic economic database for SNA based on Exchange Algebra. The approach lacks of the idea of agent design and of extending the framework to multi-nation environment. Another approach to a whole economic system is Boxed-Economy Project [23.9, 23.10, 23.11]. They design the templates of economic agents as object class in detail, in order to construct a class structure of economic agents from the most genetic form to a specific one. Our approach in World Trade League is to design both computational framework and agent class structure simultaneously. In that sense, our approach contains both directions of the above two approaches, and it provides a flexible common framework for network game, simulation, education, training, entertainment, and economic experiment.

216

K. Kurumatani and A. Ohuchi

World Trade League as a network game has a common characteristics with RoboCup Soccer [23.12], because both games consist of multi-players and each player itself is a multi-agent system, although World Trade League is a heterogeneous multi-agent system. World Trade League and RoboCup Rescue [23.13] uses multiple evaluation functions in order to evaluate complex aspects of target social systems.

23.7 Conclusion We have proposed a framework called World Trade League which provides a standard problem in multi-agent economics. In the network game of World Trade League, several countries compete in order to achieve economic development of each own country with keeping cooperative relations with other countries and protecting environment. A question is frequently asked: “Why do you call the game ‘World Trade League’ instead of ‘World Trade Game’ ?” The answer is that we hope we find the path to achieve sustainable economic development in the game, with keeping the development of the whole world economic system, and with conserving the natural environment. The game should not become a field where a country which pursuits its own benefit obtain the best position. We are now designing the detail of the network game in World Trade League, implementing common libraries X-Economy System based on XSS protocols, and verifying the game as standard problem in detail. The regulations and game settings for public will be announced in the coming papers and on the web sites.

References 23.1 URL: http://www.w-econ.org 23.2 Kawamura, H., Yamamoto, M., Ohuchi, A., Kurumatani, K.: Development of X-Economy System for Introduction of Artificial Market; in the Proceedings of the First International Workshop on Agent-based Approach to Economic and Social Complex Systems (AESCS’01), pp.51–58 (2001). 23.3 URL: http://www.x-econ.org 23.4 URL: http://www.x-ss.org 23.5 Kurumatani, K., et al.: U-Mart: A Virtual Stock Market as a Forum for Market Structure Analysis and Engineering, in the Proceedings of the Fifth Joint Conference on Information Sciences, JCIS’00 (the First International Workshop on Computational Intelligence in Economics and Finance, CIEF’00), (Atlantic City), Vol.2, pp.957–960 (2000).

23. World Trade League

217

23.6 Sato, H., Koyama, Y., Kurumatani, K., Shiozawa, Y., Deguchi, H.: U-Mart: A Test Bed for Interdisciplinary Research into Agent-Based Artificial Market; Aruka, Y. (ed.): Evolutionary Controversies in Economics, pp.179–190, Springer-Verlag (2001). 23.7 Deguchi, H., Terano, T., Kurumatani, K., Yuzawa, T., Hashimoto, S., Matsui, H., Sashima, A., Kaneda, T.: Virtual Economy Simulation and Gaming – An Agent Based Approach; in the Proceedings of the First International Workshop on Agent-based Approach to Economic and Social Complex Systems (AESCS’01), pp.169–174 (2001). 23.8 URL: http://www.v-econ.org 23.9 Iba, T., et al.: Boxed Economy Model: Fundamental Concepts and Perspectives, in the Proceedings of the First International Workshop on Computational Intelligence in Economics and Finance (CIEF’00), (2000). 23.10 Iba, T., et al.: Boxed-Economy Foundation Model: Toward Simulation Platform for Agent-Based Economic Simulations; in the Proceedings of the First International Workshop on Agent-based Approach to Economic and Social Complex Systems (AESCS’01), pp.186–193 (2001). 23.11 URL: http://www.boxed-economy.org 23.12 Kitano, H., Asada, M., Kuniyoshi, Y., Noda, I., Osawa, E., Matsubara, H.: RoboCup – A Challenge Problem for AI; AI Magazine, Vol.18, No.1, pp.73– 85 (1997). 23.13 Kitano, H., Tadokoro, S., Noda, I., Matsubara, H., Takahashi, T., Shinjou, A., Shimada, S.: RoboCup Rescue: Search and Rescue in Large-Scale Disasters as a Domain for Autonomous Agents Research; in the Proceedings of IEEE Conf. on Man, Systems and Cybernetics (1999), http://www.robocup.org/games/Rescue.ps.

24. Virtual Economy Simulation and Gaming —An Agent Based Approach— Hiroshi Deguchi1 , Takao Terano2 , Koichi Kurumatani3 , Taro Yuzawa4 , Shigeji Hashimoto5 , Hiroyuki Matsui1 , Akio Sashima3 , and Toshiyuki Kaneda6 1 2 3 4 5 6

Graduate School of Economics, Kyoto University Graduate School of Systems Management, University of Tsukuba Cyber Assist Research Center, AIST. Foundation for Fusion Of Science & Technology KOEI NET Co.,Ltd. Department of Systems Management and Engineering, Nagoya Institute of Technology

In this paper we analyze an economic systems as agent based bottom up models. For the purpose we introduce a small national economy called a virtual economy and an exchange algebra for state space description. We construct dynamical agent based simulation model and analyze it.

24.1 Introduction In this paper we construct a Simulation & Gaming model of a virtual economy. The virtual Economy consists of nine agents such as Agriculture, Milling Industry, Bread Industry (Bakery), Steel Industry, Machinery Industry, Government, Household, Bank and Central Bank. For the purpose an algebraic abstraction of bookkeeping system, which is called an exchange algebra, is introduced for describing micro economic exchange among agents. An economic state of each agent is also described by the algebra. Exchange algebra is an extension of accounting vector space [24.1, 24.2]. By using this algebras we describe systemic properties of economic exchange and properties of economic field. The economic field gives a formal model of SNA (System of National Account). The virtual economy model is illustrated with Fig.24.1. In the model economy agriculture grows wheat, milling industry makes wheat flour of wheat, bread industry (bakery) makes bread from flour, steel industry makes steel and machinery industry makes machinery from steel. In the model we assume that there are no materials for steel industry. Household purchases and consumes bread. Machines are purchased by industries as capital investments. The machines are used for production. Machines are also purchased by government or household. The machines that are purchased by government or household are considered as infrastructure and houses respectively. A machine depreciates according to a scenario. Population increases by a scenario. Household supplies workers to each industry and a government. T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 218 − 226, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

24. Virtual Economy Simulation and Gaming

219

Fig. 24.1. Virtual Economy

Then household receives a wage. A government can issue national bonds. The central bank issues a bank note and fixes the official bank rate. Household and industries deposit money in a bank. A bank lends money.

24.2 Agent Based Simulation Model for Virtual Economy In the virtual economy gaming players act as government, agriculture milling industry, bakery, steel industry, machinery industry, household, bank and central bank depending on their roles. This virtual economy becomes a multi agent model of an economic system of a country. In the economy players or machine agents act as decision makers. The game needs some basic assumptions. For example we have five products and one currency in this economy. We also assume proper units for the products and currency as follows. “MOU” stands for money unit such as dollar. “WHU” stands for wheat unit, “FLU”

220

H. Deguchi et al.

stands for flour unit, “BRU” stands for bread unit, “STU” stands for steel unit and “MAU” stands for machine unit. They regard a machine as a house in the household. We try to construct an agent based simulation model for this economy. Fig.24.2 shows the total design for the agent based simulation of the virtual economy.

C la s s fo r E x c h a n g e A lg e b r a

M u C o S a tis fa ( e .x . M

C la s s fo r A g e n t A c tiv itie s C la s s fo r S N A

V ir tu a l E c o n o m y S ta te S p a c e K e r n e l

lti A g e n s tr a in c tio n S a r k e t S o lv e r )

n t t o lv e r D M

R u le B a s e d D e c is io n M a k in g M o d e l L e r n in g M o d e l o f R u le s Fig. 24.2. Basic Design of the Agent Based Simulation Model of Virtual Economy

Fig.24.3 shows a prototype decision making model for a single human player. In this paper we introduce two types of dynamical models for this virtual economy gaming. The one is called the dictator’s view model. In this model a player has to make all decisions for transactions among agents of this economy in a term like a dictator. Table 24.1 shows decision making items for a player in a term. The other is called the bird’s eye view model across the terms. In this model some decision is are made automatically depending on hidden decision making rules. A player makes a decision across the time in this model. In the former model decisions are made step by step in terms. But in the latter model a player has a bird’s eye view across the terms. In the model a player can observe total periods of economic development and makes a decision across the terms for achieving his aim in the economy. Table 24.2 shows institutional parameters such as subsidy policy and national bond policy. A decision is made such as true or false. Table 24.3 shows

24. Virtual Economy Simulation and Gaming

Table 24.1. Decision Making for Dictator’s View Model

g

In com e Ta x R a te

A g r i . F l o u . B a k e . * * *

S t e e l *

M a c h . *

H o u s e

Cor p or a te Ta x R a te

*

*

*

*

*

*

N a ti on a l B on d R a te *

*

*

*

*

*

Offi ci a l B a n k R a te *

*

D ep os i ts i n CB Wi th d r a w fr om

*

*

*

CB

*

*

*

G o v . 0.1

*

0.2

*

* *

*

*

B a n k *

*

0 . 0 1

*

*

C B *

0.01

*

*

*

0

0

*

*

*

*

*

*

*

0

0

CB *

*

*

*

*

*

*

1000

1000

R efu n d to CB *

*

*

*

*

*

*

0

0

R ecei ve Su b s i d y

0

0

10

10

10

0

3 0

D ep os i t In ter es t *

*

*

*

*

*

*

L oa n In ter es t *

*

*

*

*

* *

B u y N a ti on a l B on d (N B )

0

0

0

0

0

0

0 *

0

0

0

0

0

0

0 *

*

*

*

*

*

* 0

0

*

0

0

*

L oa n fr om

R ed eem

N B

Accep t N B b y CB R ed eem

N B fr om

CB

*

*

*

*

*

B a n k

0

3 00

100

3 00

3 00

0 *

to B a n k

0

0

0

0

0

0 *

D ep os i t i n B a n k

0

0

0

0

0

0 *

*

0

0

0

0

0

0 *

*

0.2

0.5 1

6.25

11

Ca p i ta l In ves tm en t (n u m b er s ) 2 3

2

4

8

Sa l es of P r od u cts (Q u a n ti ty )

770

58 0

4 20

14

21

N u m b er s of E m p l oy m en t

70

70

60

3 0

6 5

Tota l Wa g e

90

90

8 0

3 0

90

Wi th d r a w fr om

B a n k

P r od u ct P r i ce p er Un i t

*

*

1000

*

0 0 0

*

2

*

0 *

* 3 3 0 4 3 0

*

*

*

*

0.03 *

*

R ed eem

0.01 *

*

L oa n fr om

0

*

* 3 5 50

* *

* *

*

Table 24.2. Institutional Parameters A d o p tio n o f a P o lic y : I n s titu tio n a l P a r a m e te r s 1

2

3

4

5

6

7

8

9

10

Su b s i d y for Ha l f th e Ca p i ta l In ves tm en t Su b s i d y for th e D efi ci t of Ma k er s

F AL S

TR UE

TR UE

TR UE

TR UE

TR UE

F AL SE

F AL SE

F AL SE

F AL SE

F AL SE

TR UE

TR UE

TR UE

TR UE

TR UE

TR UE

TR UE

TR UE

TR UE

Su b s i d y for Ha l f th e Hou s e In ves tm en t

F AL SE

F AL SE

F AL SE

F AL SE

TR UE

TR UE

TR UE

TR UE

TR UE

TR UE

Is s u e N a ti on a l B on d u n d er g u a r a n tee of Cen tr a l B a n k

F AL SE

F AL SE

F AL SE

F AL SE

F AL SE

F AL SE

TR UE

TR UE

TR UE

TR UE

221

222

H. Deguchi et al.

C la s s fo r E x c h a n g e A lg e b r a C la s s fo r A g e n t A c tiv itie s C la s s fo r S N A

V ir tu a l E c o n o m y S ta te S p a c e K e r n e l P r o c e d u M a k in S p r e T y p e

r a l D e c is io n g P r o to ty p e a d sh e e t In te r fa c e

Fig. 24.3. Prototype Decision Making Model for a Single Human Player

Table 24.3. Capital Investment Ter m 1 2 3 4

Ma ch i n e Ag r i a va i l a b l e .

F l ou r

26

4

2

B a k el y

Stee l

1

7

Ma ch i n e Hou s e h ol d

G ov. Su m

Stock of Ma ch i n e

11

0

0

1

26

3 5

2

5

4

10

14

0

0

3 5

0

4 4

2

5

6

10

19

2

0

4 4

0

55

2

10

10

23

1

0

54

1

5

62

3

15

10

10

22

2

0

62

0

6

75

2

15

17

12

25

3

1

75

0

8 1

4

17

15

10

28

4

3

8 1

0

8 1

5

21

13

16

24

2

0

8 1

0

22

13

10

3 3

2

8 8

0

3 0

15

13

3 2

0

96

1

7 8 9

8 8

10

97

5 3

8

3 3

24. Virtual Economy Simulation and Gaming

223

Table 24.4. Management and Political Decisions M a n a g e m e n t Ter m

1

2

In com e Ta x R a te

0.1

0.05

&

P o l i t i c a l 3

4

D e c i s i o n s

o f

A g e n t s

5

6

7

9

10

0.05

0.05

0.02

0.02

0.05

8 0.05

0.05

0.05 0.1

Cor p or a te Ta x R a te

0.2

0.1

0.2

0.2

0.1

0.1

0.1

0.1

0.1

N a ti on a l B ob d R a te

0.01

0.01

0.01

0.01

0.01

0.01

0.01

0.01

0.01

0.01

Offi ci a l B a n k R a te

0.01

0.01

0.01

0.01

0.01

0.01

0.01

0.01

0.01

0.01

D ep os i t In ter es t

0.01

0.01

0.01

0.01

0.01

0.01

0.01

0.01

0.01

0.01

L oa n In ter es t

0.03

0.03

0.03

0.03

0.03

0.03

0.03

0.03

0.03

0.03

P r i ce Wh ea t

0.2

0.2

0.2

0.2

0.2

0.2

0.2

0.2

0.2

0.2

P r i ce F l ou r

0.5

0.5

0.5

0.52

0.52

0.52

0.52

0.52

0.52

0.52

P r i ce B r ea d

1

1

1

1

1

1

1.1

1.1

1.1

1.1

P r i ce Steel

6.25

6.25

6.25

6.5

6.5

6.5

6.5

6.5

6.5

6.5

P r i ce Ma ch i n e

10

10

10

10

10

10

10

10

10

10

Wa g e(Ag r .)/P er s on

1.2

1.2

1.2

1.3

1.3

1.3

1.3

1.3

1.3

1.3

Wa g e(F l ou .)/P er s on

1.3

1.3

1.3

1.3

1.3

1.3

1.3

1.3

1.3

1.3

Wa g e(B a k e.)/P er s on

1.3

1.3

1.3

1.4

1.4

1.4

1.4

1.4

1.4

1.4

Wa g e(Steel )/P er s on

1

1.1

1.2

1.3

1.5

1.5

1.5

1.5

1.5

1.5

Wa g e(Ma ch .)/P er s on

1.3

1.3

1.4

1.6

1.6

1.6

1.6

1.6

1.6

1.6

Wa g e(G ov.)/P er s on

1.4

1.4

1.4

1.4

1.5

1.5

1.5

1.5

1.5

1.5

capital investment of each agent in each term. Table 24.4 shows management and political parameters. A player can observe several types of economic developments across the terms while he change these parameters. A player can set up different goals depending on the social indexes on which he is focussing such as the numbers of residents per house, GDP per person and foods consumption per person.

24.3 Result of Simulation We show the results of economic developments in ten terms. The following figures show the results of simulation by bird’s eye view model across the terms under the parameters of the previous tables. Fig.24.4 shows the numbers of residents per house. Fig.24.5 shows GDP per person and food consumption per person. Fig.24.6 shows price index. Fig.24.7 shows cash in the government and issued national bonds which are accepted in the central bank.

224

H. Deguchi et al.

3 5 . 3 0 . 2 5 . 2 0 . 1 5 . 1 0 . 5 . 0 .

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1

2

3

4

5

6

7

8

9

1 0

Fig. 24.4. The Numbers of Residents per House

3 . 0 0 2 . 5 0

M O U

2 . 0 0 1 . 5 0 1 . 0 0 0 . 5 0 0 . 0 0 1

2

3

4

5

6

7

8

9

1 0

T e r m G D P

( R e a l ) / P e r s o n

F o o d

C o n s u m p t i o n ( B R U ) / P e r s o n

Fig. 24.5. GDP (Real) per Person & Food Consumption (BRU) per Person

1 . 0 1 . 0 1 . 0 1 . 0 1 . 0 1 . 0 0 . 9 0 . 9 0 . 9

5 4 3 2 1 0 9 8 7 1

2

3

Fig. 24.6. Price Index

4

5

6

7

8

9

1 0

24. Virtual Economy Simulation and Gaming

225

Fig. 24.7. Cash in Government and Issued National Bond Accepted in Central Bank 1 0 0 9 0 8 0 7 0 6 0 5 0 4 0 3 0 2 0 1 0 0 1

2

3 P r o d u c t

4 S t o c k

o f

5 S t e e l

( S T U )

6

7 P r o d u c t

8 S t o c k

o f

9 M a c h i n e

1 0 ( M A U )

Fig. 24.8. Product Stock of Steel and Machine

24.4 Conclusion We investigated an agent based simulation model of a small national economy. The model is different from usual macro economic model. We assumed bottom up state description of an economic agent by exchange algebra. We can add multi agent decision making mechanism in the model from bottom up point of view as is shown in Fig.24.2. We want to express and design the institutional and structural varieties of real economy in agent based models. This is a first step for our research program.

226

H. Deguchi et al.

References 24.1 H.Deguchi and B.Nakano (1986): Axiomatic Foundations of Vector Accounting, Systems Research, Vol.3, No.1, pp.31–39 24.2 Hiroshi Deguchi(1996): Multi Agent Economics and Its Gsming Simulation. Modeling and Control of National and Regional Economies 1995. edited by Vlacic et al.,Pergamon Press, pp.269–274, 24.3 Hiroshi Deguchi(1998): Agent Based Approach for Social Complex Systems – Management of Constructed Social World, Community Computing and Support Systems, edited by Toru Ishida, LNCS 1519, Springer, pp.62–77

25. Boxed Economy Foundation Model: Model Framework for Agent-Based Economic Simulations Takashi Iba1,2,3 , Yohei Takabe4 , Yoshihide Chubachi4 , Junichiro Tanaka1 , Kenichi Kamihashi1 , Ryunosuke Tsuya1 , Satomi Kitano1 , Masaharu Hirokane1 , and Yoshiaki Matsuzawa1 1 2 3 4

Keio University, 5322 Endo, Fujisawa, Kanagawa, Japan JSPS Research Fellow Research Associate of Fujita Institute of Future Management Research, Hibiya bldg. 1-1, 1-chome, Shinbashi, Minato, Tokyo, Japan Keio Research Institute at SFC, 5322 Endo, Fujisawa, Kanagawa, Japan

25.1 Introduction The recent advancement of the agent-based modeling and simulation has been revolutionizing the social sciences and other research fields. The agentbased approach enables us to deal with the model that generates macroscopic phenomena by allowing numbers of agents to act at the micro level within the simulation. Therefore, in the social sciences, we can trace and understand the internal mechanisms in society. Since some interesting implications have been derived from the former researches with agent-based approach, expectations are rising in social sciences. In the last some years, several tools for agent-based simulations have been proposed: Swarm Simulation System[25.10], Ascape[25.11], RePast[25.12], MAML[25.7] and so on. Especially Swarm Simulation System has become one of the most famous and most growth toolkit in many research fields. Although these tools have promoted to share some kind of components such as Graphical User Interfaces (GUI) among researchers, they have been less successful in sharing and cumulating the parts of simulation models. It is because the provided basis are too high abstract for users to follow when they build the sharable components. To design the sharable and reusable models, the domain-specific design is required at the level of social model rather than the level of abstract general-purpose model. To put it another way, the social scientists really need not only the abstract basis, such as mathematical operators, but also the model components, such as production function or consumption function in economics. Indeed, economists usually specify their model with using the typical model components. There are, for example, some types of production functions: Cobb-Douglas type, CES type,

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 227− 23 6, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

228

T. Iba et al.

and Translog type in economics. They hardly ever make the model components from scratch each time1 . We, then, would like to provide the model framework specializing in agent-based model of economic society, incorporating the idea of objectoriented framework that define the basic architecture of economic and social models[25.8]2 . We call our model framework “Boxed Economy Foundation Model”.

25.2 Model Framework for Agent-Based Economic Simulations Boxed Economy Foundation Model provides the framework for modeling the economic society. The foundation model is an abstract model of a real society from the viewpoint of economy. We would like to suggest especially that the design with object-oriented framework is more significant than the design simply with components or objects in the field of the economic and social simulations. This is because the introduction of the frameworks makes it easier for the simulation builders to build, share and co-improve the economic simulations. Framework is the architecture that is specialized to a certain domain. Framework provides many kinds of plug-points (container) to connect the components that would be implemented by the simulation builders in each simulation(Fig. 25.1). Frameworks is important for reusing and co-improving due to define a “context” for the components developed in the future, although it is usually difficult to combine the components developed by independent groups, because they have inconsistent assumptions each other. To build a realistic model step by step, it is necessary to urge the researchers and some businessperson from other areas to participate in the development. Boxed Economy introduces the idea of framework to simulate the economic society and keep the architecture on one track. Therefore, the simulation builders can make the models in parallel as long as they keep the same framework, and they can concentrate on the object related to their major: consumer, corporation, and so on.

25.3 Boxed Economy Foundation Model Boxed Economy Foundation Model has the definition of the fundamental relationship between each part of artificial economy model. Fig. 25.2 shows 1 2

Most of the simulation models are built from scratch each time in agent-based research. Enormous developing time and costs are required in this style. C.Bruun[25.4] has similar motivation and also try to make model framework for agent-based economic simulations. There is, however, critical differences in regard to the agent design, as we will mention later.

25. Boxed Economy Foundation Model

229

Fig. 25.1. ObjectOriented Framework and Components

the classes and their relationships in the Boxed Economy Foundation Model which is expressed in Unified Modeling Language (UML)[25.2]. The main part of the foundation model currently contains 14 classes and they are called “foundation model class”. The classification is as follows3 : EconomicActor, SocialGroup, Individual Goods, Information, Possession Behavior, BehaviorManagement, Memory, Needs Relation, Path Clock, Location

– – – – –

An “agent” can be a representative of any autonomous subjects in the economy. It means that each individuals and social groups such as government or corporations are all dealt as “agent” in the model. The “agent” which is defined in the Boxed Economy, is formed by the following classes: [EconomicActor] as its core, [Behavior], [BehaviorManagement] and [Memory]. [EconomicActor] reacts with these classes that surround it and becomes an agent in artificial economy. In the rest of this section, we would like to introduce the definition of some classes, their correspondence to the real society and the relationship with other classes in the model by catalog style. 25.3.1 EconomicActor, SocialGroup, Individual [EconomicActor] Definition: Correspondence: 3

An actor who carries economic activities in the artificial society. Human or social group as consumer, corporation, bank, government, etc.

The design and definitions of Boxed Economy Foundation Model are a temporary statement and they might be changed in the future.

230

T. Iba et al.

Fig. 25.2. Main Architecture of Boxed Economy Foundation Model

Explanation:

Related Class:

[EconomicActor] is the core element that executes the economic activities. [Behavior], [BehaviorManagement] and [Memory] are added in order to create an “agent”. [EconomicActor] stands for the [Individual] and the [SocialGroup]. [EconomicActor] owns one [Memory], [BehaviorManagement] and more than one [Behavior]. Also it owns many [Goods] and exchange them through [Path] which will be created based on the [Relation] they have.

[SocialGroup] Definition: Correspondence: Explanation:

Related Class:

A group which is formed by the [EconomicActor]. Social group as corporation, regional community, etc. [SocialGroup] is one kind of [EconomicActor]. [SocialGroup] consists of [EconomicActor] or other [SocialGroup]. Note that it is possible to have [SocialGroup] inside another group. This class is extended from the [EconomicActor] class and inherits all the characteristics, it holds [Memory], [Behavior], [BehaviorManagement], [Goods], [Relation] and [Path].

25. Boxed Economy Foundation Model

231

[Individual] Definition: Correspondence: Explanation:

Related Class:

A single human being in the artificial society. Human being. [Individual] is one kind of [EconomicActor]. The differences between [Individual] and [SocialGroup] is that [Individual] may have the [Needs]. [Individual] is the minimum unit to form [SocialGroup]. This class is extended from the [EconomicActor] class and inherit its characteristics, then contains [Memory], [Behavior], [BehaviorManagement], [Goods], [Relation], and [Path].

25.3.2 Goods, Information, Possession [Goods] Definition: Correspondence: Explanation:

Related Class:

Everything that is owned or exchanged by [EconomicActor]. Also can be something that is invisible. Commodities, service, money, etc. [Goods] has the following attributes, name, kind, visibility, date of produce, basic endurance, portability, divisibility, amount, unit of measurement, etc. Career of information and also money as well are treated as a kind of [Goods]. [Goods] is named as [Possession] when it is owned by [EconomicActor]. [Information] is always exchanged with some kind of [Goods] as a carrier not by itself.

[Information] Definition: Correspondence: Explanation:

Related Class:

Knowledge which is an expression of many facts. Knowledge stored in documents, the contents of communications and advertisement, etc. [Information] does not stand by itself, but is always a thing which is contained by [Goods]. For example when papers contain the [Information], it will be a document, and when voice becomes the carrier it will be a verbal communication. When information reaches the [EconomicActor], it will be decoded into [Memory]. [Information] is always exchanged with some kind of [Goods] as a carrier not by itself.

232

T. Iba et al.

25.3.3 Behavior, BehaviorManagement, Memory, Needs [Behavior] Definition: Correspondence:

Explanation:

Related Class:

An element to construct the decision and action of the economic actor. The corporate behaviors of strategic decision-making, production, sales, etc. And the individual behavior of purchase decision-making, information processing, etc. Each of decision-making and behavior is defined as the behavior. Each [EconomicActor] is able to execute the decision-making and behavior which is defined by [Behavior] it has. It is held in [EconomicActor].

[Memory] Definition: Correspondence: Explanation:

Related Class:

Knowledge that is stored in the economic actor. Things that somebody knows, etc. [Memory] would be referred to when the agent has to make a decision. By time to time, memory would be refreshed by its experience. It is stored in [EconomicActor].

[Needs] Definition: Correspondence: Explanation:

Related Class:

A drive that motivates individual to an action. Desire of human. [Needs] is a thing that [Individual] holds as a mechanism of action, but a [SocialGroup] does not have this. The state of lack drives the [individual] to some kind of action and the desire would be fulfilled. It is held by [individual].

25.3.4 Relation, Path [Relation] Definition: Correspondence: Explanation:

Related Class:

A state that [EconomicActor] knows some other [EconomicActor]. The relationship of family, friends, labor, neighborhood, etc. Having [Relation] is a state that the communication is enabled. By the [Information] which the agent gains, there would be a new [Relation] constructed. [Relation] would be normally expressed as a one-way but when both of them connects each other it will be two-way. It is held by an [EconomicActor].

25. Boxed Economy Foundation Model

233

[Path] Definition: Correspondence: Explanation:

Related Class:

A path created with its relation to communicate with other economic actor. A path to exchange items or to communicate with others. Items or contents of verbal communication we would be exchanged through out this path. For example, retailer will open a path to the customer to give the item to him/her. [EconomicActor] will create a path by its [Relation] and the [Path] enables to pass the [Goods] to one another.

25.4 Applying Boxed Economy Foundation Model 25.4.1 Modeling Behavior Rather than Agent When you want to create a simulation based on Boxed Economy, you will be describing the details of the agents by using the class definition, which you have just read through. We would like to emphasize that it is important to characterize the agent as an object that has more than one behavior. This representation of the agent is epoch-making and has more advantage than the conventional models which also handle the agent as a minimum indivisible unit in a simulation[25.3][25.1][25.6]. The advantage is that in this way it will be possible to describe an agent to act more than one social role. For example, most of the individuals would act as “consumers” if they buy some items from the store, and would act as “labors” if they work to earn money. The point is that we do not have the subject called consumers or labors, but the subject we have in our society is only individual persons which act the role of consumer or labor in each scenes. In the Boxed Economy, we follow this idea and create the agent as an individual person that has the behavior of consumer, and we do not create a consumer agent. As a summary, to create the model of economic actor by using the Boxed Economy Foundation Model would be the modeling the behaviors that the economic actor has. 25.4.2 Flexibility on the Boundary of Agent Boxed Economy provides the ability to the agents to be dynamic inside it. In other words, the agent based on the foundation model will be able to decide its own boundary. There are three ways of changing boundary. The first way of changing the boundary is to increase/decrease the number of actors inside it. Corporation agent, for example, will be able to change the number of workers by hiring and firing.

234

T. Iba et al.

Fig. 25.3. Representation of Wholesaler or Retailer

The second way of changing the boundary would be done by exchanging, increasing or decreasing the behavior that the agent has. Since the agent in the model is defined as an object that has the behaviors to make decisions or doing some kinds of actions, the functional boundary of the agent can be changed by adding/deleting its behaviors. For instance, if you want to let the seller agent to obtain the part of banking functions, it will be realized by adding the behavior of banking function to the seller agent. The third way of changing the boundary is to generate a new agent (can be individual or social group) or to disappear the existing agent. It may be birth or death for individuals, marriage or divorce will apply to families, foundation or bankruptcy for corporations. By providing the agents with the ability mentioned above, the agents in the simulation will be able to change and adjust themselves to the situation as time goes by. Since the analysis with artificial economy is often focused to observe the long-term movements in the whole economy, we need to implement this behavior to the agent. 25.4.3 Example: Sellers in Distribution Mechanism The mechanism of distribution include corporations which stand for producer, wholesaler and retailer in its structure. Both wholesaler and retailer mostly has the same behavior, but retailer only sells its items to the consumer and wholesaler is a reseller of products to anyone except the consumers. In the Boxed Economy we do not model the agents as wholesalers or retailers, instead we define the agents by dividing their decision-making and action by its behaviors (Fig. 25.3). In this way, it will be possible for many subjects to have the same behavior, and will provide expandability to the agents. In the fundamental model, we can create the social group within the social group: for instance, if you imagine departments in a corporation, both departments and corporation would be a [SocialGroup] (Fig. 25.4). And by using this idea we will be able to out-source some of the functions to others or we can also create a transportation business that only has the function of transportation. In the real world, there are movements of out-sourcing the

25. Boxed Economy Foundation Model

235

Fig. 25.4. Representation of Wholesaler or Retailer (in the Case of Department Structure)

function, or merge the whole structure of the corporation, and we will be ready to simulate such situations by modeling them with its behavior4 .

25.5 Conclusion In this paper, we proposed the concept and design of “Boxed Economy Foundation Model”, which is a sharable model framework for agent-based economic simulations. Here we should note that we have developed “Boxed Economy Simulation Platform”, which realizes the simulation environment for the simulation model based on Boxed Economy Foundation Model[25.9]. The platform is implemented by Java, which is portable and independent of the computer platform, and will be opened to public before soon (Fig. 25.5). Creating the foundation for the social simulation researches is an oversized project for our members to complete. We would like to realize this by collaborating with many researchers in various fields. Please contact us on http://www.boxed-economy.org/, if you are interested in our challenge. Acknowledgment. This research was partly supported by a grant from the Ministry of Education, Science, Sports and Culture, Grant-in-Aid for Encouragement of Young Scientists, 1999 and 2000. It was also supported by Fujita Institute of Future Management Research, Japan, since 1997. Thank you also for other members of Boxed Economy Project: Asaka, K., Kaiho, K., Dr.Takenaka, H., Dr.Takefuji, Y. and Dr.Oiwa, H. 4

The design that separates behaviors from the class is of great advantage not only to build flexible social models but also to build flexible software. To delegate the role to other objects, which is called “composition”, is more flexible than inheritance, and is known as close way to the essence of object-oriented design[25.5].

236

T. Iba et al.

Fig. 25.5. Boxed Economy Simulation Platform

References 25.1 Basu, N., Pryor, R.J., Quint, T. (1998): ASPEN: A Microsimulation Model of the Economy. Computational Economics. 12, 223–241 25.2 Booch, G., Rumbaugh, J., Jacobson, I. (1999): The Unified Modeling Language User Guide. Addison-Wesley 25.3 Bruun, C. (1997): Agent-Based Keynesian Economics. Simulating Social Phenomena. Springer-Verlag, 279–285 25.4 Bruun, C. (2000): Prospect for an Economics Framework for Swarm. http:// www.socsci.auc.dk/˜cbruun/. 25.5 Coad, P., Mayfield, M. (1999): Java Design: Building Better Apps & Applets. 2nd edition. Yourdon Press, Prentice Hall PTR. 25.6 Deguchi, H. (1998): Agent Based Approach for Social Complex Systems: Management of Constructed Social World. Community computing and support systems: social interaction in network communities. 25.7 Gulyas, L., Kozsik, T., Corliss, J.B. (1999): The Multi-Agent Modelling Language and the Model Design Interface. Journal of Artificial Societies and Social Simulation, 2(3), http://www.soc.surrey.ac.uk/JASSS/2/3/8.html 25.8 Iba, T., Hirokane, M., Takabe, Y., Takenaka, H., Takefuji, Y. (2000): Boxed Economy Model: Fundamental Concepts and Perspectives. Proceedings of Computational Intelligence in Economics and Finance. 25.9 Iba, T., Takabe Y., Chubachi, Y., Takefuji, Y. (2001): Boxed Economy Simulation Platform and Foundation Model. Workshop of Emergent Complexity of Artificial Markets, 4th International Conference on Computational Intelligence and Multimedia Applications. 25.10 Minar, N., Burkhart, R., Langton, C., Askenazi, M. (1996): The Swarm Simulation System:A Toolkit for Building Multi-agent Simulations. http://www. santafe.edu/projects/ swarm/overview/overview.html 25.11 Parker, M.T. (2001): What is Ascape and Why Should You Care?, Journal of Artificial Societies and Social Simulation. 4(1), http://www.soc.surrey. ac.uk/JASSS/ 4/1/5.html 25.12 The University of Chicago’s Social Science Research: RePast. http://repast. sourceforage.net/

26. Workshop on Rough Set Theory and Granular Computing – Summary Shusaku Tsumoto1 , Shoji Hirano1 , and Masahiro Inuiguchi2 1 2

Shimane Medical University, Izumo, 693-8501 Japan {tsumoto,hirano}@shimane-med.ac.jp Graduate School of Engineering, Osaka University, Suita, Osaka 565-0871, Japan [email protected]

Rough sets was proposed by Z. Pawlak in 1980 as the way how real-world concepts can be approximated by human measurements. For example, in a database, real-world concepts were approximated by the combination of attributes, as lower and upper approximation. The formal studies on this approximation can be viewed as the computation of information granularity, which are closely related with data mining, machine learning, multi-valued logic and fuzzy sets. The workshop on rough sets and granular computing started from May 20 due to the number of paper submissions (30). The workshop consisted of three invited talks by Z. Pawlak, A. Skowron and S.K.Pal and 30 presentations of regular papers (3: inductive logic programming (ILP), 3: decision making, 5: rule induction, 3: fuzzy logic, 3: granular computing, 5: fundamentals of rough sets, 6: applications, 2: conflict analysis). The number of attendees in this workshop was 42 in total(22:Japan, 5: Poland, 3: India, 2: US, 1: Korea, 6: PhD students of Shimane University and Shimane Medical University). In the invited talks, Pawlak discussed the relations between rough sets and Bayesian inference and the Lukasiewicz multi-valued logic as a key notion of the bridge between rough sets and Bayesian reasoning. In the second talk, Skowron reviewed the studies on rough sets which plays important roles in the estimation of information granularity and discussed the potentials of granular computing in multi-agent systems. In the final talk, Pal discussed the importance of rough sets, fuzzy sets and genetic algorithms in data mining. In the regular sessions, not only the applications of rough sets but also several fundamental studies on the extensions of rough sets were presented. Also, theoretical studies on the combinations of rough sets and other methods, such as inductive logic programming and fuzzy reasoning were shown. These invited talks and regular papers showed that rough sets are widely used as an important tool for data mining and data analysis and that rough sets should be recognised as a fundamental tool for the theoretical studies on approximate reasoning.

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p . 23 9, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

27. Bayes’ Theorem Revised – The Rough Set View Zdzislaw Pawlak Institute of Theoretical and Applied Informatics, Polish Academy of Sciences, ul. Baltycka 5, 44 100 Gliwice, Poland

Rough set theory offers new insight into Bayes’ theorem. The look on Bayes’ theorem offered by rough set theory is completely different from that used in the Bayesian data analysis philosophy. It does not refer either to prior or posterior probabilities, inherently associated with Bayesian reasoning, but it reveals some probabilistic structure of the data being analyzed. It states that any data set (decision table) satisfies total probability theorem and Bayes’ theorem. This property can be used directly to draw conclusions from data without referring to prior knowledge and its revision if new evidence is available. Thus in the presented approach the only source of knowledge is the data and there is no need to assume that there is any prior knowledge besides the data. We simply look what the data are telling us. Consequently we do not refer to any prior knowledge which is updated after receiving some data.

27.1 Introduction This paper is an abbreviation of [27.8] Bayes’ theorem is the essence of statistical inference. ”The result of the Bayesian data analysis process is the posterior distribution that represents a revision of the prior distribution on the light of the evidence provided by the data” [27.5]. ”Opinion as to the values of Bayes’ theorem as a basic for statistical inference has swung between acceptance and rejection since its publication on 1763” [27.4]. Rough set theory offers new insight into Bayes’ theorem. The look on Bayes’ theorem offered by rough set theory is completely different to that used in the Bayesian data analysis philosophy. It does not refer either to prior or posterior probabilities, inherently associated with Bayesian reasoning, but it reveals some probabilistic structure of the data being analyzed. It states that any data set (decision table) satisfies total probability theorem and Bayes’ theorem. This property can be used directly to draw conclusions from data without referring to prior knowledge and its revision if new evidence is available. Thus in the presented approach the only source of knowledge is the data and there is no need to assume that there is any prior knowledge besides the data. We simply look what the data are telling us. Consequently T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 24 0− 250, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

27. Bayes’ Theorem Revised – The Rough Set View

241

we do not refer to any prior knowledge which is updated after receiving some data. Moreover, the rough set approach to Bayes’ theorem shows close relationship between logic of implications and probability, which was first observed by L  ukasiewicz [27.6] and also independly studied by Adams [27.1] and others. Bayes’ theorem in this context can be used to ”invert” implications, i.e. to give reasons for decisions. This is a very important feature of utmost importance to data mining and decision analysis, for it extends the class of problem which can be considered in these domains. Besides, we propose a new form of Bayes’ theorem where basic role plays strength of decision rules (implications) derived from the data. The strength of decision rules is computed from the data or it can be also an subjective assessment. This formulation gives new look on Bayesian method of inference and also essentially simplifies computations.

27.2 Bayes’ Theorem In this section we recall basic ideas of Bayesian inference philosophy, after recent books on Bayes’ theory citeber:smi,box:tia,bert:han. In his paper [27.2] Bayes considered the following problem: ”Given the number of times in which an unknown event has happened and failed: required the chance that the probability of its happening in a single trial lies somewhere between any two degrees of probability that can be named.” ”The technical results at the heart of the essay is what we now know as Bayes’ theorem. However, from a purely formal perspective there is no obvious reason why this essentially trivial probability result should continue to excite interest” [27.3]. ”In its simplest form, if H denotes an hypothesis and D denotes data, the theorem says that P (H|D) = P (D|H) × P (H) /P (D) . With P (H) regarded as a probabilistic statement of belief about H before obtaining data D, the left-hand side P (H|D) becomes an probabilistic statement of belief about H after obtaining D. Having specified P (D|H) and P (D), the mechanism of the theorem provides a solution to the problem of how to learn from data. In this expression, P (H), which tells us what is known about H without knowing of the data, is called the prior distribution of H, or the distribution of H a priori. Correspondingly, P (H|D), which tells us what is known about H given knowledge of the data, is called the posterior distribution of H given D, or the distribution of H a posteriori” [27.3]. ”A prior distribution, which is supposed to represent what is known about unknown parameters before the data is available, plays an important role in

242

Z. Pawlak

Baysian analysis. Such a distribution can be used to represent prior knowledge or relative ignorance” [27.4]. Let us illustrate the above by a simple example taken from [27.5]. Example 27.2.1. ”Consider a physician’s diagnostic test for presence or absence of some rare disease D, that only occurs in 0.1% of the population, i.e., P (D) = .001. It follows that P (D) = .999, where D indicates that a person does not have the disease. The probability of an event before the evaluation of evidence through Bayes’ rule is often called the prior probability. The prior probability that someone picked at random from the population has the disease is therefore P (D) = .001. Furthermore we denote a positive test result by T + , and a negative test result by T − . The performance of the test is summarized in Table 1. Table 27.1. Performance of diagnostic test T+

T−

D

0.95

0.05

D

0.02

0.98

What is the probability that a patient has the disease, if the test result is positive? First, notice that D, D is a partition of the outcome space. We apply Bayes’ rule to obtain   P D|T + =

P (T + |D) P (D)    = P (T + |D) P (D) + P T + |D P D .95 · .001 = = .045. .95 · .001 + .02 · .999

Only 4.5% of the people with a positive test result actually have the disease. On the other hand, the posterior probability (i.e. the probability after evaluation of evidence) is 45 times as high as the prior probability”.  

27.3 Information Systems and Approximation of Sets In this section we define basic concepts of rough set theory: information system and approximation of sets. Rudiments of rough set theory can be found in [27.7, 27.10]. An information system is a data table, whose columns are labeled by attributes, rows are labeled by objects of interest and entries of the table are attribute values.

27. Bayes’ Theorem Revised – The Rough Set View

243

Formally, by an information system we will understand a pair S = (U, A), where U and A, are finite, nonempty sets called the universe, and the set of attributes, respectively. With every attribute a ∈ A we associate a set Va , of its values, called the domain of a. Any subset B of A determines a binary relation I(B) on U , which will be called an indiscernibility relation, and defined as follows: (x, y) ∈ I(B) if and only if a(x) = a(y) for every a ∈ A, where a(x) denotes the value of attribute a for element x. Obviously I(B) is an equivalence relation. The family of all equivalence classes of I(B), i.e., a partition determined by B, will be denoted by U/I(B), or simply by U/B; an equivalence class of I(B), i.e., block of the partition U/B, containing x will be denoted by B(x). If (x, y) belongs to I(B) we will say that x and y are B-indiscernible (indiscernible with respect to B). Equivalence classes of the relation I(B) (or blocks of the partition U/B) are referred to as B-elementary sets or Bgranules. If we distinguish in an information system two disjoint classes of attributes, called condition and decision attributes, respectively, then the system will be called a decision table and will be denoted by S = (U, C, D), where C and D are disjoint sets of condition and decision attributes, respectively. Thus the decision table determines decisions which must be taken, when some conditions are satisfied. In other words each row of the decision table specifies a decision rule which determines decisions in terms of conditions. Observe, that elements of the universe are in the case of decision tables simply labels of decision rules. Suppose we are given an information system S = (U, A), X ⊆ U , and B ⊆ A. Our task is to describe the set X in terms of attribute values from B. To this end we define two operations assigning to every X ⊆ U two sets B∗ (X) and B ∗ (X) called the B-lower and the B-upper approximation of X, respectively, and defined as follows:

{B (x) : B (x) ⊆ X}, B∗ (X) = x∈U

B ∗ (X) =

{B (x) : B (x) ∩ X = ∅}.

x∈U

Hence, the B-lower approximation of a set is the union of all B-granules that are included in the set, whereas the B-upper approximation of a set is the union of all B-granules that have a nonempty intersection with the set. The set BNB (X) = B ∗ (X) − B∗ (X) will be referred to as the B-boundary region of X. If the boundary region of X is the empty set, i.e., BNB (X) = ∅, then X is crisp (exact) with respect to B; in the opposite case, i.e., if BNB (X) = ∅, X is referred to as rough (inexact) with respect to B.

244

Z. Pawlak

27.4 Rough Membership Rough sets can be also defined employing instead of approximations rough membership function [27.9], which is defined as follows: μB X : U → [0, 1] and μB X (x) =

|B (x) ∩ X| , |B (x) |

where X ⊆ U and B ⊆ A. The function measures the degree that x belongs to X in view of information about x expressed by the set of attributes B. The rough membership function, can be used to define approximations and the boundary region of a set, as shown below: B∗ (X) = {x ∈ U : μB X (x) = 1}, B ∗ (X) = {x ∈ U : μB X (x) > 0}, BNB (X) = {x ∈ U : 0 < μB X (x) < 1}.

27.5 Information Systems and Decision Rules Every decision table describes decisions (actions, results etc.) determined, when some conditions are satisfied. In other words each row of the decision table specifies a decision rule which determines decisions in terms of conditions. In what follows we will describe decision rules more exactly. Let S = (U, C, D) be a decision table. Every x ∈ U determines a sequence c1 (x), . . . , cn (x), d1 (x), . . . , dm (x) where {c1 , . . . , cn } = C and {d1 , . . . , dm } = D. The sequence will be called a decision rule (induced by x) in S and denoted by c1 (x), . . . , cn (x) → d1 (x), . . . , dm (x) or in short C →x D. Decision rules are often presented as logical implications in the form ”if...then...”. A set of decision rules corresponding to a decision table will be called a decision algorithm. The number suppx (C, D) = |C(x) ∩ D(x)| will be called a support of the decision rule C →x D and the number σx (C, D) =

suppx (C, D) , |U |

will be referred to as the strength of the decision rule C →x D, where |X| denotes the cardinality of X. With every decision rule C →x D we associate

27. Bayes’ Theorem Revised – The Rough Set View

245

the certainty factor of the decision rule, denoted cerx (C, D) and defined as follows: suppx (C, D) |C (x) ∩ D (x) | cerx (C, D) = = = |C (x) | |C (x) | σx (C, D) = , π (C (x)) where π (C (x)) = |C(x)| |U | . The certainty factor may be interpreted as a conditional probability that y belongs to D (x) given y belongs to C (x), symbolically πx (D|C) . If cerx (C, D) = 1, then C →x D will be called a certain decision rule in S; if 0 < cerx (C, D) < 1 the decision rule will be referred to as an uncertain decision rule in S. Besides, we will also use a coverage factor of the decision rule, denoted covx (C, D) defined as |C (x) ∩ D (x) | suppx (C, D) = = |D (x) | |D (x) | σx (C, D) , = π (D (x))

covx (C, D) =

where π (D (x)) = Similarly

|D(x)| |U | .

covx (C, D) = πx (C|D) . If C →x D is a decision rule then D →x C will be called an inverse decision rule. The inverse decision rules can be used to give explanations (reasons) for decisions. Let us observe that D cerx (C, D) = μC D(x) (x) and covx (C, D) = μC(x) (x) .

That means that the certainty factor expresses the degree of membership of x to the decision class D (x), given C, whereas the coverage factor expresses the degree of membership of x to condition class C (x), given D.

27.6 Probabilistic Properties of Decision Tables Decision tables have important probabilistic properties which are discussed next. Let C →x D be a decision rule in S and let Γ = C (x) and let Δ = D (x) . Then the following properties are valid:  cery (C, D) = 1 (27.1) y∈Γ

246

Z. Pawlak



covy (C, D) = 1

(27.2)

y∈Δ

π (D (x)) =



cery (C, D) · π (C (y)) =

(27.3)

y∈Γ

=



σy (C, D)

y∈Γ

π (C (x)) =



covy (C, D) · π (D (y)) =

(27.4)

y∈Δ

=



σy (C, D)

y∈Δ

covx (C, D) · π (D (x)) = cerx (C, D) =  covy (C, D) · π (D (y))

(27.5)

y∈Δ

=

σx (C, D) π (C (x))

cerx (C, D) · π (C (x)) = covx (C, D) =  cery (C, D) · π (C (y))

(27.6)

y∈Γ

=

σx (C, D) π (D (x))

That is, any decision table, satisfies (1),...,(6). Observe that (3) and (4) refer to the well known total probability theorem, whereas (5) and (6) refer to Bayes’ theorem. Thus in order to compute the certainty and coverage factors of decision rules according to formulas (5) and (6) it is enough to know the strength (support) of all decision rules only. The strength of decision rules can be computed from data or can be a subjective assessment. Let us observe that the above properties are valid also for syntactic decision rules, i.e., any decision algorythm satisfies (1),...,(6). Thus, in what follows, we will use the concept of the decision table and the decision algorithm equivalently.

27.7 Decision Tables and Flow Graphs With every decision table we associate a flow graph, i.e., a directed acyclic graph defined as follows: to every decision rule C →x D we assign a directed branch x connecting the input node C (x) and the output node D (x) . Strength

27. Bayes’ Theorem Revised – The Rough Set View

247

of the decision rule represents a throughflow of the corresponding branch. The throughflow of the graph is governed by formulas (1),...,(6). Formulas (1) and (2) say that an outflow of an input node or an output node is equal to their inflows. Formula (3) states that the outflow of the output node amounts to the sum of its inflows, whereas formula (4) says that the sum of outflows of the input node equals to its inflow. Finally, formulas (5) and (6) reveal how throughflow in the flow graph is distributed between its inputs and outputs.

27.8 Comparison of Bayesian and Rough Set Approach Now we will illustrate the ideas considered in the previous sections by means of the example considered in section 2. These examples intend to show clearly the difference between ”classical” Bayesian approach and that proposed by the rough set philosophy. Observe that we are not using data to verify prior knowledge, inherently associated with Bayesian data analysis, but the rough set approach shows that any decision table safisties Bayes’ theorem and total probability theorem. These properties form the basis of drawing conclusions from data, without referring either to prior or posterior knowledge. Example 27.8.1. This example, which is a modification of example 1 given in section 2, will clearly show the different role of Bayes’ theorem in classical statistical inference and that in rough set based data analysis. Let us consider the data table shown in Table 2. Table 27.2. Data table T+

T−

D

95

5

D

1998

97902

In Table 2, instead of probabilities, like those given in Table 1, numbers of patients belonging to the corresponding classes are given. Thus we start from the original data (not probabilities) represanting outcome of the test. Now from Table 2 we create a decision table and compute strength of decision rules. The results are shown in Table 3. In Table 3 D is the condition attribute, wheras T is the decision attribute. The decision table is meant to represent a ”cause–effect” relation between the disease and result of the test. That is, we expect that the disease causes positive test result and lack of the disease results in negative test result.

248

Z. Pawlak

Table 27.3. Decision table fact

D

T

support

strength

1

+

2

−

+

95

0.00095

+

1998

0.01998

3

+

−

5

0.00005

4

−

−

97902

0.97902

The decision algorithm is given below: 1’) 2’) 3’) 4’)

if if if if

(disease, (disease, (disease, (disease,

yes) then (test, positive) no) then (test, positive) yes) then (test, negative) no) then (test, negative)

The certainty and coverage factors of the decision rules for the above decision algorithm are given is Table 4. Table 27.4. Certainty and coverage rule

strength

certainty

coverage

1

0.00095

0.95

0.04500

2

0.01998

0.02

0.95500

3

0.00005

0.05

0.00005

4

0.97902

0.98

0.99995

The decision algorithm and the certainty factors lead to the following conclusions: -

95% persons suffering from the disease have positive test results 2% healthy persons have positive test results 5% persons suffering from the disease have negative test result 98% healthy persons have negative test result

That is to say that if a person has the disease most probably the test result will be positive and if a person is healthy the test result will be most probably negative. In other words, in view of the data there is a causal relationship between the disease and the test result. The inverse decision algorithm is the following: 1) if (test, positive) then (disease, yes) 2) if (test, positive) then (disease, no)

27. Bayes’ Theorem Revised – The Rough Set View

249

3) if (test, negative) then (disease, yes) 4) if (test, negative) then (disease, no) From the coverage factors we can conclude the following: - 4.5% persons with positive test result are suffering from the disease - 95.5% persons with positive test result are not suffering from the disease - 0.005% persons with negative test results are suffering from the disease - 99.995% persons with negative test results are not suffering from the disease That means that if the test result is positive it does not necessarily indicate the disease but negative test results most probably (almost for certain) does indicate lack of the disease. It is easily seen from Table 4 the negative test result almost exactly identifies healthy patients. For the remaining rules the accuracy is much smaller and consequently test results are not indicating the presence or absence of the disease.   It is clearly seen from examples 1 and 2 the difference between Bayesian data analysis and the rough set approach. In the Bayesian inference the data is used to update prior knowledge (probability) into a posterior probability, whereas rough sets are used to understand what the data are telling us.

27.9 Conclusion From examples 1 and 2 it is easily seen the difference between employing Bayes’ theorem in statistical reasoning and the role of Bayes’ theorem in rough set based data analysis. Bayesian inference consists in updating prior probabilities by means of data to posterior probabilities. In the rough set approach Bayes’ theorem reveals data patterns, which are used next to draw conclusions from data, in form of decision rules. In other words, classical Bayesian inference is based rather on subjective prior probability, whereas the rough set view on Bayes’ theorem refers to objective probability inherently associated with decision tables. Acknowledgments. The author wishes to express his gratitude to Professor Andrzej Skowron for many critical remarks.

References 27.1 Adams, E. W.: The logic of conditionals, an application of probability to deductive Logic. D. Reidel Publishing Company, Dordrecht, Boston (1975)

250

Z. Pawlak

27.2 Bayes, T.: An essay toward solving a problem in the doctrine of chances, Phil. Trans. Roy. Soc. 53 (1763) 370–418; Reprint Biometrika 45 (1958) 296–315 27.3 Bernardo, J. M., Smith, A. F. M.: Baysian theory, Wiley series in probability and mathematical statistics. John Wiley & Sons, Chichester, New York, Brisbane, Toronto, Singapore (1994) 27.4 Box, G.E.P., Tiao, G.C.: Bayesiaon inference in statistical analysis. John Wiley and Sons, Inc., New York, Chichester, Brisbane, Toronto, Singapore (1992) 27.5 Berthold, M., Hand, D.J.: Intelligent data analysis, an introduction. SpringerVerlag, Berlin , Heidelberg, New York (1999) 27.6 L  ukasiewicz, J.: Die logishen Grundlagen der Wahrscheinilchkeitsrechnung. Krak´ ow (1913). In: L. Borkowski (ed.), Jan L  ukasiewicz – Selected Works, North Holland Publishing Company, Amsterdam, London, Polish Scientific Publishers, Warsaw (1970) 27.7 Pawlak, Z.: Rough sets – theoretical aspect of reasoning about data, Kluwer Academic Publishers, Boston, Dordrech, London (1991) 27.8 Pawlak, Z.: New look on Bayes’ theorem – the rough set outlook. In: S. Hirano, M. Inuiguchi, S. Tsumoto (eds.), Proceedings of the International Workshop on Rough Set Theory and Granular Computing (RSTGC-2001), Ball of the International Set Society, Vol. 5 No.1/2, Matsue, Shimane, Japan, May 20-22 (2001) 1–8 27.9 Pawlak, Z.: Rough sets and decision algorithms. Springer-Verlag, Berlin, Heidelberg, New York (to appear) 27.10 Pawlak, Z., Skowron, A.: Rough membership functions. Advances in the Dempster-Shafer Theory of Evidence, R, Yager, M. Fedrizzi, J. Kacprzyk (eds.), John Wiley & Sons, Inc. ew York (1994) 251–271 27.11 Skowron, A.: Rough Sets in KDD (plenary talk); 16-th World Computer Congress (IFFIP’2000), Beijing, August 19-25, 2000, In:Zhongzhi Shi, Boi Faltings, Mark Musem (eds.) Proceedings of the Conference on Intelligent Information Processing (IIP2000), Publishing Hous of Electronic Industry, Beijing (2000) 1–17 27.12 Tsumoto, S., Tanaka, H.: Discovery of functional components of proteins based on PRIMEROSE and domain knowledge hierarchy. Proceedings of the Workshop on Rough Sets and Soft Computing (RSSC-94) (1994): Lin, T.Y., and Wildberger, A.M.(eds.) Soft Computing (1995) 280–285

28. Toward Intelligent Systems: Calculi of Information Granules Andrzej Skowron Institute of Mathematics, Warsaw University Banacha 2, 02-097 Warsaw, Poland [email protected]

We present an approach based on calculi of information granules as a basis for approximate reasoning in intelligent systems. Approximate reasoning schemes are defined by means of information granule construction schemes satisfying some robustness constraints. In distributed environments such schemes are extended to rough neural networks. Problems of learning in rough neural networks from experimental data and background knowledge are discussed. The approach is based on rough mereology.

28.1 Introduction Computing with Words (CWW) (see, e.g., [28.38], [28.39], [28.40]) is one among a collection of recently emerging computing paradigms. The goal of this new research direction is to build foundations for future intelligent computers and information systems performing computations on words from natural language representing concepts rather than on numbers. Information granulation belongs to intensively studied topics in soft computing (see, e.g., [28.38], [28.39], [28.40]). One of the recently emerging approaches to deal with information granulation is based on information granule calculi (see, e.g., [28.24], [28.33]). The development of such calculi is important for making progress in many areas like object identification by autonomous systems (see, e.g., [28.3], [28.36]), web mining (see, e.g., [28.8]), spatial reasoning (see, e.g., [28.4]) or sensor fusion (see, e.g., [28.2], [28.16], [28.19]). One way to achieve CWW is through Granular Computing (GC). The main concepts of GC are related to information granulation and in particular to information granules [28.24]. Any approach to information granulation should make it possible to define complex information granules (e.g., in spatial and temporal reasoning, one should be able to determine if the situation on the road (see Fig. 28.1) is safe on the basis of sensor measurements or to classify situations in complex games, like soccer [28.35]). These complex information granules consitute a form of information fusion. Any calculus of complex information granules should permit to (i) deal with vagueness of information granules, (ii) develop strategies of inducing multi-layered schemes of complex granule construction, (iii) derive robust (stable) information granule construction schemes with respect to deviations of granules from which they are constructed, and (iv) T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 251− 260, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

252

A. Skowron

develop adaptive strategies for reconstruction of induced schemes of complex information granule synthesis.

Fig. 28.1. Classification of situations

To deal with vagueness, one can adopt fuzzy set theory [28.37] or rough set theory [28.15] either separately or in combination [28.13]. The second requirement is related to the problem of understanding of reasoning from measurements to perception (see, e.g., [28.40]) and to concept approximation learning in layered learning [28.35] as well as to fusion of information from different sources (see, e.g., [28.38], [28.39], [28.40]). The importance of searching for Approximate Reasoning Schemes (AR-schemes, for short) as schemes of new information granule construction, is stressed in rough mereology (see, e.g., [28.20], [28.21], [28.21], [28.22], [28.26], [28.27]). In general, this leads to hierarchical schemes of new information granule construction. This process is closely related to ideas of co-operation, negotiations and conflict resolution in multi-agent systems [28.7]. Among important topics studied in relation to AR-schemes are methods for specifying operations on information granules; in particular, for their construction from data and background knowledge, and methods for inducing these hierarchical schemes of information granule construction. One of the possible approaches is to learn such schemes using evolutionary strategies [28.10]. Robustness of the scheme means that any scheme produces rather a higher order information granule that is a clump (e.g., a set) of close information granules rather than a single information granule. Such a clump is constructed by means of the scheme from the Cartesian product of input clumps (e.g., clusters) satisfying some constraints. The input clumps are defined by deviations (up to acceptable degrees) of input information granules. It is worthwhile to mention that modeling complex phenomena requires to use complex information granules representing local models (perceived by local agents) which next should be fused. This process involves the negotiations between agents [28.7] to resolve contradictions and conflicts in local modeling. This kind of modeling will become more and more important in solving

28. Toward Intelligent Systems

253

complex real-life problems which we are unable to model using traditional analytical approaches. If the latter approaches can be applied to modeling of such problems they lead to exact models . However, the necessary assumptions used to build them in case of complex real-life problems are often causing the resulting solutions to be too far from reality to be accepted as solutions of such problems. Let us also observe, using multi-agent terminology, that local agents perform operations on information granules from granule sets that are understandable by them. Hence, granules submitted as arguments by other agents should be approximated by means of properly tuned approximation spaces creating interfaces between agents. The process of tuning of the approximation space [28.32], [28.27] parameters in AR-schemes corresponds to the tuning of weights in neural networks. The methods for inducing of AR-schemes transforming information granules into information granules studied using rough set (see, e.g., [28.15], [28.9]) and rough mereological methods in hybridization with other soft computing approaches create a core for Rough Neurocomputing (RNC) (see, e.g., [28.14], [28.27]). In RNC, computations are performed on information granules. Another important problem concerns relationships between information granules and words (linguistic terms) in a natural language and also a possibility to use induced AR-schemes as schemes matching up to a satisfactory degree reasoning schemes in natural language. Further research in this direction will create strong links between RNC and CWW. The results of such research will be of great importance for many applications (e.g., web mining problems, Fig. 28.2).

Fig. 28.2. Web mining

RNC is attempting to define information granules using rough sets [28.15], [28.9] and rough mereology (see, e.g., [28.21], [28.21], [28.22], [28.26], [28.27]) introduced to deal with vague concepts in hybridization with other soft computing methods like neural networks [28.29], fuzzy sets [28.13], [28.37], [28.39]

254

A. Skowron

and evolutionary programming [28.14], [28.10]. The methods based on the above mentioned approaches can be used for constructing of more complex information granules by means of schemes analogous to neural networks. We outline a rough neurocomputing model as a basis for granular computing.

28.2 AR-Schemes AR-schemes are the basic constructs used in RNC. We assume each agent ag from a given collection Ag of agents [28.7] is equipped with a system of information granules S(ag) specifying information granules the agent ag is perceiving and the inclusion (or closeness) relations to a degree used by ag to measure the degree of inclusion (or closenees) between information granules. A formal definition of information granule system the reader can find, e.g., in [28.31]. Using such system S(ag) the agent ag creates a representation for all components of S(ag). The details of such representation the reader can find, e.g., in [28.22], [28.24]. From such representations agents are able to extract local schemes of approximate reasoning called productions. Algorithmic methods for extracting such productions from data are discussed in [28.21], [28.30], [28.34], [28.17], [28.18]. The left hand side of each production is (in (1) (1) (k) (k) the simplest case) of the form (st1 (ag), (1 , ·, r ), ·, (stk (ag), (1 , ·, r ) and the right hand side is of the form (st(ag), (1 , ·, r ) for some positive integers k, r. Such production represents an information about an operation o which can be performed by the agent ag. In the production k denotes the arity of operation. The operation o represented by the production is transforming standard (prototype) input information granules st1 (ag), · · · , stk (ag) into the standard (prototype) information granule st(ag). Moreover, if input informa(1) (k) tion granules g1 , · · · , gk are close to st1 (ag), · · · , stk (ag) to degrees j , ·, j then the result of the operation o on information granules g1 , · · · , gk is close to the standard st(ag) to a degree at least j where 1 ≤ j ≤ k. Standard (prototype) granules can be interpreted in different ways. In particular they can correspond to concept names in natural language. The described above productions are basic components of reasoning system over an agent set Ag. An important property of such productions is that they are expected to be discovered from available experimental data and background knowledge. Let us also observe that the degree structure is not necessarily restricted to reals from the interval [0, 1]. The inclusion degrees can have a structure of complex information granules used to represent the degree of inclusion. It is worthwhile to mention that the productions can also be interpreted as a constructive description of some operations on fuzzy sets. The methods for such constructive description are based on rough sets and Boolean reasoning (see, e.g., [28.9], [28.15]).

28. Toward Intelligent Systems

255

AR-schemes can be treated as derivations obtained by using productions from different agents. The relevant derivations defining AR-schemes are satisfying so called robustness (or stability) condition. It means that at any node of derivation the inclusion (or closeness) degree of constructed granule to the prototype (standard) granule is higher than required by the production to which the result should be sent. This makes it possible to obtain a sufficient robustness condition for the whole derivations. For details the reader is referred to, e.g., [28.22], [28.24], [28.25], [28.26]. In case where standards are interpreted as concept names in natural language and there is given a reasoning scheme in natural language over the standard concepts the corresponding AR-scheme represents a cluster of reasoning (constructions) approximately following (by mens of other information granule systems) the reasoning in natural language.

28.3 Rough Neural Networks We extend AR-schemes for synthesis of complex objects (or granules) developed in [28.24] and [28.22] by adding one important component. As a result we obtain granule construction schemes that can be treated as a generalization of neural network models. The main idea is that granules sent by one agent to another are not, in general, exactly understandable by the receiving agent. This is because these agents are using different languages and usually does not exist any translation (from the sender language to the receiver language) preserving exactly semantical meaning of formulas. Hence, it is necessary to construct interfaces that will make it possible to understand received granules approximately. These interfaces can be, in the simplest case, constructed on the basis of information exchanged by agents and stored in the form of decision data tables. From such tables the approximations of concepts can be constructed using rough set approach [28.33]. In general, it is a complex process because a high quality approximation of concepts can be often obtained only in dialog (involving nagotiations, conflict resolutions and cooperation) among agents. In this process the approximation can be constructed gradually when dialog is progressing. In our model we assume that for any n-ary operation o(ag) of an agent ag there are approximation spaces AS1 (o(ag), in), ..., ASn (o(ag), in) which will filter (approximate) the granules received by the agent for performing the operation o(ag). In turn, the granule sent by the agent after performing the operation is filtered (approximated) by the approximation space AS(o(ag), out). These approximation spaces are parameterized. The parameters are used to optimize the size of neighborhoods in these spaces as well as the inclusion relation [28.26]. A granule approximation quality is taken as the optimization criterion. Approximation spaces attached to any operation of ag correspond to neuron weights in neural networks whereas the operation performed by the agent ag on information granules corresponds to the operation realized on vectors of

256

A. Skowron

real numbers by the neuron. The generalized scheme of agents is returning a granule in response to input information granules. It can be for example a cluster of elementary granules. Hence, our schemes realize much more general computations than neural networks operating on vectors of real numbers. We call extended schemes for complex object construction rough neural networks (for complex object construction). The problem of deriving such schemes is closely related to perception (see, e.g., [28.1], [28.40]). The stability of such networks corresponds to the resistance to noise of classical neural networks. Let us observe that in our approach the deductive systems are substituted by productions systems of agents linked by approximation spaces, communication strategies and mechanism of derivation of AR-schemes. This revision of classical logical notions seems to be important for solving complex problems in distributed environments.

28.4 Decomposition of Information Granules Information granule decomposition methods are important components of methods for inducing of AR-schemes from data and background knowledge. Such methods are used to extract from data, local decomposition schemes called produtions [28.25]. The AR-schemes are constructed by means of productions. The decomposition methods are based on searching for the parts of information granules that can be used to construct relevant higher level patterns matching up to a satisfactory degree the target granule. One can distinguish two kinds of parts (represented, e.g., by sub-formulas or sub-terms) of AR-schemes. Parts of the first type are represented by expressions from a language, called the domestic language Ld , that has known semantics (consider, for example, semantics defined in a given information system [28.15]). Parts of the second type of AR-scheme are from a language, called foreign language Lf (e.g., natural language), that has semantics definable only in an approximate way (e.g., by means of patterns extracted using rough, fuzzy, rough–fuzzy or other approaches). For example, the parts of the second kind of scheme can be interpreted as soft properties of sensor measurements [28.3]. For a given expression e, representing a given scheme that consists of subexpressions from Lf first it is necessary to search for relevant approximations in Ld of the foreign parts from Lf and next to derive global patterns from the whole expression after replacing the foreign parts by their approximations. This can be a multilevel process, i.e., we are facing problems of discovered pattern propagation through several domestic-foreign layers. Productions from which AR-schemes are built can be induced from data and background knowledge by pattern extraction strategies. Let us consider some of such strategies. The first one makes it possible to search for relevant approximations of parts using the rough set approach. This means that each

28. Toward Intelligent Systems

257

part from Lf can be replaced by its lower or upper approximation with respect to a set B of attributes. The approximation is constructed on the basis of relevant data table [28.15], [28.9]. With the second strategy parts from Lf are partitioned into a number of sub-parts corresponding to cuts (or the set theoretical differences between cuts) of fuzzy sets representing vague concepts and each sub-part is approximated by means of rough set methods. The third strategy is based on searching for patterns sufficiently included in foreign parts. In all cases, the extracted approximations replace foreign parts in the scheme and candidates for global patterns are derived from the scheme obtained after the replacement. Searching for relevant global patterns is a complex task because many parameters should be tuned, e.g., the set of relevant features used in approximation, relevant approximation operators, the number and distribution of objects from the universe of objects among different cuts and so on. One can use evolutionary techniques [28.10] in searching for (semi-) optimal patterns in the decomposition. It has been shown that the decomposition strategies can be based on the developed rough set methods for decision rules generation and Boolean reasoning [28.21], [28.12], [28.17], [28.33]. In particular, methods for decomposition based on background knowledge can be developed [28.30], [28.18]. Conclusions. We have discussed a methodology for synthesis of AR-schemes and rough neural networks. For more details the reader is referred to [28.21], [28.22], [28.23], [28.24], [28.26], [28.27], [28.32], [28.33], [28.34]. We enclose a list of research directions related to the synthesis and analysis of AR-schemes and rough neural networks. 1. Developing foundations for information granule systems. Certainly, still more work is needed to develop solid foundations for synthesis and analysis of information granule systems. In particular, methods for construction of hierarchical information granule systems, and methods for representation of such systems should be developed. 2. Algorithmic methods for inducing parameterized productions. Some methods have already been reported such as discovery of rough mereological connectives from data (see, e.g., [28.21]) or methods based on decomposition (see, e.g., [28.22], [28.30], [28.34], [28.17]). However, these are only initial steps toward algorithmic methods for inducing of parameterized productions from data. One interesting problem is to determine how such productions can be extracted from data and background knowledge. A method in this direction has been proposed in [28.3]. 3. Algorithmic methods for synthesis of AR-schemes. It was observed (see, e.g., [28.22], [28.27]) that problems of negotiations and conflict resolutions are of great importance for synthesis of AR-schemes. The problem arises, e.g., when we are searching in a given set of agents for a granule sufficiently included or close to a given one. These agents, often working with different systems of information granules, can derive different gra-

258

A. Skowron

nules and their fusion will be necessary to obtain the relevant output granule. In the fusion process, the negotiations and conflict resolutions are necessary. Much more work should be done in this direction by using the existing results on negotiations and conflict resolution. In particular, Boolean reasoning methods seem to be promising ([28.22]) for solving such problems. Another problem is related to the size of production sets. These sets can be of large size and it is important to develop learning methods for extracting small candidate production sets in the process of extension of temporary derivations out of huge production sets. For solving this kind o problems methods for clustering of productions should be developed to reduce the size of production sets. Moreover, dialog and cooperation strategies between agents can help to reduce the search space in the process of AR-scheme construction from productions. 4. Algorithmic methods for learning in rough neural networks. A basic problem in rough neural networks is related to selecting relevant approximation spaces and to parameter tuning. One can also look up to what extent the existing methods for classical neural methods can be used for learning in rough neural networks. However, it seems that new approach and methods for learning of rough neural networks should be developed to deal with real-life applications. In particular, it is due to the fact that high quality approximations of concepts can be often obtained only through dialog and negotiations processes among agents in which gradually the concept approximation is constructed. Hence, for rough neural networks learning methods based on dialog, negotiations and conflict resolutions should be developed. In some cases, one can use directly rough set and Boolean reasoning methods (see, e.g., [28.33]). However, more advanced cases need new methods. In particular, hybrid methods based on rough and fuzzy approaches can bring new results [28.13]. 5. Fusion methods in rough neural neurons. A basic problem in rough neurons is fusion of the inputs (information) derived from information granules. This fusion makes it possible to contribute to the construction of new granules. In the case where the granule constructed by a rough neuron consists of characteristic signal values made by relevant sensors, a step in the direction of solving the fusion problem can be found in [28.19], [28.6]. Acknowledgements. I would like to thank to Professor Lech Polkowski for the years of close cooperation on rough mereology, to Professor Jaroslaw Stepaniuk for the cooperation on information granule models and to Professor James F. Peters for cooperation on sensor fusion methods, insightful comments and a number of clarifying discussions on the presented paper. The research has been supported by the State Committee for Scientific Research of the Republic of Poland (KBN) research grant 8 T11C 025 19 and by the Wallenberg Foundation grant.

28. Toward Intelligent Systems

259

References 28.1 Barsalou, L.W. (1999): Perceptual Symbol Systems, Behavioral and Brain Sciences 22, 577–660 28.2 Brooks, R.R., Iyengar, S.S. (1998): Multi-Sensor Fusion, Prentice-Hall PTR, Upper Saddle River, NJ 28.3 Doherty, P., L  ukaszewicz, W., Skowron A., Szalas, A. (2001): Combining Rough and Crisp Knowledge in Deductive Databases (submitted) 28.4 D¨ untsch I. (Ed.)(2001): Spatial Reasoning, Fundamenta Informaticae 46(12) (special issue) 28.5 Hirano, S., Inuiguchi, M., Tsumoto, S. (Eds.) (2001): Proc. RSTGC’01, Bulletin of International Rough Set Society 5(1-2) 28.6 Han, L., Peters, J.F., Ramanna, S., Zhai, R. (1999): Classifying Faults in High Voltage Power Systems: A Rough–Fuzzy Neural Computational Approach, Proc. RSFDGrC’99, Lecture Notes in Artificial Intelligence 1711, Springer Verlag, Berlin 47–54 28.7 Huhns, M.N., Singh, M.P. (Eds.) (1998): Readings in Agents, Morgan Kaufmann, San Mateo 28.8 Kargupta, H., Chan, Ph. (2001): Advances in Distributed and Parallel Knowledge Discovery, AAAI Press/MIT Press, Cambridge 28.9 Komorowski, J., Pawlak, P., Polkowski, L., and Skowron A. (1999): Rough Sets: A Tutorial, in [28.13] 3–98 28.10 Koza, J. R. (1994): Genetic Programming II: Automatic Discovery of Reusable Programs, MIT Press, Cambridge, MA 28.11 Lin T.Y. (1998): Granular Computing on Binary Relations I. Data Mining and Neighborhood Systems, in: [28.23] 18, 107–121 28.12 Nguyen, H.S., Nguyen, S.H., Skowron, A. (1999): Decomposition of Task Specification, Proc. ISMIS’99, Lecture Notes in Artificial Intelligence 1609, Springer-Verlag, Berlin, 310–318 28.13 Pal, S.K., Skowron, A. (Eds.) (1999): Rough-Fuzzy Hybridization: A New Trend in Decision Making, Springer-Verlag, Singapore 28.14 Pal, S.K., Pedrycz, W., Skowron, A., Swiniarski, R. (Eds.) (2001): RoughNeuro Computing, Neurocomputing 36, 1–262 (special issue) 28.15 Pawlak, Z. (1991): Rough Sets. Theoretical Aspects of Reasoning about Data, Kluwer Academic Publishers, Dordrecht 28.16 Peters, J.F., Ramanna, S., Skowron, A., Stepaniuk, J., Suraj, Z., Borkowsky, M. (2001): Sensor Fusion: A Rough Granular Approach, Proc. of Int. Fuzzy Systems Association World Congress (IFSA’01), Vancouver, July 2001 (to appear) 28.17 Peters, J.F., Skowron, A. Stepaniuk, J. (2001): Rough Granules in Spatial Reasoning, Proc. of Int. Fuzzy Systems Association World Congress (IFSA’01), Vancouver, July 2001 (to appear) 28.18 Peters, J.F., Skowron, A. Stepaniuk, J. (2001): Information Granule Decomposition, Fundamenta Informaticae (to appear) 28.19 Pawlak, Z., Peters, J.F., Skowron, A., Suraj, Z., Ramanna, S., Borkowsky, M. (2001): Rough Measures: Theory and Applications, in: [28.5] 177–183 28.20 Polkowski, L., Skowron, A. (1996): Rough Mereology: A New Paradigm for Approximate Reasoning, International J. Approximate Reasoning 15(4), 333–365 28.21 Polkowski, L., Skowron, A. (1996): Rough Mereological Approach to Knowledge-Based Distributed AI, (Eds.) J.K. Lee, J. Liebowitz, and J.M. Chae, Critical Technology, Proc. of the Third World Congress on Expert Systems, February 5-9, Seoul, Korea, Cognizant Communication Corporation, New York, 774–781

260

A. Skowron

28.22 Polkowski, L., Skowron, A. (1998): Rough Mereological Foundations for Design, Analysis, Synthesis, and Control in Distributed Systems, Information Sciences An International Journal 104(1-2), 129–156 28.23 Polkowski, L., Skowron, A. (Eds.) (1998): Rough Sets in Knowledge Discovery, Studies in Fuzziness and Soft Computing 18-19, Physica-Verlag / Springer-Verlag, Heidelberg (1998) 28.24 Polkowski, L., Skowron, A. (1999): Towards adaptive calculus of granules, in: [28.39] 30, 201–227 28.25 Polkowski, L., Skowron, A. (1999): Grammar Systems for Distributed Synthesis of Approximate Solutions Extracted from Experience, (Eds.) Paun, G., Salomaa, A., Grammar Systems for Multiagent Systems, Gordon and Breach Science Publishers, Amsterdam, 316–333 28.26 Polkowski, L., Skowron, A. (2000): Rough Mereology in Information Systems. A Case Study: Qualitative Spatial Reasoning, in [28.28] 89–135 28.27 Polkowski, L., Skowron, A. (2001): Rough-Neuro Computing, in: [28.42] 25– 32 (to appear) 28.28 Polkowski, L., Tsumoto, S., Lin, T.Y. (Eds.) (2000): Rough Set Methods and Applications. New Developments in Knowledge Discovery in Information Systems, Physica–Verlag, Heidelberg 28.29 Ripley, B.D. (1996): Pattern Recognition and Neural Networks, Cambridge University Press 28.30 Skowron, A. (2001): Toward Intelligent Systems: Calculi of Information Granules, in: [28.5] 9–30 28.31 Skowron, A. (2001): Approximate Reasoning by Agents in Distributed Environments, Proc. IAT’01 (to appear) 28.32 Skowron, A., Stepaniuk, J. (1996): Tolerance Approximation Spaces Fundamenta Informaticae 27(2-3), 245–253 28.33 Skowron, A., Stepaniuk, J. (2001): Information Granules: Towards Foundations of Granular Computing, International Journal of Intelligent Systems 16(1), 57–86 28.34 Skowron A., Stepaniuk, J., Peters, J.F. (2001): Extracting Patterns Using Information Granules, in: [28.5] 135–142 28.35 Stone, P. (2000): Layered Learning in Multiagent Systems: A Winning Approach to Robotic Soccer, MIT Press, Cambridge 28.36 WITAS project web page: http://www.ida.liu.se/ext/witas/eng.html 28.37 Zadeh, L.A. (1965): Fuzzy Sets, Information and Control 8 333–353 28.38 Zadeh, L.A. (1996): Fuzzy Logic = Computing with Words, IEEE Trans. on Fuzzy Systems 4, 103–111 28.39 Zadeh, L.A., Kacprzyk, J. (Eds.) (1999): Computing with Words in Information/Intelligent Systems, Studies in Fuzziness and Soft Computing 30-31, Physica–Verlag, Heidelberg 28.40 Zadeh, L.A. (2001): A New Direction in AI: Toward a Computational Theory of Perceptions, AI Magazine 22(1), 73–84 28.41 Zhong, N., Skowron, A., Ohsuga, S. (Eds.) (1999): Proc. RSFDGr’99, Lecture Notes in Artificial Intelligence 1711 Springer–Verlag, Berlin 28.42 Ziarko, W., Yao, Y.Y. (Eds.) (2001): Proc. RSCTC’2000, Lecture Notes in Artificial Intelligence 2005 Springer-Verlag, Berlin, 33–39 (to appear)

29. Soft Computing Pattern Recognition: Principles, Integrations, and Data Mining Sankar K. Pal Machine Intelligence Unit Indian Statistical Institute Calcutta 700035, India [email protected]

Relevance of fuzzy logic, artificial neural networks, genetic algorithms and rough sets to pattern recognition and image processing problems is described through examples. Different integrations of these soft computing tools are illustrated. Evolutionary rough fuzzy network which is based on modular principle is explained, as an example of integrating all the four tools for efficient classification and rule generation, with its various characterstics. Significance of soft computing approach in data mining and knowledge discovery is finally discussed along with the scope of future research.

29.1 Introduction Soft computing is a consortium of methodologies which work synergestically and provides in one form or another flexible information processing capabilities for handling real life ambiguous situations. Its aim is to exploit the tolerance for imprecision, uncertainty, approximate reasoning and partial truth in order to achieve tractability, robustness, low cost solutions, and close resemblance to human like decision making. In other words, it provides the foundation for the conception and design of high MIQ (Machine IQ) systems, and therefore forms the basis of future generation computing systems. At this juncture, Fuzzy Logic (FL), Rough Sets (RS), Artificial Neural Networks (ANN) and Genetic Algorithms (GA) are the principal components where FL provides algorithms for dealing with imprecision and uncertainty arising from vagueness rather than randomness, RS for handling uncertainty arising from limited discernibility of objects, ANN the machinery for learning and adaptation, and GA for optimization and searching [29.1, 29.2]. Machine recognition of patterns [29.3, 29.4] can be viewed as a two-fold task, consisting of learning the invariant and common properties of a set of samples characterizing a class, and of deciding that a new sample is a possible member of the class by noting that it has properties common to those of the set of samples. Therefore, the task of pattern recognition by a computer can be described as a transformation from the measurement space M to the feature space F and finally to the decision space D. Depending on the type of input patterns, one may have speech recognition system, image recognition or vision system, medical diagnostic system etc. T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 261− 271, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

262

S.K. Pal

In this article we first describe the relevance of different soft computing tools to pattern recognition problems with examples. Different integration among them are then described. As an example we explain an evolutionary rough fuzzy MLP, which has been designed based on modular concept for pattern classification and rule generation. Finally the significance of soft computing in data mining and knowledge discovery is discussed.

29.2 Relevance of Fuzzy Set Theory in Pattern Recognition Fuzzy sets were introduced in 1965 by Zadeh [29.5] as a new way to represent vagueness in everyday life. They are generalizations of conventional (crisp) set theory. Conventional sets contain objects that satisfy precise properties required for membership. Fuzzy sets, on the other hand, contain objects that satisfy imprecisely defined properties to varying degrees. A fuzzy set A of the universe X is defined as a collection of ordered pairs A = {(μA (x), x), ∀x ∈ X} where μA (x), (0 ≤ μA (x) ≤ 1) gives the degree of belonging of the element x to the set A or the degree of possession of an imprecise property represented by A. Different aspects of fuzzy set theory including membership functions, basic operations and uncertainty measures can be found in [29.5, 29.6]. In this section we explain some of the uncertainties which one often encounters while designing a pattern recognition system and the relevance of fuzzy set theory in handling them. Let us consider, first of all, the case of processing and recognition of a gray-tone image pattern. Conventional approaches to image analysis and recognition [29.7, 29.8] consist of segmenting the image into meaningful regions, extracting their edges and skeletons, computing various features (e.g., area, perimeter, centroid etc.) and primitives (e.g., line, corner, curve etc.) of and relationships among the regions, and finally, developing decision rules and grammars for describing, interpreting and/or classifying the image and its sub-regions. In a conventional system each of these operations involves crisp decisions (i.e., yes or no, black or white, 0 or 1) to make regions, features, primitives, properties, relations and interpretations crisp. Since the regions in an image are not always crisply defined, uncertainty can arise within every phase of the aforesaid tasks. Any decision made at a particular level will have an impact on all higher level activities. An image recognition system should have sufficient provision for representing and manipulating the uncertainties involved at every processing stage; i.e., in defining image regions, features and relations among them, so that the system retains as much of the ‘information content’ of the data as possible. If this is done, the ultimate output (result) of the system will possess minimal uncertainty

29. Soft Computing Pattern Recognition

263

(and unlike conventional systems, it may not be biased or affected as much by lower level decision components). In Short, gray information is expensive and informative. Once it is thrown away, there is no way to get it back. Therefore one should try to retain this information as long as possible throughout the decision making tasks for its full use. When it is required to make a crisp decision at the highest level one can always through away or ignore this information. Let us now consider the case of a decision-theoretic approach to pattern classification. With the conventional probabilistic and deterministic classifiers [29.3, 29.4], the features characterizing the input patterns are considered to be quantitative (numeric) in nature. The patterns having imprecise or incomplete information are usually ignored or discarded from their designing and testing processes. The impreciseness (or ambiguity) may arise from various causes. For example, instrumental error or noise corruption in the experiment may lead to only partial or partially reliable information being available on a feature measurement F . Again, in some cases it may become convenient to use linguistic variables and hedges. In such cases, it is not appropriate to give exact representation to uncertain feature data. Rather, it is reasonable to represent uncertain feature information by fuzzy subsets. Again, uncertainty in classification or clustering of patterns may arise from the overlapping nature of the various classes. This overlapping may result from fuzziness or randomness. In the conventional technique, it is usually assumed that a pattern may belong to only one class, which is not necessarily true in real life applications. A pattern can and should be allowed to have degrees of membership in more than one class. It is, therefore, necessary to convey this information while classifying a pattern or clustering a data set. From the aforementioned examples, we see that the concept of fuzzy sets can be used at the feature level in representing input data as an array of membership values denoting the degree of possession of certain properties, in representing linguistically phrased input features for their processing, in weakening the strong commitments for extracting ill-defined image regions, properties, primitives, and relations among them, and at the classification level, for representing class membership of objects in terms of membership values. In other words, fuzzy set theory provides a notion of embedding: We find a better solution to a crisp problem by looking in a large space at first, which has different (usually less) constraints and therefore allows the algorithm more freedom to avoid errors forced by commission to hard answers in intermediate stages. The capability of fuzzy set theory in pattern recognition problems has been reported adequately since late sixties. A cross-section of the advances with applications is available in [29.6, 29.2, 29.9].

264

S.K. Pal

29.3 Relevance of Neural Network Approaches Neural network (NN) models [29.10, 29.11] try to emulate the biological neural network/nervous system with electronic circuitry. NN models have been studied for many years with the hope of achieving human-like performance (artificially), particularly in the field of pattern recognition, by capturing the key ingredients responsible for the remarkable capabilities of the human nervous system. Note that these models are extreme simplifications of the actual human nervous system. NNs are designated by the network topology, connection strength between pairs of neurons (called weights), node characteristics and the status updating rules. Node characteristics mainly specify the primitive types of operations it can perform, like summing the weighted inputs coming to it and then amplifying it or doing some fuzzy aggregation operations. The updating rules may be for weights and/or states of the processing elements (neurons). Normally an objective function is defined which represents the complete status of the network and the set of minima of it corresponds to the set of stable states of the network. Since there are interactions among the neurons the collective computational property inherently reduces the computational task and makes the system fault tolerant. Thus NN models are also suitable for tasks where collective decision making is required. Hardware implementations of neural networks are also attempted. Neural network based systems are usually reputed to enjoy the following major characteristics: – – – – –

adaptivity- adjusting the connection strengths to new data/information, speed- due to massively parallel architecture, robustness- to missing, confusing, ill-defined/noisy data, ruggedness- to failure of components, optimality- as regards error rates in performance.

For any pattern recognition system, one desires to achieve the above mentioned characteristics. More over, there exists some direct analogy between the working principles of many pattern recognition tasks and neural network models. For example, image processing and analysis in the spatial domain mainly employ simple arithmetic operations at each pixel site in parallel. These operations usually involve information of neighboring pixels (co-operative processing) in order to reduce the local ambiguity and to attain global consistency. An objective measure is required (representing the overall status of the system), the optimum of which represents the desired goal. The system thus involves collective decisions. On the other hand, we notice that neural network models are also based on parallel and distributed working principles (all neurons work in parallel and independently). The operations performed at each processor site are also simpler and independent of the others. The overall status of a neural network can also be measured.

29. Soft Computing Pattern Recognition

265

Again, the task of recognition in a real-life problem involves searching a complex decision space. This becomes more complicated particularly when there is no prior information on class distribution. Neural network based systems use adaptive learning procedures, learn from examples and attempt to find a useful relation between input and output, however complex it may be, for decision-making problems. Neural networks are also reputed to model complex non-linear boundaries and to discover important underlying regularities in the task domain. These characteristics demand that methods are needed for constructing and refining neural network models for various recognition tasks. In short, neural networks are natural classifiers having resistance to noise, tolerance to distorted images/patterns (ability to generalize), superior ability to recognize partially occluded or degraded images/overlapping pattern classes or classes with highly nonlinear boundaries, and potential for parallel processing.

29.4 Genetic Algorithms for Pattern Recognition Genetic Algorithms (GAs) [29.12, 29.13, 29.14, 29.15] are adaptive computational procedures modeled on the mechanics of natural genetic systems. They express their ability by efficiently exploiting the historical information to speculate on new offspring with expected improved performance [29.12]. GAs are executed iteratively on a set of coded solutions, called population, with three basic operators: selection/reproduction, crossover and mutation. They use only the payoff (objective function) information and probabilistic transition rules for moving to the next iteration. They are different from most of the normal optimization and search procedures in four ways: – GAs work with the coding of the parameter set, not with the parameter themselves. – GAs work simultaneously with multiple points, and not a single point. – GAs search via sampling (a blind search) using only the payoff information. – GAs search using stochastic operators, not deterministic rules. One may note that the methods developed for pattern recognition and image processing are usually problem dependent. Moreover, many tasks involved in the process of analyzing/identifying a pattern need appropriate parameter selection and efficient search in complex spaces in order to obtain optimal solutions. This makes the process not only computationally intensive, but also leads to a possibility of losing the exact solution. Therefore, the application of genetic algorithms for solving certain problems of pattern recognition, which need optimization of computation requirements, and robust, fast and close approximate solution, appears to be appropriate and natural [29.13].

266

S.K. Pal

29.5 Integration and Hybrid Systems Integration of the individual soft computing tools help in designing hybrid systems which are more versatile and efficient compared to stand alone use of the tools. The most visible integration in soft computing community is that of neural networks and fuzzy sets [29.2]. Neuro-fuzzy systems has been successfully developed for decision making, pattern recognition and image processing tasks. The hybridization falls in two major categories: a neural network equipped with the capability of handling fuzzy information (termed fuzzy neural network) to augment its application domain, and a fuzzy system augmented by neural networks to enhance some of its characterstics like flexibility, speed, adaptivility, learning (termed neural-fuzzy systems). Both the classes of hybridisation and their application to various pattern recognition problem are described in [29.2]. There are some applications where the integration of GAs with fuzzy sets and ANNs is found to be effective. For example GAs are found sometimes essential for overcoming some of the limitations of fuzzy set theory, specifically to reduce the ‘subjective’ nature of membership functions. Note that the other way of integration, i.e., incorporating the concept of fuzziness into GAs has not been tried seriously. Synthesis of ANN architectures can be done using GAs as an example of neuro-genetic systems. Such an integration may help in designing optimum ANN architecture with appropiate parameter sets. Methods for designing neural network architectures using GAs are primarily divided into two parts. In one part the GA replaces the learning method to find appropiate connection weights of some predefined architecture. In another part, GAs are used to find the architecture itself and it is then evaluated using some learning algorithms. Literature is also available on integration of fuzzy sets, neural networks and genetic algorithms [29.2, 29.16, 29.17]. The theory of rough sets [29.18] has emerged as another major mathematical approach for managing uncertainty that arises from inexact, noisy, or incomplete information. It is turning out to be methodologically significant to the domains of artificial intelligence and cognitive sciences, especially in the representation of and reasoning with vague and/or imprecise knowledge, data classification, data analysis, machine learning, and knowledge discovery [29.19]. Recently, rough sets have been integrated with both fuzzy sets and neural networks. Several rough-fuzzy hybrid systems are discussed in [29.2]. In the framework of rough-neuro integration [29.20], two broad approaches are available, namely, use of roughs set for encoding weights of knowledge based networks [29.21], and designing neural network architectures which incorporate roughness in the neuronal level. Genetic algorithms have also been used for fast generation of rough set reducts from an indiscernibility matrix. In the next section we describe, as an example, a methodology for integrating all the four soft computing tools, viz., fuzzy sets, ANN, rough sets

29. Soft Computing Pattern Recognition

267

and GAs for classification and rule generation. Here rough sets are used to encode domain knowledge in network parameters of a fuzzy MLP. GAs are used to evolve the optimal architecture based on modular concept.

29.6 Evolutionary Rough Fuzzy MLP The evolutionary rough fuzzy MLP utilises the concept of modular learning for better integration and performance enhancement [29.22]. The knowledge flow structure of evolutionary rough fuzzy MLP is illustrated in Figure 29.1. Here each of the soft computing tools act synergestically to contribute to the final performance of the system as follows. Rough set rules are used for extracting crude domain knowledge, which when encoded in a fuzzy MLP not only results in fast training of the network, but also automatic determination of the network size. The GA operators are adaptive and use the domain knowledge extracted with rough sets for even faster learning. The fuzziness incorporated at the input and outputs helps in better handling of uncertainties and overlapping classes. The nature of integration is illustrated in Figure 29.2. R o u g h S e t R u le s C1

(L 1

M2 )

F e a tu r e S p a c e

(M1

H 2 ) (R 1)

C 1 (R 1 )

C2

M2

C2

L 2

H1 L 1

(R 2) F

2

C 1

C 2 (R 2 )

(R 3 )

N e tw o r k M a p p in g

C 2

(R 3 ) F1

R 1 (S u b n e t 1 )

R 2 (S u b n e t 2 )

R 3 (S u b n e t 3 )

P a r tia l T r a in in g w ith O r d in a r y G A (S N 1 )

(S N 2 )

(S N 3 )

F e a tu r e S p a c e (S N 1 ) (S N 2 )

P a r tia lly R e fin e d S u b n e tw o r k s

(S N 3 )

C o n c a te n a tio n o f S u b n e tw o r k s

E v o lu tio n o f th e P o p u la tio n o f C o n c a te n a te d n e tw o r k s w ith G A h a v in g v a r ia b le m u ta tio n o p e r a to r

F e a tu r e S p a c e

C 1

C 2

F in a l S o lu tio n N e tw o r k

Fig. 29.1. Knowledge Flow in Modular Rough Fuzzy MLP

268

S.K. Pal I n c o r p o r a te D o m a in K n o w le d g e U s in g R o u g h S e ts

L F

M H

F

F

j

F

M e m b e r s h ip

h -1

jL

w ji

h

x j

y jh

μ 1

jM

μ 2

jH

μ 3

P a r a m e te r s

μ 1 μ 2

.

μ 3

B o u n d a r y

G A T u n in g X X | 0 0 0 | X X | | 0 0 | X X X | 0 0

Fig. 29.2. Components of the Modular Rough-fuzzy MLP

The evolutionary modular rough fuzzy MLP has been applied to a number of real world problems like speech recognition and medical diagnosis. In case of speech recognition [29.22], the system is found to correctly classify 84% of the samples, while the fuzzy MLP correctly classifies only 78% and the MLP only 59%. The system also gained in computation time significantly. For determining the stages of Cervical Cancer [29.22], the system provides results identical to that of medical experts in 83% of the cases. In other cases also the stagings were close. In addition to the above performance logical rules were extracted from the trained system. It was found that the rules coincided with the guidelines adopted by medical practicioners for staging. In the rough fuzzy MLP, the final network has a structure imposed on the weights. Hence, crisp logical rules can be easily extracted from the networks. This makes the system suitable for Knowledge Discovery in Databases. The rules obtained are found to be superior to those of several popular methods, as measured with some quantitative indices. For example, on the speech recognition data, the rules obtained using the modular rough-fuzzy MLP have an accuracy of 81.02% with 10 rules, while the popular C4.5 rule generation algorithm have accuracy of 75.00% using 16 rules. Fraction of samples which are ‘uncovered’ by the rules obtained by us is only 3.10%, whereas the C4.5 rules have 7.29% uncovered samples. The ‘confusion index’ is also low for the proposed method (1.4) compared to C4.5 (2.0).

29.7 Data Mining and Knowledge Discovery In recent years, the rapid advances being made in computer technology have ensured that large sections of the world population have been able to gain easy access to computers on account of falling costs worldwide, and their use is now commonplace in all walks of life. Government agencies, scientific, business and commercial organizations are routinely using computers not just for computational purposes but also for storage, in massive databases, of the immense volumes of data that they routinely generate, or require from other sources. Large-scale computer networking has ensured that such data has become accessible to more and more people. In other words, we are in the

29. Soft Computing Pattern Recognition

269

D a ta M in in g (D M )

.

D a ta C le a n in g

.

D a ta C o n d e n s a tio n

. H u g e H e te r o g e n e o u s R a w D a ta

.

D im e n s io n a lity R e d u c tio n

D a ta W r a p p in g

M a c h in e L e a r n in g

P r e p r o c e sse d D a ta

. .

.

C la s s ific a tio n C lu s te r in g

M a th e m a tic a l M o d e l o f D a ta

K n o w le d g e I n te r p r e ta tio n U se fu l

.

(P a tte r n s)

R u le G e n e r a tio n

.

K n E K n E

o w le d g e x tr a c tio n o w le d g e v a lu a tio n

K n o w le d g e

K n o w le d g e D is c o v e r y in D a ta b a s e (K D D )

Fig. 29.3. Block diagram for Knowledge Discovery in Databases (KDD)

midst of an information explosion, and there is urgent need for methodologies that will help us bring some semblance of order into the phenomenal volumes of data that can readily be accessed by us with a few clicks of the keys of our computer keyboard. Traditional statistical data summarization and database management techniques are just not adequate for handling data on this scale, and for extracting intelligently, information or, rather, knowledge that may be useful for exploring the domain in question or the phenomena responsible for the data, and providing support to decision-making processes. This quest had thrown up some new phrases, for example, data mining and knowledge discovery in databases (KDD), which are perhaps self-explanatory, but will be briefly discussed in the next few paragraphs. Their relationship with the discipline of pattern recognition will also be examined. The massive databases that we are talking about are generally characterized by the presence of not just numeric, but also textual, symbolic, pictorial and aural data. They may contain redundancy, errors, imprecision, and so on. KDD is aimed at discovering natural structures within such massive and often heterogeneous data. Therefore PR plays a significant role in KDD process. However, KDD is being visualized as not just being capable of knowledge discovery using generalizations and magnifications of existing and new pattern recognition algorithms, but also the adaptation of these algorithms to enable them to process such data, the storage and accessing of the data, its preprocessing and cleaning, interpretation, visualization and application of the results, and the modeling and support of the overall human-machine interaction. What really makes KDD feasible today and in the future is the rapidly falling cost of computation, and the simultaneous increase in computational power, which together make possible the routine implementation of sophisticated, robust and efficient methodologies hitherto thought to be too computation-intensive to be useful. A block diagram of KDD is given in Figure 29.3. Data mining is that part of knowledge discovery which deals with the process of identifying valid, novel, potentially useful, and ultimately understandable patterns in data, and excludes the knowledge interpretation part

270

S.K. Pal

of KDD. Therefore, as it stands now, data mining can be viewed as applying PR and machine learning principles in the context of voluminous, possibly heterogeneous data sets. Furthermore, soft computing-based (involving fuzzy sets, neural networks, genetic algorithms and rough sets) PR methodologies and machine learning techniques seem to hold great promise for data mining. The motivation for this is provided by their ability to handle imprecision, vagueness, uncertainty, approximate reasoning and partial truth and lead to tractability, robustness and low-cost solutions. In this context, case-based reasoning [29.17], which is a novel Artificial Intelligence (AI) problem-solving paradigm, has a significant role to play, as is evident from the recent book edited by Pal, Dillon and Yeung [29.17]. Some of the challenges that researchers in this area are likely to deal with, include those posed by massive data sets and high dimensionality, nonstandard and incomplete data, and overfitting. The focus is most likely to be on aspects like user interaction, use of prior knowledge, assessment of statistical significance, learning from mixed media data, management of changing data and knowledge, integration of tools, ways of making knowledge discovery more understandable to humans by using rules, visualization, etc., and so on. We believe the next decade will bear testimony to this.

References 29.1 L. A. Zadeh. Fuzzy logic, neural networks, and soft computing. Communications of the ACM, 37:77–84, 1994. 29.2 S. K. Pal and S. Mitra. Neuro-fuzzy Pattern Recognition: Methods in Soft Computing. John Wiley, New York, 1999. 29.3 R. O. Duda and P. E. Hart. Pattern Classification and Scene Analysis. John Wiley, New York, 1973. 29.4 J. T. Tou and R. C. Gonzalez. Pattern Recognition Principles. AddisonWesley, London, 1974. 29.5 L. A. Zadeh. Fuzzy sets. Information and Control, 8:338–353, 1965. 29.6 S. K. Pal and D. Dutta Majumder. Fuzzy Mathematical Approach to Pattern Recognition. John Wiley (Halsted Press), New York, 1986. 29.7 A. Rosenfeld and A. C. Kak. Digital Picture Processing, volume 1-2. Academic Press, New York, 1982. 29.8 R. C. Gonzalez and P. Wintz. Digital Image Processing. Addison-Wesley, Reading, MA, 1987. 29.9 J. C. Bezdek and S. K. Pal, editors. Fuzzy Models for Pattern Recognition: Methods that Search for Structures in Data. IEEE Press, New York, 1992. 29.10 D. E. Rumelhart and J. L. McClelland, editors. Parallel Distributed Processing: Explorations in the Microstructures of Cognition, volume 1. MIT Press, Cambridge, MA, 1986. 29.11 R. P. Lippmann. Pattern classification using neural networks. IEEE Communications Magazine, pages 47–64, 1989. 29.12 D. E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley, Reading, MA, 1989. 29.13 S. K. Pal and P. P. Wang, editors. Genetic Algorithms for Pattern Recognition. CRC Press, Boca Raton, 1996.

29. Soft Computing Pattern Recognition

271

29.14 L. B. Booker, D. E. Goldberg, and J. H. Holland. Classifier systems and genetic algorithms. Artificial Intelligence, 40:235–282, 1989. 29.15 J. H. Holland. Adaptation in Natural and Artificial Systems. University of Michigan Press, Ann Arbor, 1975. 29.16 S.K. Pal, A. Ghosh, and M.K. Kundu, editors. Soft Computing for Image Processing. Physica Verlag, Heidelberg, 2000. 29.17 S.K. Pal, T.S. Dillon, and D.S. Yeung. Soft Computing in Case Based Reasoning. Springer Verlag, London, 2000. 29.18 Z. Pawlak. Rough Sets, Theoretical Aspects of Reasoning about Data. Kluwer Academic, Dordrecht, 1991. 29.19 R. Slowi´ nski, editor. Intelligent Decision Support, Handbook of Applications and Advances of the Rough Sets Theory. Kluwer Academic, Dordrecht, 1992. 29.20 S. K. Pal, W. Pedrycz, A. Skowron, and R. Swiniarski (eds). Spl. issue on rough-neuro computing. Neurocomputing, 36(1-4), 2001. 29.21 M. Banerjee, S. Mitra, and S. K. Pal. Rough fuzzy MLP: Knowledge encoding and classification. IEEE Transactions on Neural Networks, 9(6):1203–1216, 1998. 29.22 P. Mitra, S. Mitra, and S. K. Pal. Staging of cervical cancer using soft computing. IEEE Transactions on Biomedical Engineering, 47(7):934–940, 2000.

3 0 . I d e n tify in g U p p e r a n d L o w e r P o s s ib ility D is tr ib u tio n s w ith R o u g h S e t C o n c e p t 1

2

P . G u o a n d Hi d eo Ta n a k a 1

2

F a cu l ty g u o @ e Toy oh a Ja p a n ,

of E con om i cs , Ka g a wa Un i ver s i ty , Ta k a m a ts u , Ka g a wa , 760-8 523 Ja p a n , c . k a g a w a - u . a c . j p s h i Soz o Col l eg e, 20-1 Ma ts u s h i ta , Us h i k a wa ch o, Toy oh a s h i , 4 4 0-8 511 t a n a k a h @ s o z o . a c . j p

3 0 .1 C o n c e p t s o f U p p e r a n d L o w e r P o s s ib ilit y D is t r ib u t io n s on e ex p er t ca n b e r ep r es en ted b y a d a ta s et { (x i , h i ) | i = 1, ..., m }

Kn owl ed g e fr om

= [ x i 1 , L , x i n ] t i s a n n -d i m en s i on a l vector to ch a r a cter i z e s om e s p eci fi ed even t, h i i s a n a s s oci a ted p os s i b i l i ty g r a d e g i ven b y a n ex p er t to r efl ect h i s j u d g em en t on wh a t th e p os s i b i l i ty g r a d e of th e i th s a m p l e i s for th i s even t, a n d m i s th e n u m b er of s a m p l es . Th e d a ta s et ( x i , h i ) (i = 1, ..., m ) ca n b e a p p r ox i m a ted b y a d u a l d a ta s ets ( x i , h l i ) a n d ( x i , h u i ) (i = 1, ..., m ) wi th th e con d i ti on h l i ≤ h i ≤ h u i . As s u m e th a t th e va l u es h l i a n d h u i a r e fr om a cl a s s of th e fu n cti on s G (x , θ ) wi th p a r a m eter vector θ . L et G (x i , l ) a n d G (x i , u ) cor r es p on d to h l i a n d h u i (i = 1, ..., m ), r es p ecti vel y a n d s i m p l y d en ote a s π l (x i ) a n d π u (x i ) . G i ven th e d a ta s et ( x i , h i ) (i = 1, ..., m ), th e ob j ecti ve of es ti m a ti on i s to ob ta i n two op ti m a l p a wh er e x

i

r a m eter vector s fr om u p p er a n d d u a l op ti m a l p a h ol d for a n y a r b Su p p os e th

θ

a n d θ l * fr om th e p a l ower d i r ecti on s a ccor d i n g r a m eter vector s ( θ u * , θ l * ) i tr a r y n -d i m en s on a l vector G (x a t th e fu n cti on

ex p { − (x − a ) t D m u l a s h ol d .

− 1 A

π

l

(x i ) = ex p { − (x u

(x i ) = ex p { − (x

(x i ) ≤ h i

≤ π u

− 1

− a )tD i

i

l

(x i ) a n d π

(x

− 1

− a )tD u

l

a ce to a p p r ox i m a te ( x i , h i ) i ven m ea s u r e. Mor eover , th e r el a ti on G (x , θ l * ) ≤ G (x , θ u * )

r a m eter s p to s om e g m a k e th e x . ,θ ) is

( x − a )} , s i m p l y d en oted a s (a , D

l

π π

*

u

(x

i

i

(x ) ≤ π

a n A

ex p on en ti a l

fu n cti on

) e . Th en th e fol l owi n g for -

− a )} , i = 1, … , m ,

(1)

− a )} , i = 1, … , m ,

(2)

u

(x ) ,

i = 1, …

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 272-277, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

,m ,

(3 )

3 0. Id en ti fy i n g Up p er a n d L ower P os s i b i l i ty D i s tr i b u ti on s

wh er e a = [ a 1 , a 2 , L, a n ]

D

tr i ces , d en oted a s

t

i s a cen ter vector , D

> 0 a n d u

D l

a n d D u

273

a r e p os i ti ve d efi n i te m a l

> 0 , r es p ecti vel y . It ca n b e s een th a t i n th e a n d m a tr i ces D

a b ove ex p on en ti a l fu n cti on , vector a b e s ol ved . D i ffer en t p a r a m eter s a , D

a n d u

D

a n d D u

a r e p a r a m eter s to l

l ea d to d i ffer en t va l u es π l

l

(x i )

a n d π u (x i ) wh i ch a p p r ox i m a te th e g i ven p os s i b i l i ty d eg r ee h i of x i to th e d i ffer en t ex ten t. D e f i n i t i o n 1 . G i ven th e for m u l a s (1), (2) a n d (3 ), th e fi tn es s of a p p r ox i m a ti on b a s ed on p a r a m eter s a , D u a n d D l , d en oted a s β , i s d efi n ed a s fol l ows :

β =

∏

i

m

va l u es π

π l

l

(x i ) a n d π

(x i ) a n d

i= 1

∏

m

∏

m i= 1

l u

(4 )

(x i ) . (x i )

D efi n i ti on 1 th a t th e h i g h er th e p a r a m eter β

It i s k n own fr om

h

m

π

u

(x i ) a r e fr om

i s , th e cl os er to

l ower a n d u p p er d i r ecti on s , r es p ecti vel y .

(x i ) ca n b e r eg a r d ed a s l i k el i h ood fu n cti on s for l ower a n d u

i= 1

u p p er p os s i b i l i ty d i s tr i b u ti on s . D e f i n i t i o n 2 . D en ote th e op ti m a l s ol u ti on s of a , D

D * l , r es p ecti vel y , wh i ch m a x i m i z e β fu n cti on s

π

* l

π

* u

wi th

− 1 * l

(x ) = ex p { − (x − a * ) t D

− 1 * u

*

D l

a s a *

, D

* u

a n d

th e con s tr a i n t (3 ). Th e fol l owi n g

) tD

(x ) = ex p { − (x − a

a n d u

)}

(5)

( x − a * )}

(6)

(x − a *

a r e ca l l ed l ower a n d u p p er ex p on en ti a l p os s i b i l i ty d i s tr i b u ti on s of th e p os s i b i l i ty vector X , r es p ecti vel y . F or s i m p l i ci ty a fter wa r d s we wr i te π u (x ) a n d π l (x ) i n s tea d of π

* u

(x ) a n d π

* l

(x ) .

3 0 .2 C o m p a r is o n o f D u a l P o s s ib ilit y D is t r ib u t io n s w it h D u a l A p p r o x im a tio n s in R o u g h S e ts T h e o r y R ou g h s ets th eor y h a s b een p r op os ed b y P a wl a k a n d ex ten s i vel y a p p l i ed fi ca ti on p r ob l em s , m a ch i n e l ea r n i n g , a n d d eci s i on a n a l y s i s etc. [ 1, 2] . p a r i n g th e d u a l p os s i b i l i ty d i s tr i b u ti on s wi th th e r ou g h s ets , th e b a s i c n r ou g h s ets a r e i n tr od u ced b el ow. L et U b e th e u n i ver s e of ob j ects a n d R b e a n eq u i va l en ce r el a ti on i n U U / R we m ea n th e fa m i l y of a l l eq u i va l en ce cl a s s of R . E q u i va l en ce cl a s r el a ti on R a r e ca l l ed el em en ta r y s ets . An y fi n i te u n i on of el em en ta r y s ets b e a d efi n a b l e s et. G i ven a s et Z , th e u p p er a n d l ower a p p r ox i m a ti on s

to cl a s s i F or com oti on s of . Th en b y s es of th e i s s a i d to of Z , d e-

274

P . G u o a n d H. Ta n a k a *

R

n oted a s l ows :

(Z ) a n d

R * ( Z ) , r es p ecti vel y a r e two d efi n a b l e s ets d efi n ed a s fol -

( Z ) = U{ Y ∈ U *

R

R wh er e ∅ i s th e em p ty It ca n b e s een th a t th s et con ta i n i n g th e s et Z d efi n a b l e s et con ta i n ed r a cy m ea s u r e of a s et Z

*

s et. eu p a n d in Z , d en

( Z ) = U{ Y ∈ U

R

*

*

(7)

/R : Y ⊆ Z }

(8 )

p er a p p r ox i m a ti on of Z i s th e l ower a p p r ox i m a ti on s o th a t th e con d i ti on R * oted a s α (Z ) , i s d efi n ed

α (Z ) =

wh er e C a r d ( R

/R : Y IZ ≠ ∅ } ,

C a r d (R

( Z )) a n d

C a r d (R C a r d (R *

d efi n ed a s th e l ea s t d efi n a b l e of Z i s d efi n ed a s th e g r ea tes t ( Z ) ⊇ R * ( Z ) h ol d s . An a ccu a s (9)

( Z )) * *

( Z ))

( Z )) a r e th e ca r d i n a l i ti es

of R *

(Z ) a n d

(Z ) .

3 0 .3 I d e n t if ic a t io n o f U p p e r a n d L o w e r P o s s ib ilit y D is t r ib u t io n s Th e u p p er a n d l ower a p p r ox i m a ti on s of Z ca n b e r eg a r d ed a s th e op ti m a l s ol u ti on s of th e fol l owi n g op ti m i z a ti on p r ob l em .

C a r d ( R l ( Z )) C a r d ( R u ( Z ))

m a x

a (Z ) =

s . t.

R l (Z ) ⊆ Z ⊆ R

R u ( Z ), R l ( Z )

u

(10)

(Z ) ,

wh er e R l (Z ) a n d R u (Z ) a r e d efi n a b l e s ets b y U / R . Si m i l a r l y th e m od el to i d en ti fy th e u p p er a n d l ower p os s i b i l i ty d i s tr i b u ti on s ca n b e for m u l a ted to m a x i m i z e th e fi tn es s m ea s u r e a s fol l ows :

β =

m a x

a ,D

u

,D l

s . t. l u

m

∏

π m

i= 1

π u

l

(x i ) (x i )

(11)

(x i ) ≤ h i , (x i ) ≥ h i

u (x ) ≥ l (x ) Th e cor r es p on d i n g r el a ti on s b etween d u a l a p p r ox i m a ti on s a n d d i s tr i b u ti on s a r e l i s ted i n Ta b l e 1.

d u a l p os s i b i l i ty

3 0. Id en ti fy i n g Up p er a n d L ower P os s i b i l i ty D i s tr i b u ti on s

275

T a b l e 1 . Th e s i m i l a r i ti es b etween r ou g h s et a n d p os s i b i l i ty d i s tr i b u ti on s P os s i b i l i ty d i s tr i b u ti on s Up p er d i s tr i b u ti on : u

L ower d i s tr i b u ti on : P r od u ct of

(x i ) : l

m

∏

m l

i= 1 u

(x )

L ower a p p r ox i m a ti on : R

(x i )

Ca r d i n a l i ty of R

(x i )

Ca r d i n a l i ty of R

u

i= 1

*

*

(Z ) (Z )

*

*

(Z ) : C a r d ( R

( Z ))

m

∏

l

i= 1

In eq u a l i ty r el a ti on : Mea s u r e of fi tn es s :

β =

Up p er a p p r ox i m a ti on : R

m

∏

(x i ) : u

P r od u ct of

l

R ou g h s ets

(x )

u

(x ) ≥ l

(Z ) : C a r d ( R *

In cl u s i on r el a ti on : R * ( Z ) ⊇ R Accu r a cy m ea s u r e of a s et Z :

(x )

(x i ) (x i )

C a r d (R

α (Z ) =

( Z ))

(Z )

( Z )) * *

C a r d (R

*

*

( Z ))

It i s s tr a i g h tfor wa r d th a t th e ob j ecti ve fu n cti on a n d con s tr a i n ts of (11) cor r es p on d to th e ob j ecti ve fu n cti on a n d con s tr a i n ts of (10), r es p ecti vel y . Wi th con s i d er i n g

ln

m

∏

th a t m

π l

m a x im iz in g

m (x i ) = (∑ (l n π (x i ) i= 1

π u ca n b e r ewr i tten a s fol l ows : i= 1

m in a , D u

, D

∑ l

i= 1

D u

D

i

l

− D

i= 1 u

(x i ) − l n π

− 1 l

l

− 1 u

(x i

l

(x

is

eq u i va l en t

to

m a x im iz in g

(x i ))) / m , th e op ti m i z a ti on p r ob l em u

(x

− 1

− a ) D

(x i ) (x i ) l

t

− a )t D i

m

− a )t D i

(x

s . t.

(x

l

m

(x

∏ m

− a ) -∑ i

i

m

(x i= 1

i

− a )t D

− 1 u

(x i

− a )

(11)

(12)

− a ) ≥ − l n h i , i = 1, ..., m ,

− a ) ≤ − l n h i , i = 1, ..., m ,

≥ 0,

> 0.

It s h ou l d b e n oted th a t th e op ti m i z a ti on p r ob l em (12) i s eq u i va l en t to th e i n teg r a ted m od el p r op os ed i n th e p a p er [ 3 , 4 ] i n for m . However , th ey a r i s e fr om ver y d i ffer en t con s i d er a ti on . Th e l a tter wa s u s ed to i n teg r a te two op ti m i z a ti on p r ob l em s to ob ta i n u p p er a n d l ower p os s i b i l i ty d i s tr i b u ti on s s i m u l ta n eou s l y . Th e for m er i s u s ed to s eek a n op ti m a l cen ter vector a a n d p os i ti ve d efi n i te m a tr i ces D u a n d D l to m a x i m i z e fi tn es s m ea s u r e β d efi n ed i n for m u l a (4 ). Mod el (11) m a k es i t q u i te cl ea r th a t u p p er a n d l ower p os s i b i l i ty d i s tr i b u ti on s h a ve ver y s i m i l a r s tr u ctu r e to th e u p p er a n d l ower a p p r ox i m a ti on s i n r ou g h s ets th eor y . In th e fol l owi n g , l et u s con s i d er h ow to ob ta i n cen ter vector a a n d p os i ti ve m a tr i ces D l a n d D u .

276

P . G u o a n d H. Ta n a k a

Cen ter vector a ca n b e a p p r ox i m a tel y es ti m a ted a s a = x i* , wh er e x i

d en otes th e vector wh os e g r a d e i s h *

=

i*

(13 )

m a x

h

k = 1, ..., m

k

. Th e a s s oci a ted p os -

s i b i l i ty g r a d e of x i * i s r evi s ed to b e 1 b eca u s e i t b ecom es th e cen ter vector . Ta k i n g th e tr a n s for m a ti on y = x − a , th e p r ob l em (12) i s ch a n g ed i n to th e fol l owi n g p r ob l em .

m in D u

s . t.

, D

m

∑ l

i= 1

− 1

y ti D l

y

− 1

t

y iD u

D

− D u

− 1

y ti D l

y

-∑ i

m i= 1

− 1

y ti D u

(14 )

y i

≥ − l n h i , i = 1, ..., m , i

y

≤ − l n h i , i = 1, ..., m , i

≥ 0, l

D l > 0 Th e for m u l a (14 ) i s a n on l i n ea r op ti m i z a ti on p r ob l em d u e to th e l a s t two con s tr a i n ts . To cop e wi th th i s d i ffi cu l ty , we u s e p r i n ci p l e com p on en t a n a l y s i s (P C A ) to r ota te th e g i ven d a ta (y i , h i ) to ob ta i n a p os i ti ve d efi n i te m a tr i x ea s i l y . Th e d a ta y i

(i = 1, …

, m ) ca n b e tr a n s for m ed b y a l i n ea r tr a n s for m a ti on m a tr i x T wh os e

col u m n s a r e ei g en vector s of th e m a tr i x Σ = [ σ

σ Us i n g

{ z i

= T

∑

= {

ij

m

(x

k i

− a i )( x

j

)h k

∑

} /

k = 1

th e l i n ea r tr a n s for m a ti on t

− a

k j

] , wh er e σ

ij

m a tr i x

i s d efi n ed a s

ij

(15) m

h k

k = 1

y

T , th e d a ta

i

is

tr a n s for m ed

i n to

y i } . Th en for m u l a s (1) a n d (2) ca n b e r ewr i tten a s fol l ows :

π π

(z i ) = ex p { − z

u

l

(z i ) = ex p { − z t

Si n ce T i s ob ta i n ed b y P C A , T n a l m a tr i ces a s fol l ows :

D

t

t i − 1 u

t

T

i

t

T

T

⎜ u

= T t

D

− 1 u

T z i } , i = 1, … , m ,

(16)

T z i } , i = 1, … , m .

(17)

u

− 1

D l

t

a n d T

⎛ c

C

− 1

D

D

− 1 l

T

ca n b e a s s u m ed to b e d i a g o-

⎜

(18 )

0 ⎞ ⎟

u 1

⎟

.

T = ⎜ ⎜

⎜ 0 ⎝

⎟ ,

. c

⎟ u n

⎠

⎟

3 0. Id en ti fy i n g Up p er a n d L ower P os s i b i l i ty D i s tr i b u ti on s

⎛ c ⎜

C

= T l

t

− 1

D l

⎜

(19)

0 ⎞ ⎟

l1

⎟

.

T = ⎜ ⎜

⎜ 0 ⎝

277

⎟ .

. c

⎟ ln

⎟ ⎠

Mod el (14 ) ca n b e r ewr i tten a s th e fol l owi n g L P p r ob l em :

∑

m in

C l ,C u

m i= 1

z ti C

s . t.

z ti C

z u

t

*

a n d C

l *

D D

u * l

i= 1

z ti C u

z

(20) i

i

≥ − l n h i , i = 1, ..., m ,

u j

, j = 1, ..., n ,

lj

≥ c

u j

≥ ε , j = 1, ..., n , c

d efi n i te a n d m a tr i ces D u

m

≤ − l n h i , i = 1, ..., m i

wh er e th e con d i ti on *

-∑

l

c

C

i

z iC c

z

z l

lj

≥ c

u

a n d D

u j

≥ l

ε > 0 m a k es th e m a tr i x D

u

− D l

s em i -p os i ti ve

p os i ti ve. D en ote th e op ti m a l s ol u ti on s of (20) a s

. Th u s , we h a ve

= T C

* − 1 u

= T C

* − 1 l

t

T

T

t

,

(21)

.

3 0 .4 C o n c lu s io n s In th i s p a p er , fr om u p p er a n d l ower d i r ecti on s th e u p p er a n d l ower p os s i b i l i ty d i s tr i b u ti on s a r e i d en ti fi ed to a p p r ox i m a te th e g i ven p os s i b i l i ty g r a d es , wh i ch i s r eg a r d ed a s th e ex p er t’ s k n owl ed g e. Th e u p p er p os s i b i l i ty d i s tr i b u ti on r efl ects th e op ti m i s ti c vi ewp oi n t of th e ex p er t a n d th e l ower p os s i b i l i ty d i s tr i b u ti on r efl ects p es s i m i s ti c on e. Th e s i m i l a r i ti es b etween d u a l p os s i b i l i ty d i s tr i b u ti on s a n d u p p er a n d l ower a p p r ox i m a ti on s i n r ou g h s ets th eor y a r e i n ves ti g a ted . It i s ob vi ou s th a t th ey h a ve h om og en ou s s tr u ctu r es .

R e fe r e n c e s 3 0.1 Z . P a wl a k , R ou g h Sets (Kl u wer Aca d em i c P u b l i s h er s , N 3 0.2 Z . P a wl a k a n d R . Sl owi n s k i , R ou g h s et a p p r oa ch to m u E u r op ea n Jou r n a l of Op er a ti on a l R es ea r ch 7 2 (1994 ) 4 4 3 0.3 H. Ta n a k a a n d P . G u o, P os s i b i l i s ti c D a ta An a l y s i s for b er g ; N ew Yor k ; P h y s i ca -Ver l a g , F eb ., 1999). 3 0.4 H. Ta n a k a a n d P . G u o, P or tfol i o s el ecti on b a s ed on u p p s i b i l i ty d i s tr i b u ti on s , E u r op ea n Jou r n a l of Op er a ti on a l R

eth er l a n d s , 1991). l ti -a ttr i b u te d eci s i on a n a l y s i s , 3 -4 59. Op er a ti on s R es ea r ch (Hei d el er a n d l ower ex p on en ti a l p os es ea r ch 1 1 4 (1999) 115-126

31. On Fractals in Information Systems: The First Step Lech Polkowski1,2 1 2

Polish–Japanese Institute of Information Technology Koszykowa 86, 02-008 Warsaw, Poland Department of Mathematics and Information Sciences Warsaw University of Technology Pl. Politechniki 1,00-650 Warsaw, Poland [email protected]

We introduce the notion of a fractal in an information system and we define a dimension function of a fractal in an information system parallel to the Minkowski dimension in Euclidean spaces. We prove basic properties of this new dimension.

31.1 Introduction Objects called now ”fractals” have been investigated since 1920’s (cf. [31.3], [31.6]) yet the renewed interest in them goes back to 1970’s in connection with studies of chaotic behavior, irregular non–smooth sets, dynamic systems, information compression and computer graphics (cf. [31.9]). The basic characteristics of ”fractals” are rooted in dimension theory. The topological dimension theory assigns to any subset T of a (sufficiently regular) topological space X an integer dimT ≥ −1 called the dimension of T (cf. [31.7]). This dimension function, however, does not capture peculiar features of fractals among them the periodicity of local structure and appearance of details at any scale; for this reason, fractals are evaluated by means of other functions e.g. Hausdorff dimension or Minkowski (box) dimension better suited at capturing the peculiarities of local structure. Many fractal objects can be generated by means of iterations of affine mappings (iterated function systems (cf. [31.8]) hence they allow for knowledge compression algorithms (cf. [31.1]; cf. also [31.11] for a rough set counterpart of the fractal collage theorem). We are interested here in transferring the notion of a fractal to the general framework of rough set theory and we examine here some propositions for a counterpart of fractal dimension in this general framework.

31.2 Fractal Dimensions For a set (for properties of fractal dimensions see [31.4], [31.5]) T ⊆ E n ,  s and s ≥ 0, δ > 0, one lets Hδ (T ) = inf i diams (Qi ), the infimum taken T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 278 − 28 2, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

31. On Fractals in Information Systems: The First Step

279



over all families {Qi : i = 1, 2, ...} of sets in E n such that (i) T ⊆ i Qi (ii) diam(Qi ) ≤ δ. Then the limit H s (T ) = limδ→0+ Hδs (T ) exists and it follows easily that there exists a unique s∗ with the property that H s (T ) = ∞ for s < s∗ and H s (T ) = 0 for s > s∗ . The real number s∗ is the Hausdorff dimension of the set T , denoted dimH (T ). The Hausdorff dimension is too closely related to the metric structure of the underlying space to admit any substantial abstraction. For our purposes, the other function, the Minkowski dimension seems to be better suited. This dimension has an information theoretic content and may be transferred–with changes relaxing its geometric content–into a universe of a general information system. For a bounded set T ⊆ E n (i.e. diam(T ) < ∞), and δ > 0, we denote by nδ (T ) the least number of n–cubes of diameter less than δ that cover T . Then δ (T ) we may consider the fraction logn −logδ and evaluate its limit. When the limit δ (T ) limδ→0+ logn −logδ exists, it is called the Minkowski dimension of the set T and it is denoted dimM (T ). One may interpret this dimension as an information content (cf.[31.2]) of T : the shortest description of T over an alphabet of δ–cubes has length of order of dimM (T ). Both dimensions agree on ”standard” fractal objects like the Cantor set (cf. [31.4], [31.5]) in general they disagree. An advantage of the Minkowski dimension is that families of δ–cubes in its dimension may be selected in many ways, one among them is to consider a δ– grid of cubes of side length δ on E n and to count the number Nδ (T ) of those δ (T ) among them which intersect T ; then (cf. [31.4], [31.5]) if limδ→0+ logN −logδ exists it is equal to dimM (T ).

31.3 Rough Sets and Topologies on Rough Sets Rough sets arise in an attempt at formalization of the notion of uncertain knowledge (cf. [31.10], [31.13]). In this paradigm, knowledge base is an information system A = (U, A) where U is the set of objects described by means of attributes (features, properties) collected in the set A. For an object x ∈ U and an attribute a ∈ A we denote by the symbol a(x) the value of a on x. We admit here the case when the set U is infinite (e.g. E n ) and the set A consists of countably many attributes an where n = 1, 2, ... . Each attribute an induces on U the {an }–indiscernibility relation Indan viz. xIndan y ⇔ an (x) = an (y) which partitions U into classes [x]an ; Pn is the resulting partition. We may assume that Pn+1 ⊆ Pn for each n. A subset (concept) Z ⊆  U is n–exact in case it is a union of a family of classes of Indan i.e. Z = {[z]an : z ∈ Z}. Otherwise, Zis said to be n–rough. Rough sets  are approximated by exact n = {[x]an : [x]an ⊆ Z} and Zn+ = {[x]an : [x]an ∩ Z = ∅}. The sets : Z− n set Z− is the lower an –approximation of Z and the set Zn+ is the upper an – n , Zn+ as interior Int, approximation of Z. A topological interpretation of Z− resp. closure Cl of Z in topology Pn induced by the partition Pn suggests (cf.

280

L. Polkowski

[31.12]) a topology Π A on the set U by taking as an open base for this topology the family P = n Pn . A set Z is ΠA –exact in case IntΠA Z = ClΠA Z otherwise it is ΠA –rough. In this way, we define a taxonomy of sets in U : they may be divided into three classes: sets which are Πn –exact, sets which are ΠA –exact and sets which are ΠA –rough (for a detailed study of topologies on rough sets see [31.12]). We now consider an information system AC on the Euclidean space E n ; this system consists of the universe U = E n and of attributes ak for k = 1, 2, .... defined via partitions Pk induced by relations Indak . The partition Pk consists of n–cubes of the form (c)

n  i=1

[mi +

ji ji + 1 , mi + ) 2k 2k

where mi is an integer for each i = 1, 2, ..., n and 0 ≤ ji ≤ 2k − 1 is an integer. From the definition of the Minkowski dimension we have Proposition 1 If the Minkowski dimension dimM (T ) exists then dimM (T ) = logN  limk→∞ klog2k where Nk is the number of cubes in Pk which do intersect T . Proposition 2 For any ΠA –exact set Z, we have dimM (Z) = n. Proof. Indeed, if a set Z is ΠM –exact then Z is a union of a family {Qj : j = 1, 2, ...} of n–cubes of the form (c) and thus n ≥ dimM (T ) ≥ dimM (Q1 ) = n by the monotonicity and stability of dimM (cf. [31.5] Sect. 3.2 and Thm.3.4). Corollary 1 Any set Z of fractional dimension dimM is a ΠA –rough set. The last fact directs us towards general information systems and rough sets resulting in them.

31.4 Fractals in Information Systems For an information system A = (U, A) with the countable set A = {an : n = 1, 2, ...} of attributes such that Indan +1 ⊆ Indan for n = 1, 2, ..., we will define the notion of an A–dimension, denoted dimA . We will observe the information–theoretic content of the Minkowski dimension and thus– refraining from any geometric content, we introduce a normalization condition (N ) dimA (Q) = 1 for every equivalence class Q of any relation Indan . The condition (N ) assures us that any equivalence class carries with itself a single bit of information, thus playing a role of an alphabet symbol.

31. On Fractals in Information Systems: The First Step

281

We restrict ourselves to bounded subsets Z ⊆ U i.e. such Z which for each n are covered by a finite number of equivalence classes of Indan . We may therefore assume that (i) the number of equivalence classes of Inda1 is k1 (ii) each class of Indan ramifies into kn+1 classes of Indan+1 . We will say that the information system A is of type κ =(ki )i . log n l For a bounded set Z ⊆ U , we let dimA (Z) = limn→∞ log ni=1 kii where li i=1 is the number of classes of Indai that intersect Z i.e. the number of classes in the upper approximation Zi+ of Z. Then we have Proposition 3 In case A is of type κ with kj ≥ 2 for infinitely many j, dimA does satisfy (N ). Proof. Consider Q, a basic openset so that Q = [x]ak . Wehave  log

n

n

log

l

ki

log

k

k

i = 1−limn→∞ log i=1 = 1 where limn→∞ log ni=1 kii = limn→∞ log i=k+1 n n i=1 i=1 ki i=1 ki li is the number of Indai classes that intersect Q. Thus dimA (Q) = 1.

Let us observe that – as with the Minkowski dimension– the A – dimension may be ramified into two  weaker notions viz. the upper A – dimenlog n l  sion dimA = limsupn→∞ log ni=1 kii and the lower A–dimension dimA = log

n

i=1

l

liminfn→∞ log ni=1 kii . i=1 Basic properties of dimA parallel the respective properties of the Minkowski dimension. Proposition 4 dimA satisfies the following 1. dimA (Z) ≤ dimA (T ) whenever Z ⊆ T 2. dimA (Z ∪ T ) = max{dimA (Z), dimA (T )} in case A is of type κ with ki ≥ 2 for infinitely many i 3. dimA (Z) = dimA (ClΠA Z) Proof. Indeed, (i) follows by the very definition of dimA . For (ii), by (i) it follows that dimA (Z ∪ T ) ≥ max{dimA (Z), dimA (T )}. To prove the converse let us assume that dimA (Z) ≥ dimA (T ) and split infinite sequences of natural numbers into two classes (pj denotes the number of classes of Indaj intersecting Z and qj means the same for T ): (I) a sequence (nj )j falls here in case pnj < qnj for infinitely many j and pnj ≥ qnj for infinitely many j (II) a sequence falls here in case pnj < qnj for almost every j (III) a sequence falls here in case pnj ≥ qnj for almost every j. We assume that lj is the number of classes of Indaj intersecting Z ∪ T ; clearly lj ≤ pj + qj for each j. Now consider a sub–sequence nj for which log log

nj

i=1 n j i=1

li ki

converges. In case it falls into (II), we have lnj ≤ 2qnj for al-

most every j and thus limj→∞ dimA (Z).

log log

nj i=1 n j i=1

li ki

≤ limj→∞

log

nj

log

i=1 nj i=1

2qi ki

≤ dimA (T ) ≤

282

L. Polkowski

Similarly in case the sequence falls into (III), lnj ≤ 2pnj for almost every j and thus limj→∞

log log

nj

i=1 n j i=1

li ki

≤ dimA (Z).

In case the sequence is in (I), by its convergence we have lim   2qi log u 2pi log v i=1 , lim limu→∞ log i=1 u v v→∞ log k i=1 i i=1 ki

log log

nj i=1 n j

li

i=1 ki

≤

≤ max{dimA (Z), dimA (T )} = dimA (Z) where u, v run respectively over indices nj where pnj < qnj , pnj ≥ qnj . Finally, (iii) follows from the fact that Q ∩ ClΠA Z = ∅ if and only if Q ∩ Z = ∅ for every Q, a class of Indan , any n.

31.5 Conclusions We have examined the notion of a fractal in the universe of an information system, and we have defined the A–dimension proving its basic properties. Acknowledgement. This work has been supported by the Grant No. 8T11C 024 17 from the State Committee for Scientific Research (KBN) of the Republic of Poland.

References 31.1 M. F. Barnsley, Fractals Everywhere, Academic Press, 1988. 31.2 P. Billingsley, Ergodic Theory and Information, John Wiley, 1985. 31.3 C. Car´ atheodory, Ueber das lineare Mass von Punktmenge eine Verallgemeinerung des Langenbegriffs, Nach.Gessell.Wiss.G¨ ottingen, 1914, 406-426. 31.4 K. J. Falconer, The Geometry of Fractal Sets, Cambridge U. Press, 1990. 31.5 K. J. Falconer, Fractal Geometry. Mathematical Foundations and Applications, Wiley and Sons, 1990. 31.6 F. Hausdorff, Dimension und ausseres Mass, Math.Annalen, 79, 1919, 157179. 31.7 W. Hurewicz and H. Wallman, Dimension Theory, Princeton U. Press, 1941. 31.8 J. E. Hutchinson, Fractals and self–similarity, Indiana Math. Journal, 30, 1981, 713–747. 31.9 B. Mandelbrot, Les Objects Fractals: Forme, Hasard et Dimension, Flammarion, Paris, 1975. 31.10 Z. Pawlak, Rough Sets: Theoretical Aspects of Reasoning about Data, Kluwer, 1991. 31.11 L. Polkowski, Approximation mathematical morphology. Rough set approach, in: A. Skowron, S. K. Pal (eds.), Rough Fuzzy Hybridization. A New Trend in Decision–Making, Springer Verlag Singapore, 1999, 151-162. 31.12 L. Polkowski, Mathematical morphology of rough sets, Bull. Polish Acad. Sci. Math., 41, 1993, 241–273. 31.13 L. Polkowski, S. Tsumoto, and T. Y. Lin, (eds.), Rough Set Methods and Applications. New Developments in Knowledge Discovery in Information Systems, Physica Verlag/Springer Verlag, 2000.

32. Generalizations of Fuzzy Multisets for Including Infiniteness Sadaaki Miyamoto Institute of Engineering Mechanics and Systems University of Tsukuba, Ibaraki 305-8573, Japan [email protected]

This paper aims at discussing two generalizations of fuzzy multisets in order to take infinite features into account. First, a class of fuzzy multisets having an infinite membership set for an element of the universe and finite cardinality is introduced. The sum, union, intersection as well as most t-norm and conorm operations except the drastic sum keep the property of the finite cardinality of the derived set. Second, the membership sequence is generalized to a closed set on the plane whereby both the fuzzy multiset and another fuzzification of multisets using the fuzzy number are discussed within this framework.

32.1 Introduction Multisets, sometimes called bags, have been considered by many authors (e.g., [32.3, 32.1]) and used in a number of applications. Fuzzy multisets have also been considered by several researchers [32.6, 32.4]. An application of fuzzy multisets is information retrieval on Web, since an information item may appear more than once with possibly different degrees of relevance to a query [32.5]. This application invokes interesting problems. Huge amount, almost infinite, of information items exists in the space of WWW. A query may search a very large number of the items wherefrom all information is unable to be obtained by human capability. We thus observe a small part of the obtained information pieces. Such experiences lead us to consideration of infinite fuzzy multisets. The infiniteness implies that although the information pieces may be finite but the number of information items is very large and there is no fixed upper bound to this number. We are concerned with infinite fuzzy multisets in this paper. The infiniteness does not mean the universal space on which fuzzy multisets are discussed is infinite. It means that a membership set for an element the universe is infinite even when the underlying crisp multisets cannot have infinite multiplicity. We introduce a class of infinite fuzzy multisets for which the cardinality is finite, and shows that most t-norm and conorm operations for two sets in this class keep the derived set within this class. T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 28 3 − 28 8 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

284

S. Miyamoto

Second generalization of fuzzy multisets is moreover considered. There is another fuzzification of a multiset using the fuzzy number. This generalization essentially include the both fuzzifications using the membership sets and the fuzzy number in a unified framework.

32.2 Multisets and Fuzzy Multisets A multiset M of X is characterized by the count function CM : X → {0, 1, 2, . . . }. Thus, CM (x) is the number of copies of the element x ∈ X. The followings are basic relations and operations for crisp multisets; (inclusion): M ⊆ N ⇔ CM (x) ≤ CN (x), ∀x ∈ X. (equality): M = N ⇔ CM (x) = CN (x), ∀x ∈ X. (union): CM ∪N (x) = CM (x) ∨ CN (x). (intersection): CM ∩N (x) = CM (x) ∧ CN (x). (sum): CM +N (x) = CM (x) + CN (x). It is reasonable to assume that the number CM (·) should be finite. Moreover we assume X is finite: X = {x1 , . . . , xn }. A fuzzification of the multiset is to define CM (x) in terms of fuzzy numbers. We thus use the above definitions but CM (x) and CN (x) are assumed to be nonnegative fuzzy numbers. A fuzzy multiset A of X (more often called fuzzy bag) is characterized by the function CA (·) of the same symbol, but the value CA (x) is a finite set in I [32.6]. Given x ∈ X, CA (x) = {μ, μ , . . . , μ }, μ, μ , . . . , μ ∈ I. For two fuzzy multisets A and B of X such that CA (x) = {μ, μ , . . . , μ }, CB (x) = {ν, ν  , . . . , ν  }, the sum A + B is CA+B (x) = {μ, μ , . . . , μ , ν, ν  , . . . , ν  }, but other operations need another representation called membership sequence [32.4]. A membership sequence is defined for each CA (x) = {μ, μ , . . . , μ }; the set {μ, μ , . . . , μ } is arranged into the decreasing order denoted by 1 2 m   μ1A (x), μ2A (x), . . . , μm A (x): {μA (x), μA (x), . . . , μA (x)} = {μ, μ , . . . , μ } 1 2 m (μA (x) ≥ μA (x) ≥ · · · ≥ μA (x)). The followings are other basic relations and operations for fuzzy multisets [32.4]; they are given in terms of the membership sequences. 1. 2. 3. 4. 5.

inclusion: A ⊆ B ⇔ μjA (x) ≤ μjB (x), j = 1, . . . , m, ∀x ∈ X. equality: A = B ⇔ μjA (x) = μjB (x), j = 1, . . . , m, ∀x ∈ X. union: μjA∪B (x) = μjA (x) ∨ μjB (x), j = 1, . . . , m, ∀x ∈ X. intersection: μjA∩B (x) = μjA (x) ∧ μjB (x), j = 1, . . . , m, ∀x ∈ X. t-norm and conorm: μjATB (x) = t(μjA (x), μjB (x)), j = 1, . . . , m, ∀x ∈ X. μjASB (x) = s(μjA (x), μjB (x)), j = 1, . . . , m, ∀x ∈ X.

32. Generalizations of Fuzzy Multisets

285

Remark that there are different types of t-norms and conorms: we consider the algebraic product and sum, the bounded product and sum, the Frank family, the Hamacher family, the Yager family, the Sugeno family, and lastly the drastic product and sum [32.2]. All t-norms and conorms are denoted by single letters T and S except the last one; the drastic product and sum are denoted by TD and SD , respectively.

32.3 Infinite Memberships Even when crisp multisets cannot admit infinite values of the function CM (x), fuzzy multisets are capable of having infinite number of memberships. Remark that every infinite set does not provide a well-defined fuzzy multiset, since an α-cut of a fuzzy multiset should give a crisp multiset of the finite count. Instead of the finite set, infinite CA (x) = {μ, μ , . . . } is used. We assume that the members {μ, μ , . . . } of CA (x) can be arranged into the decreasing order: CA (x) = {μ1A (x), μ2A (x), . . . },

μ1A (x) ≥ μ2A (x) ≥ . . .

In order that the α-cuts provide well-defined crisp multisets, it is necessary and sufficient that μjA (x) → 0, as j → ∞, for all x ∈ X. This class of fuzzy multiset of X is denoted by FM0 (X). The operations such as A + B, A ∪ B, etc. are defined in the same way as above except that m → ∞ in the definitions. We have Proposition 1. For A, B ∈ FM0 (X), A+B ∈ FM0 (X), A∪B ∈ FM0 (X), A ∩ B ∈ FM0 (X), ATB ∈ FM0 (X), ASB ∈ FM0 (X), except the drastic sum: ASD B ∈ FM0 (X) does not necessarily hold.  A basic measure of a fuzzy set F is its cardinality defined by |F | = x∈X μF (x). When a fuzzy multiset A of finite membership sets is considered, its generalization is immediate: |A| =

m 

μjA (x).

x∈X j=1

Let us consider the cardinality for the infinite memberships. We define |A|x =

∞ 

μjA (x).

(32.1)

j=1

 Then, |A| = x∈X |A|x . It is easy to see that |A| is finite if and only if |A|x is finite for all x ∈ X, since we are considering finite X.

286

S. Miyamoto

Note that for some sets, say B ∈ FM0 (X), |B|x = +∞. (Consider μjB (x) = 1/j.) We hence introduce a subclass FM1 (X) for which the cardinality is finite: FM1 (X) = {A ∈ FM0 (X) : |A|x < ∞, ∀x ∈ X}.

(32.2)

We now have the following proposition. Proposition 2. For arbitrary A, B ∈ FM1 (X), A + B ∈ FM1 (X), A ∪ B ∈ FM1 (X), A ∩ B ∈ FM1 (X), ATB ∈ FM1 (X), ASB ∈ FM1 (X), except the drastic sum: ASD B is not necessarily in FM1 (X). It should be noted that most, but not all, t-conorms keep the derived sets within FM1 .

32.4 A Set-Valued Multiset It seems that nothing is in common between fuzzy multisets and fuzzification by fuzzy numbers. On the contrary, there is a generalized framework in which the two kinds of fuzzified multisets are put. Let us notice that the membership sequence, whether it is finite or infinite, is regarded as a nonincreasing step function. In view of this, we first consider a monotone nonincreasing function ζA (y; x) of the variable y ∈ [0, +∞) with the values in [0, +∞) for every x ∈ X as a parameter. Moreover the function is assumed to satisfy ζA (y; x) → 0 as y → ∞. Even if we do not assume any kind of continuity, it is well-known that the function ζA (y; x) is continuous almost everywhere due to the monotone property. We moreover assume, for the next step, that the function is upper-semicontinous. Second, this function ζA (y; x) is transformed to a closed set νA (y, z; x) on the (y, z)-plane; we use the set νA (·, ·; x) as the membership for the generalized fuzzy multiset. This set is defined by νA (y, z; x) = {(y, z) ∈ [0, ∞)2 : ζA (y; x) ≥ z}. Another function ηA (z; x) with the variable z derived from νA is moreover defined: ηA (z; x) = sup{y ∈ νA (y, z; x)},

(z ∈ (0, ∞)).

It is evident that if we define  νA (z, y; x) = {(y, z) ∈ [0, ∞) × (0, ∞) : ηA (z; x) ≥ y}

∪ {(y, 0) : y ∈ [0, ∞)},

(32.3)

 then νA (y, z; x) = νA (z, y; x). The generalized fuzzy multiset A is characterized by νA (y, z; x). For two generalized fuzzy multisets A and B of X, the basic relations and operations are defined by the operations on the sets νA and νB .

32. Generalizations of Fuzzy Multisets

287

(I) (inclusion). A ⊆ B ⇔ νA (·, ·; x) ⊆ νB (·, ·; x), ∀x ∈ X. (II) (equality). A = B ⇔ νA (·, ·; x) = νB (·, ·; x), ∀x ∈ X.  from (III) (sum). Define ηA+B (z; x) = ηA (z; x) + ηB (z; x) and derive νA+B ηA+B (z; x) using (32.3).  (z, y; x). Define νA+B by νA+B (y, z; x) = νA+B (IV) (union). Define ζA∪B (y; x) = ζA (y; x) ∨ ζB (y; x) and derive νA∪B from ζA∪B (y; x). (V) (intersection). Define ζA∩B (y; x) = ζA (y; x) ∧ ζB (y; x) and derive νA∩B from ζA∩B (y; x). (VI) (t-norm and conorm). Define ζATB (y; x) = t(ζA (y; x), ζB (y; x)); ζASB (y; x) = s(ζA (y; x), ζB (y; x)) and derive νATB and νASB from ζATB (y; x) and ζASB (y; x), respectively. It is evident that this generalization includes the fuzzy multisets and positive real-valued multisets [32.1], whereas it is not obvious that this also includes the fuzzification by the fuzzy number. A simple mapping from the class of CA (x) as a fuzzy number to ζA (·; x) is used for showing this generalization encompasses the fuzzification of multisets using fuzzy numbers. Notice that CA (x), a fuzzy number, consists of two upper-semicontinuous functions L(y) and R(y): CA (x) = L(y), (0 ≤ y ≤ c) and CA (x) = R(y), (c ≤ y), where L(c) = R(c) = 1. ˜ First, L(y) is transformed into a lower-semicontinuous function L(y) which is equal to L(y) on all continuity points. Then CA (x) is mapped to ζA (·; x) by the next rule: ⎧ ⎪ ˜ ⎨1 − 1 L(y), 0 ≤ y ≤ c, 2 ζA (y; x) = 1 ⎪ ⎩ R(y), c < y. 2 It is immediate to see that the inclusion and equality as well as the operations of the sum, union, and intersection for the fuzzification by the fuzzy number is expressed in terms of the present generalization by the above mapping.

32.5 Conclusion We have discussed two generalizations which include infinite features in fuzzy multisets. In the first generalization a subclass of finite cardinality has been introduced and it has been shown that the standard set operations are performed within this class, whereas an exceptional t-conorm of the drastic sum may put the derived set out of this class. More general results will be expected about t-conorms. In the second generalization two fuzzifications of the crisp multiset are considered in the unified framework. When compared with the first generalization, the latter is more general.

288

S. Miyamoto

Multisets have close relationships with rough sets and their generalizations [32.7]. Theoretical aspects of fuzzy multisets in relation to rough sets should further be considered. We have suggested application of infinite fuzzy multisets to information retrieval on WWW. More efforts should be concentrated on such applications as future studies.

References 32.1 W.Blizard, Real-valued multisets and fuzzy sets, Fuzzy Sets and Systems, Vol. 33, pp. 77–97, 1989. 32.2 A.di Nola, S.Sessa, W.Pedrycz, E.Sanchez, Fuzzy Relation Equations and Their Applications to Knowledge Engineering, Kluwer, Dordrecht, 1989. 32.3 Z.Manna, R.Waldinger, The Logical Basis for Computer Programming, Vol. 1: Deductive Reasoning, Addison-Wesley, 1985. 32.4 S.Miyamoto, Basic operations of fuzzy multisets, J. of Japan Soc. for Fuzzy Theory and Systems, Vol.8, pp.639–645,1996(in Japanese). 32.5 S.Miyamoto, Rough sets and multisets in a model of information retrieval, in F.Crestani et al. eds., Soft Computing in Information Retrieval: Techniques and Applications, Springer, pp.373–393,2000. 32.6 R.R.Yager, On the theory of bags, Int. J. General Systems, Vol. 13, pp. 23–37, 1986. 32.7 Y.Y.Yao, S.K.M.Wong, T.Y.Lin, A review of rough set models, in T.Y.Lin, N.Cercone, eds., Rough Sets and Data Mining: Analysis of Imprecise Data, Kluwer, Boston, pp.47–75,1997.

33. Fuzzy c-Means and Mixture Distribution Model for Clustering Based on L1 -Space Takatsugu Koga1 , Sadaaki Miyamoto2 , and Osamu Takata3 1

2 3

Graduate School of Systems and Information Engineering University of Tsukuba, Ibaraki 305-8573, Japan [email protected] Institute of Engineering Mechanics and Systems, University of Tsukuba Doctoral Program in Engineering, University of Tsukuba

This paper aims at proposing and comparing two fuzzy models and a statistical model for clustering based on L1 -space. Clustering methods in the fuzzy models are the standard fuzzy c-means and an entropy regularization method based on L1 -space. Furthermore, we add new variables to them for improving the cluster division. In the statistical model, a mixture distribution model based on L1 -space is proposed and the EM algorithm is applied.

33.1 Introduction A characteristic of methods of data clustering is that various measures of distance and similarity between objects can be employed [33.1, 33.5]. For example, the L1 space, instead of the most known Euclidean space, is sometimes useful in crisp and fuzzy c-means. Several results have been published in fuzzy c-means based on the L1 space [33.3, 33.7, 33.9], and studies are ongoing in order to improve the method and to investigate the properties of the clusters theoretically. For example, the method of entropy regularization and fuzzy classification functions [33.9, 33.8] should be studied; additional variables for clustering can be taken into account [33.6]. The aim of the present paper is to include new variables into the methods of the standard fuzzy c-means [33.2] and the entropy fuzzy c-means [33.10] based on the L1 -metric. In addition, a new mixture distribution model on the L1 -space is proposed in which the EM algorithm [33.4, 33.11] is used to estimate parameters.

33.2 Fuzzy c-Means Based on L1 -Space Assume that the p-dimensional space Rp is equipped with the weighted L1 norm: for x = (x1 , . . . , xp ) and y = (y 1 , . . . , y p ) in Rp , x − y =

p 

wj |xj − y j |,

j=1

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 28 9− 294 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

290

T. Koga, S. Miyamoto, and O. Takata

where (w1 , . . . , wp ) is the weight vector. A set X = {x1 , . . . , xn } of objects xk = (x1k , . . . , xpk ) ∈ Rp should be divided into c clusters. Clusters are denoted by Gi (i = 1, . . . , c) or simply i. Center for cluster i is denoted by vi = (vi1 , . . . , vip ); we write V = (v1 , . . . , vc ) for simplicity. The membership matrix is U = (uik ); uik is the degree of membership of xk to cluster i. The method of fuzzy c-means uses an alternative minimization of an objective function J(U, V ). In addition to U and V , we use more variables α = (α1 , . . . , αc ) for controlling the sizes of clusters and η = (ηij ) for controlling the scatters of them. We consider the following two objective functions. m  p n  c   uik αi ηij |xjk − vij | Jstd (U, V, α, η) = α i i=1 j=1 k=1

Jent (U, V, α, η) =

n c   i=1 k=1

uik

p  j=1

ηij |xjk − vij | + λ−1

n c   i=1 k=1

uik log

uik αi

The subscript std and ent imply that the methods are standard fuzzy cmeans [33.2] and the method of entropy regularization [33.10, 33.8], respectively. Each function has its parameter: m(> 1) in Jstd and λ(> 0) in Jent . The constraints for U , α, and η are   c M = (uik ) | uik ∈ [0, 1], i=1 uik = 1, k = 1, . . . , n ,   c A = (αi ) | αi ∈ [0, 1], i=1 αi = 1 ,   p H = (ηij ) | ηij > 0, j=1 ηij = 1, i = 1, . . . , c . The next alternative optimization algorithm FCM is used for clustering in which J = Jent or J = Jstd . Algorithm FCM. FCM1. Set initial values V¯ and α ¯ C¯ η. ¯. FCM2. Solve min J(U, V¯ , α ¯ , η¯) and let the solution be U U ∈M ¯ , V, α FCM3. Solve min J(U ¯ , η¯) and let the solution be V¯ . V

FCM4. Solve FCM5. Solve

¯ , V¯ , α, η¯) and let the solution be α. min J(U ¯ α∈A ¯ , V¯ , α min J(U ¯ , η) and let the solution be η¯.

η∈H

¯ , V¯ , α FCM6. If the solution (U ¯ , η¯) is convergent,stop; otherwise go to FCM2. The optimal solutions of U , α, and η for J = Jent and J = Jstd are as follows. For the cluster centers V , we do not have a closed formula. Instead, an efficient p algorithm can be employed. For simplicity we put Dik = j=1 ηij |xjk − vij |.

33. Fuzzy c-Means and Mixture Distribution Model

291

(i) J = Jstd :  c  c  1 −1  1 −1 n m  α  Dik  m−1  m k=1 (uk ) Dk  , αi = , uik = n mD αi Dk (u ) ik ik k=1 =1

⎡ ηij = ⎣

p

=1

=1



n (uik )m |xjk k=1 n m  k=1 (uik ) |xk

− vij | − vi |

1 ⎤−1 p

⎦

.

(ii) J = Jent :

−1  p  n n j j

1 αi e−λDik 1 j k=1 uik |xk − vi | p n uik = c , αi = uik , ηi = .    n k=1 uik |xk − vi | −λDk k=1 =1 αl e =1

Calculation of V (cf. [33.9]). First, xj1 , xj2 , . . . , xjn−1 , xjn are sorted into the increasing order.

xjq(1)

xj1 , xj2 , . . . , xjn−1 , xjn ↓ SORT ≤ xjq(2) ≤ · · · ≤ xjq(n−1) ≤ xjq(n)

Algorithm C: begin n 1 (¯ uik )m ; S := − 2 k=1 r := 0; while (S < 0) do begin r := r + 1; S := S + (¯ uiq(r) )m ; end; output v¯ij = xjq(r) end. uik )m should be replaced by This algorithm is for V in Jstd ; For Jent , (¯ u ¯ik . Notice that this algorithm is very fast, since the computation of O(np) is sufficient in the main loop of iteration except the initial sorting.

33.3 Mixture Distribution Based on L1 -Space Mixture distribution model can be used for clustering [33.5, 33.11]. Our purpose is to develop a mixture distribution model for L1 -space, in contrast to the Gaussian mixture model for the Euclidean space. Three elements are used in clustering by a mixture distribution.

292

T. Koga, S. Miyamoto, and O. Takata

(i)

the prior probability of occurrence of the cluster Gi :P (Gi ) = αi , x (ii) the conditional probability of x given Gi : P (x|Gi ) = pi (x|φi ), −∞

(iii) the probability P (Gi |x) by which an observation x is allotted to Gi . Notice the Bayes formula: P (Gi |x) =

αi pi (x|φi ) P (Gi )P (x|Gi ) = c . c   P (Gj )P (x|Gj ) αj pj (x|φj ) j=1

(33.1)

j=1

We must assume the density pi (x|φi ) and estimate the parameters αi and φi (i = 1, . . . , c). Since the Gaussian distribution cannot be used in L1 -space, we assume the following density function:

pi (x|φi ) = pi (x|μi , νi ) =

p  j  ν i

j=1

2

j

e−νi |x

j

−μji |



where the parameter φi = (μ1i , . . . , μpi , νi1 , . . . , νip ) is 2p-dimensional vector. In order to estimate the vector parameter Φ = (α1 , . . . , αc , φ1 , . . . , φc ), the EM algorithm [33.4, 33.11] is used. Let Q(Φ|Φ( ) ) =

c 

( )

Ψi

log αi +

i=1

c  n 

( )

ψik log pi (xk |φi ).

i=1 k=1

The EM algorithm. (O) Set initial value of Φ(0) for the parameter Φ. Put  = 0. Repeat (E) and (M) until convergence. (E) Calculate Q(Φ|Φ( ) ). ¯ (M) Solve max Q(Φ|Φ( ) )and let the optimal solution be Φ. Φ ¯ Put  =  + 1 and Φ( ) = Φ. End EM. The solution in the step (M) is given as follows. Put ( )

( )

ψik =

( )

( )

ψ αi pi (xk |φi ) ( ) , wik = n ik . c   ( ) ( ) ( ) αj pj (xk |φj ) ψik k =1

j=1

1  ( ) ψik n n

Optimal αi : αi =

k=1

Calculation of μji .

(i = 1, . . . , c).

33. Fuzzy c-Means and Mixture Distribution Model

293

This algorithm is essentially the same as the former algorithm for calculating the cluster centers. First the sorting (33.2) is performed. The algorithm C in the previous section is then applied with the obvious replacement of (¯ uik )m ( ) ( ) and (¯ uiq(r) )m into wik and wiq(r) , respectively. Lastly, νij is obtained in terms of the optimal μ ¯ji : νij =

1 n 

( ) wik |xjk

. −

μ ¯ji |

k=1

33.4 Conclusion L1 -based methods of the standard and entropy fuzzy c-means with additional variables of the sizes and scatters of the clusters as well as the mixture distribution model have been proposed and algorithms have been developed. In the mixture distribution model, it has been shown that the EM algorithm is employed. Future studies include application to real data, in particular data mining applications are promising, since binary and nominal data should be dealt with, which means that L1 -space is a suitable framework.

References 33.1 M. R. Anderberg, Cluster Analysis for Applications, Academic Press, New York, 1973. 33.2 J. C. Bezdek, Pattern Recognition with Fuzzy Objective Function Algorithms, Plenum, New York, 1981. 33.3 L. Bobrowski and J. C. Bezdek, c-means clustering with the 1 and ∞ norms, IEEE Trans. on Syst., Man, and Cybern., Vol. 21, No. 3, pp. 545–554, 1991. 33.4 A. P. Dempster, N. M. Laird, and D. B. Rubin, Maximum likelihood from incomplete data via the EM algorithm, J. of the Royal Statistical Society, B., Vol.39, pp. 1–38, 1977. 33.5 B. S. Everitt, Cluster Analysis, 3rd ed., Arnold, London, 1993. 33.6 H. Ichihashi, K. Honda, N. Tani, Gaussian mixture PDF approximation and fuzzy c-means clustering with entropy regularization, Proc. of the 4th Asian Fuzzy System Symposium, May 31-June 3, 2000, Tsukuba, Japan, pp.217– 221. 33.7 K. Jajuga, L1 -norm based fuzzy clustering, Fuzzy Sets and Systems, Vol. 39, pp. 43–50, 1991. 33.8 Z.-Q. Liu, S.Miyamoto (Eds.), Soft Computing and Human-Centered Machines, Springer, Tokyo, 2000. 33.9 S. Miyamoto and Y. Agusta, An efficient algorithm for 1 fuzzy c-means and its termination, Control and Cybernetics Vol. 24, No.4, pp. 421–436, 1995.

294

T. Koga, S. Miyamoto, and O. Takata

33.10 S. Miyamoto and M. Mukaidono, Fuzzy c - means as a regularization and maximum entropy approach, Proc. of the 7th International Fuzzy Systems Association World Congress (IFSA’97), June 25-30, 1997, Prague, Chech, Vol.II, pp. 86–92, 1997. 33.11 R. A. Redner and H. F. Walker, Mixture densities, maximum likelihood and the EM algorithm, SIAM Review, Vol.26, No.2, pp. 195–239, 1984.

34. On Rough Sets under Generalized Equivalence Relations Masahiro Inuiguchi and Tetsuzo Tanino Department of Electronics and Information Systems Graduate School of Engineering, Osaka University 2-1, Yamada-Oka, Suita, Osaka 565-0871, Japan {inuiguti, tanino}@eie.eng.osaka-u.ac.jp http://vanilla.eie.eng.osaka-u.ac.jp

We consider two generalized situations: a case when an equivalence relation is generalized to a similarity relation and a case when a partition is generalized to a cover. Two interpretations of rough sets, i.e., the approximation by means of elementary sets and the distinction among positive, negative and boundary regions, are conceivable. The relations between two generalized situations are investigated. Rough sets are generalized based on two different interpretations under two different situations. Fundamental properties and complete definability are discussed in each generalization.

34.1 Introduction Rough sets were originally proposed in the presence of an equivalence relation. An equivalence relation is sometimes difficult to be obtained in realworld problems due to the vagueness and incompleteness of human knowledge. From this point of view, the concept of rough sets has been extended to cases when a similarity relation and a fuzzy partition are given (see [34.1]-[34.4]). However we have different definitions of rough sets even under the same generalized equivalence relation. Those different definitions coincide when the generalized equivalence relation degenerate to an equivalence relation. In spite of this difference, the reason has not discussed considerably, so far. In this paper, we demonstrate that there are two interpretations of rough sets and two generalized problem settings. In crisp cases, one of the two generalized settings is a situation that a similarity relation instead of an equivalence relation is given and the other is a situation that a cover instead of the partition associated with an equivalence relation is given. Rough sets composed of lower and upper approximations are interpreted in two different ways: distinction among positive, negative and boundary elements of a given subset and approximation of a given subset by means of elementary sets obtained from a similarity relation or a cover. Restricting ourselves into crisp cases, we discuss the relations between two different settings, how definitions of rough sets are different depending on the interpretation, fundamental properties of rough sets under those interpretations and the complete definability of rough sets.

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 295− 3 00, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

296

M. Inuiguchi and T. Tanino

Table 34.1. Fundamental properties of rough sets (i) (ii) (iii) (iv) (v) (vi) (vii)

R∗ (X) ⊆ X ⊆ R∗ (X) R∗ (∅) = R∗ (∅) = ∅, R∗ (U ) = R∗ (U ) = U R∗ (X ∩ Y ) = R∗ (X) ∩ R∗ (Y ), R∗ (X ∪ Y ) = R∗ (X) ∪ R∗ (Y ) X ⊆ Y implies R∗ (X) ⊆ R∗ (Y ), X ⊆ Y implies R∗ (X) ⊆ R∗ (Y ) R∗ (X ∪ Y ) ⊇ R∗ (X) ∪ R∗ (Y ), R∗ (X ∩ Y ) ⊆ R∗ (X) ∩ R∗ (Y ) R∗ (U − X) = U − R∗ (X), R∗ (U − X) = U − R∗ (X) R∗ (R∗ (X)) = R∗ (R∗ (X)) = R∗ (X), R∗ (R∗ (X)) = R∗ (R∗ (X)) = R∗ (X)

34.2 The Original Rough Sets Let R be an equivalence relation in the finite universe U . In rough set literature, R is referred to as an indiscernibility relation and a pair (U, R) is called an approximation space. By the equivalence relation R, U can be partitioned into a collection of elementary sets, U |R = {E1 , E2 , . . . , En }. Define R(x) as R(x) = {y ∈ U | yRx}. Then we have x ∈ Ei if and only if Ei = R(x). In rough sets, we consider the approximations of an arbitrary set X ⊆ U by means of elementary sets. Then the rough set of X is defined by a pair of the following lower and upper approximations: R∗ (X) = {x ∈ U | R(x) ⊆ X},

R∗ (X) = {x ∈ U | R(x) ∩ X = ∅}. (34.1)

By the definition, R∗ (X) ⊆ R∗ (X) holds. If R∗ (X) = R∗ (X) holds then X is said to be completely definable under the approximation space (X, U ). Under indiscernible circumstances given by (U, A), we cannot recognize the difference among elements in Ei but between x ∈ Ei and y ∈ Ej (i = j). Thus, what we can specify is not a particular element x of U but a particular elementary set Ei of U |R. Consider an element to which we know only it is in Ei . If Ei ⊆ R∗ (X), we can conclude that the element belongs to X. If Ei ⊆ U − R∗ (X), we can conclude that the element does not belong to X. From those facts, R∗ (X) and U − R∗ (X) are regarded as the positive and negative regions of X, respectively. R∗ (X) − R∗ (X) is regarded as the ambiguous region. The fundamental properties of R∗ (X) and R∗ (X) are listed in Table 34.1. Let ⎧ ⎨ R(x), if ∃R(x); R(x) ⊆ X, R∗1 (X) = R(x)⊆X (34.2) ⎩ ∅, otherwise, ⎧  ⎨ (U − R(x)), if ∃R(x); R(x) ∩ (U − X) = ∅, (34.3) R∗2 (X) = R(x)∩(U −X)=∅ ⎩ U, otherwise,

34. On Rough Sets under Generalized Equivalence Relations

R1∗ (X) =

R2∗ (X) =

⎧ ⎨



⎩ ⎧ ⎨

R(x)∩X=∅

⎩

R(x)∩X=∅

(U − R(x)), if ∃R(x); R(x) ∩ X = ∅, U,

(34.4)

otherwise,

R(x), if ∃R(x); R(x) ∩ X = ∅, ∅,

297

(34.5)

otherwise.

Since R is an equivalent relation, we have R∗ (X) = R∗1 (X) = R∗2 (X) and R∗ (X) = R1∗ (X) = R2∗ (X).

34.3 Two Different Problem Settings An equivalence relation R is identified by a partition U |R = {E1 , E2 , . . . , En } and vice versa. From this fact, there are two possible generalization schemes: generalization of R and generalization of U |R (see [34.2]). Generalization of R is to drop and/or to weaken some of the requirements of R so that R can be considered the so-called similarity relation, i.e., xRy means ‘x is similar to y’. Until now R is generalized up to a relation which satisfies only the reflexivity (see [34.4]). On the other hand, generalization of U |R nis to give a cover of U , i.e., a class F = {F1 , F2 , . . . , Fn } such that U = i=1 Fi (see [34.1]). Let us discuss relations between those generalizations. First consider a case that a similarity relation R is given. When R is no longer symmetric, a set of elements similar to x, i.e., R(x) is distinct from R−1 (x) = {y | xRy} that is a set of elements to which x is similar [34.4]. If R is reflexive, we can have a cover F = {R(x) | x ∈ U }. Since a similarity relation R should satisfy the reflexivity, the situation with R is a special case of the situation with a cover F. On the other hand, when a cover F = {F1 , F2 , . . . , Fn } is given, we face a problem how we can produce a similarity relation R such that F = {R(x) | x ∈ U }. Only if there is a unique Fi such that x ∈ Fi for any x ∈ U , we can solve this problem. However, this case is nothing but a case when F is a partition. Thus, there is no R satisfies F = {R(x) | x ∈ U } whenever F is not a partition. Hence, under a finite universe U , a problem setting with a cover F seems to be more general than that with a similarity relation R. This is true in an interpretation of rough sets as approximations by means of elementary sets. However, each elementary set Fi of F is not associated with an element x ∈ U . Because of this fact, we cannot always say that a cover F is more general than a similarity relation R. Finally, we should note that R∗ , R∗1 and R∗2 are no longer equivalent in both generalized settings. Neither R∗ , R1∗ nor R2∗ are. We have R∗ (X) = U − R∗ (U − X), Ri∗ (X) = U − R∗i (U − X), i = 1, 2 and, under the reflexivity of R, we obtain R∗2 (X) ⊆ R∗ (X) ⊆ R∗1 (X) and R1∗ (X) ⊆ R∗ (X) ⊆ R2∗ (X).

298

M. Inuiguchi and T. Tanino

Table 34.2. Fundamental properties of F∗1 (X) and F1∗ (X) (i) (ii) (iii) (iv) (v) (vi) (vii)

F∗1 (X) ⊆ X ⊆ F1∗ (X) F∗1 (∅) = F1∗ (∅) = ∅, F∗1 (U ) = F1∗ (U ) = U F∗1 (X ∩ Y ) ⊆ F∗1 (X) ∩ F∗1 (Y ), F1∗ (X ∪ Y ) ⊇ F1∗ (X) ∪ F1∗ (Y ) X ⊆ Y implies F∗1 (X) ⊆ F∗1 (Y ), X ⊆ Y implies F1∗ (X) ⊆ F1∗ (Y ) F∗1 (X ∪ Y ) ⊇ F∗1 (X) ∪ F∗1 (Y ), F1∗ (X ∩ Y ) ⊆ F1∗ (X) ∩ F1∗ (Y ) F∗1 (U − X) = U − F1∗ (X), F1∗ (U − X) = U − F∗1 (X) F∗1 (F∗1 (X)) = F∗1 (X), F∗1 (X) ⊆ F1∗ (F∗1 (X)) ⊆ F1∗ (X), F1∗ (F1∗ (X)) = F1∗ (X), F1∗ (X) ⊇ F∗1 (F1∗ (X)) ⊇ F∗1 (X)

34.4 Approximation by Means of Elementary Sets In interpretation of rough sets as approximations of sets by means of elementary sets, we assume a general setting, i.e., a case when a cover F = {F1 , F2 , . . . , Fn } is given. In this case, we should consider F∗1 (X), F∗2 (X), F1∗ (X) and F2∗ (X) defined by (34.2)–(34.5) substituting Fi for R(x), respectively. We can prove F∗2 (X) ⊆ F∗1 (X) ⊆ X and X ⊆ F1∗ (X) ⊆ F2∗ (X). Hence, 1 F∗ (X) and F1∗ (X) are better lower and upper approximations of X. Thus, we define a rough set of X under F by a pair of F∗1 (X) and F1∗ (X). For F∗1 (X) and F1∗ (X), we have fundamental properties listed in Table 34.2. By the lack of disjointedness between Fi and Fj (i = j), none of F∗1 (X ∩ Y ) ⊇ F∗1 (X) ∩ F∗1 (Y ), F1∗ (X ∪ Y ) ⊆ F1∗ (X) ∪ F1∗ (Y ), F∗1 (X) ⊇ F1∗ (F∗1 (X)) and F1∗ (X) ⊇ F∗1 (F1∗ (X)) always holds. Complete definability of X in the setting where F is given can be defined as (a) (b) (c)

X is F-inner completely definable if and only if F∗1 (X) = X is satisfied. X is F-outer completely definable if and only if F1∗ (X) = X is satisfied. X is F-completely definable if and only if X is F-inner completely definable and at the same time F-outer completely definable.

34.5 Distinction among Three Regions Let X be a set corresponding to a vague concept. Then the elements of X are not always agreed by all people. A given set X includes elements on whose memberships all people agree and also elements on whose memberships some people argue. Elements of X can be divided into unquestionable and questionable members. In such a case, rough sets can be applied to classify elements into three categories: positive members, negative members and boundary members. Let X and X be sets of positive members and possible members, respectively. Here ‘possible members’ are composed of positive and boundary

34. On Rough Sets under Generalized Equivalence Relations

299

Table 34.3. Fundamental properties of R∗3 (X) and R3∗ (X) (i) (ii) (iii) (iv) (v) (vi) (vii)

R∗3 (X) ⊆ X ⊆ R3∗ (X) R∗3 (∅) = R3∗ (∅) = ∅, R∗3 (U ) = R3∗ (U ) = U R∗3 (X ∩ Y ) = R∗3 (X) ∩ R∗3 (Y ), R3∗ (X ∪ Y ) = R3∗ (X) ∪ R3∗ (Y ) X ⊆ Y implies R∗3 (X) ⊆ R∗3 (Y ), X ⊆ Y implies R3∗ (X) ⊆ R3∗ (Y ) R∗3 (X ∪ Y ) ⊇ R∗3 (X) ∪ R∗3 (Y ), R3∗ (X ∩ Y ) ⊆ R3∗ (X) ∩ R3∗ (Y ) R∗3 (U − X) = U − R3∗ (X), R3∗ (U − X) = U − R∗3 (X) R3∗ (R∗3 (X)) ⊆ X does not always hold, R∗3 (R3∗ (X)) ⊇ X does not always hold.

members. A given X should satisfy X ⊆ X ⊆ X. We assume that only elements which are similar to a member of X can be regarded as possible members. Then we have

X⊆ R(y) = {x | R−1 (x) ∩ X = ∅}. (34.6) y∈X

Since U − X = (U − X) and (U − X) = U − X, we also have

R(y) = {x | R−1 (x) ⊆ X}. X⊇U−

(34.7)

y∈X

In our problem setting, we know X such that X ⊆ X ⊆ X, only. We obtain a lower approximation of X and a upper approximation of X as follows: R∗3 (X) = {x | R−1 (x) ⊆ X},

R3∗ (X) = {x | R−1 (x) ∩ X = ∅}.

(34.8)

The fundamental properties of R∗3 (X) and R3∗ (X) are listed in Table 34.3. By the interpretation of lower and upper approximations, R∗3 (R∗3 (X)) and R3∗ (R3∗ (X)) are nonsense. Comparing to Table 34.2, property (iii) is preserved in Table 34.3. The preservation of property (vii) in Table 34.3 is worse than that in Table 34.2. From (34.6) and (34.7), a family of consistent lower regions is given as X = {X | R∗3 (X) ⊆ X ⊆ X ⊆ R3∗ (X)}. Similarly, a family of consistent upper regions is given as X = {X | R∗3 (X) ⊆ X ⊆ X ⊆ R3∗ (X)}. From (34.6) and (34.7) again, X ⊆ R3∗ (X) and R∗3 (X) ⊆ X should be satisfied. Thus, a family of consistent pairs of positive and possible regions is obtained as C = {(X, X) | X ∈ X , X ∈ X , X ⊆ R3∗ (X), R∗3 (X) ⊆ X}. Note that (X, X) ∈ C always holds. We can define the definiteness of X under a similarity relation R as follows: (d) X is said to be R-definite if and only if C is a singleton. (e) X is said to be R-inner definite if and only if X = X for all (X, X) ∈ C = ∅. (f) X is said to be R-outer definite if and only if X = X for all (X, X) ∈ C = ∅. When X is R-definite, we have C = {(X, X)}. This implies that the concept expressed by X is precise. X is R-definite whenever X is R-inner

300

M. Inuiguchi and T. Tanino

and outer definite. X is R-inner definite if R∗3 (X) = X and X is R-outer definite if R3∗ (X) = X. When R is an equivalence relation, X is R-definite if and only if R∗3 (X) = R3∗ (X) = X. Thus, the definiteness corresponds to complete definability.

References 34.1 Bonikowski, Z., Bryniarski, E., Wybraniec-Skardowska, U.: Extensions and Intensions in the Rough Set Theory. Information Sciences 107 (1998) 149– 167 34.2 Dubois, D., Prade, H.: Putting Rough Sets and Fuzzy Sets Together. In: Slowinski, R. (ed.), Intelligent Decision Support: Handbook of Applications and Advances of the Rough Sets Theory, Kluwer, Dordrecht (1992) 203–232 34.3 Inuiguchi, M., Tanino, T.: Fuzzy Rough Sets Based on Certainty Qualifications. Proceedings of the Forth Asian Fuzzy Systems Symposium 1 (2000) 433– 438 34.4 Slowinski, R., Vanderpooten, D.: A Generalized Definition of Rough Approximations Based on Similarity. IEEE Transactions on Data and Knowledge Engineering 12(2) (2000) 331–336

35. Two Procedures for Dependencies among Attributes in a Table with Non-deterministic Information: A Summary Hiroshi Sakai Department of Computer Engineering, Kyushu Institute of Technology, Tobata, Kitakyushu 804, Japan [email protected]

The data dependency among attributes is very important for the rule generation. So far, we proposed a dependency among attributes in a table with non-deterministic information, and developed some important algorithms. According to these algorithms, a procedure for dependencies has been implemented. This paper proposes new algorithms and enhances the implemented procedure. In two procedures, the manipulation of equivalence relations takes an important role.

35.1 Preliminary The rough set theory has been widely applied in the research areas in artificial intelligence such as knowledge, imprecision, vagueness, learning, induction, and so on [35.2], since it was proposed by Pawlak in around 1980. According to [35.2], we define a Deterministic Information System DIS = (OB, AT, {V ALa |a ∈ AT }, f ), where OB is a finite set whose elements we call objects, AT is a finite set whose elements we call attributes, V ALa is a finite set whose elements we call attribute values and f is a mapping such that f : OB × AT → ∪a V ALa which we call a classif ication f unction. For every object x, y(x = y) ∈ OB, if f (x, a) = f (y, a) for every a ∈ AT , we say there is a relation for x and y, which becomes an equivalence relation over OB. We express an equivalence class with an object x as [x]. If a set X(⊂ OB) is the union of some equivalence classes, we say X is def inable. Otherwise we say X is rough. Suppose CON (⊂AT ) and DEC(⊂AT ) denote condition attributes and decision attributes, respectively. We say that two objects x, y(x = y) ∈ OB are consistent for CON and DEC, if f (x, a) = f (y, a) for every a ∈ CON then f (x, a) = f (y, a) for every a ∈ DEC. In case every object is consistent with other objects in a DIS, we say the DIS is consistent for CON and DEC, and we see there exists a dependency between CON and DEC. Furthermore, we see every tuple restricted to CON and DEC is a rule. In case a DIS is not consistent for CON and DEC, a ratio |P OSCON (DEC)|/|OB| is applied to measure the degree of dependency. Here, the set P OSCON (DEC) = ∪{L ∈eq(CON )|there exists such M ∈eq(DEC) as L ⊂M } is called the positive region. T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 01− 3 05, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

302

H. Sakai

35.2 Definitions of NISs We show a framework of the Non-deterministic Information System NIS according to [35.1]. We define a N IS = (OB, AT, {V ALa |a ∈ AT }, g), where g is a mapping such that g : OB×AT → P (∪a V ALa )(A power set of ∪a V ALa ). Every set g(x, a) is interpreted as that there is an actual value in this set but we do not know it. This is called the unknown interpretation for the incomplete information. Especially if we do not know the attribute value at all, we consider g(x, a) = V ALa . This is called the null value interpretation. As for N ISs, Lipski showed the modal question-answering. Orlowska and Pawlak discussed the modal concept, especially the axiomatization of the logic in N ISs. Grzymala-Busse surveyed the unknown attribute values and studied the learning from examples with unknown attribute values. Example 1. Let’s consider the next N IS and the problem. Table 35.1. A Table of a N IS OB 1 2 3 4 5 6 7 8

A 3 5 {1, 4, 5} 4 3 4 5 1

B 1 {2, 4} 5 5 5 5 4 {1, 3, 4}

C 1 2 4 2 2 {1, 3} 1 1

Problem: In Table 1, do we see there exists a dependency between {A, B} and {C} ? Generally, how do we deal with the dependency in every N IS, and how effectively do we calculate the dependency in N IS ? For this problem, we consider every possible case in the N IS. In Table 1, 36(=3*2*3*2) possible DISs are derived by replacing g(x, a) with an element in g(x, a). Generally in every N IS, we call such a DIS a derived DIS from a NIS. According to derived DISs, we propose a new dependency in a N IS. A Proposal of New Dependency in a NIS Suppose there exist a N IS, all derived DIS1 , · · · , DISm , condition attributes CON , decision attributes DEC. For two threshold values val1 and val2 (0 ≤ val1, val2 ≤ 1), if the following conditions hold then we see there exists a dependency between CON and DEC in the NIS. (1) Suppose a set P = {DISi |DISi (1 ≤ i ≤ m) is consistent for CON and DEC}. For this set P , |P |/m > val1 . (2) mini {degree of dependency in DISi |(1 ≤ i ≤ m)} > val2 . This new dependency is calculated by each degree of dependency in every derived DIS. In Example 1, suppose val1 = 0.8 and val2 = 0.8. The condition (1) requires |P |/36 >0.8, namely more than 29 derived DIS must be consistent. The condition (2) requires the minimal degree of dependency is

35. Two Procedures for Dependencies among Attributes

303

more than 0.8. As for the implementation, the simple way is to calculate the degree of dependency for all derived DISs. However, this way is not suitable for N ISs with large number of derived DISs. We rely on another way for the implementation.

35.3 A Way to Obtain All Possible Equivalence Relations We call every equivalence relation in a derived DIS a possible equivalence relation (pe-relation), and call every element in a pe-relation a possible equivalence class (pe-class). Proposition 1. Suppose there exist a N IS and a set X(⊂ OB). If there exist subsets of OB, CL1 , · · · , CLm satisfying (1) and (2), X is definable in this N IS. (1) ∪i CLi = X. (2) {CL1 , · · · , CLm } is a subset of a pe-relation. According to this proposition, we check the definability of a set by finding sets CL1 , · · · , CLm . We have already realized this program. In order to obtain all pe-relations, we put X = OB. Then, all pe-relations are obtained as a side effect of checking the definability of the set OB [35.3].

35.4 Procedure 1 for Dependencies Let’s eq(CON ) and eq(DEC) be equivalence relations for the condition and decision attributes in a DIS, respectively. In this case, it is easy to calculate the degree of dependency |P OSCON (DEC)|/|OB| by eq(CON ) and eq(DEC) [35.3]. This property is applied to all pe-relations in every N IS, and the new dependency is calculated. The following is a procedure for it. Procedure 1 (Step 1) Prepare a data file and an attribute file. The attributes CON and DEC are defined in this attribute file. (Step 2) Translate them into internal expressions. (Step 3) Pick up all pe-relations for CON and DEC, respectively. (Step 4) Calculate criteria values by those relations. The following is the real execution of Step 4 in Example 1. Here, CON = {A, B} and DEC = {C}. % dependency Dependency Check [1,2] => [3] CRITERION 1 Degree of consistent DISs: 0.0 CRITERION 2 Minimal Degree of Dependency: 0.375 Maximal Degree of Dependency: 0.750 EXEC TIME = 0.030 (sec) %

304

H. Sakai

35.5 Procedure 2 for Dependencies Suppose it is necessary to check several kinds of dependencies in a N IS. In Procedure 1, the CON and DEC must be specified in Step 1. So, it is necessary to do a sequence from Step 1 to Step 4 for each dependency. To make matters worse, Step 3 is time-consuming. In such a situation, we revised Procedure 1 to Procedure 2. In Procedure 2, a merging algorithm for equivalence relations is employed. Suppose eq(A1 ) and eq(A2 ) be equivalence relations for A1 , A2 (⊂ AT ), respectively. The equivalence relation eq(A1 ∪A2 ) is {M ⊂ OB|M = L1 ∩ L2 (= ∅) for some L1 ∈ eq(A1 ) and L2 ∈ eq(A2 )}. Namely, an equivalence relation for any set of attributes can be produced from eq(a)(a ∈ AT ). Procedure 2 (EStep 1) Prepare data file. (EStep 2) Translate them to internal expressions for each attribute. (EStep 3) Make all pe-relations for each attribute. (EStep 4) Fix condition attributes CON , decision attributes DEC, and produce all pe-relations for CON and DEC, respectively. (EStep 5) Calculate two criteria values by those relations. In this procedure, it is enough to execute EStep 2 and EStep 3 only once. It is enough to do EStep 4 and EStep5 for each pair of CON and DEC.

35.6 Execution Time of Every Method Now, let us see the execution time of each method to calculate the degree of dependency. Four N ISs in Table 2 are used, and the dependencies between {A, B, C} and {D} are calculated. Every method is implemented on a workstation with the 450MHz UltraSparc CPU by Prolog and C language. Table 35.2. Definitions of N ISs N IS N IS1 N IS2 N IS3 N IS4

|OB| 10 100 300 1000

|AT | 4 4 4 4

|V ala |(a ∈ AT ) 10 10 10 100

Derived DISs 864 1944 3888 7776

According to Table 3, it is known that Step 3 in Procedure 1 and EStep 3 in Procedure 2 are the most time-consuming. These two steps pick up perelations from internal expressions. The execution time of Step 4, EStep4 and EStep 5 are very small for the total execution time in Table 4. As for N IS1 and N IS2 in Table 4, each execution time of the simple method for a DIS was 0.00 (sec). Suppose it is necessary to check five kinds of dependencies among attributes in N IS4 . In the simple method and Procedure 1, it is necessary to do all

35. Two Procedures for Dependencies among Attributes

305

Table 35.3. The execution time(sec) of Procedure 1 and 2 for checking the dependency {A, B, C} and {D}. Step 2, Step 3, EStep 2 and EStep 3 are realized by Prolog, Step 4, EStep 4 and EStep 5 by C. N IS N IS1 N IS2 N IS3 N IS4

Step2 0.05 0.48 2.60 26.57

Step3 0.17 0.84 8.67 122.45

Step4 0.06 0.07 0.07 0.14

EStep2 0.07 0.70 3.61 31.32

EStep3 0.16 2.07 5.96 45.70

EStep4 0.01 0.10 1.03 0.82

EStep5 0.05 0.12 1.00 1.37

Table 35.4. The total execution time(sec) of the simple method, Procedure 1 and 2 for checking the dependency {A, B, C} and {D}. The 2nd column simple shows such a value as (execution time to calculate the degree of dependency in a derived DIS)×(the number of all derived DISs). N IS N IS1 N IS2 N IS3 N IS4

T otal(Simple) − − 38.88 933.12

T otal(P rocedure1) 0.28 1.39 11.34 149.16

T otal(P rocedure2) 0.29 2.99 11.60 79.21

steps. Therefore, it will take about 4665.60(=933.12×5)(sec) by the simple method and about 745.80(=149.16×5)(sec) by Procedure 1, respectively. In Procedure 2, it is enough to do EStep 2 and EStep 3 only once. It is enough to repeat the EStep 4 and EStep 5 for 5 times. In this case, it will take about 87.97 (=31.32+45.70+5×(0.82+1.37))(sec).

35.7 Concluding Remarks This paper proposed a dependency in non-deterministic information systems and two procedures for calculating this new dependency. We conclude that for checking a dependency in small size data like N IS1 and N IS2 , every three method will be applicable. However for large size data like N IS4 , Procedure 1 and 2 are applicable. It will be hard to apply the simple method. For checking several kinds of dependencies, Procedure 2 is much better than Procedure 1.

References 35.1 Orlowska, E.(Ed.) (1998): Incomplete Information: Rough Set Analysis. Physica-Verlag 35.2 Pawlak, Z. (1991): Rough Sets. Kluwer Academic Publisher 35.3 Sakai, H. (2001): An Enhancement of a Procedure for Checking Dependencies of Attributes in a Table with Non-deterministic Information. Bulletin of International Rough Set Society 5(1-2), 81-87

3 6 . A n A p p lic a tio n o f E x te n d e d S im u la te d A n n e a lin g A lg o r ith m to G e n e r a te th e L e a r n in g D a ta S e t fo r S p e e c h R e c o g n itio n S y s te m Ch i -Hwa Son g a n d Won D on L ee D ep t. of Com p u ter Sci en ce, Ch u n g n a m N a t’ l Un i v. D a ej eon 3 05-764 , KOR E A, (ch s on g @ cn u .a c.k r ) (wd l ee@ cn u .a c.k r )

A b s t r a c t . In th is p a p e r , w e s u g g e s t a m e th o d o f d a ta e x tr a c tio n fo r c o n s tr u c tin g th e s p e e c h r o n th e E x te n d e d S im u r e a n te x t d a ta , d r a w n th e p r o p o s e d a lg o r ith a lp h a b e ts .

e c la ra m

o g n itio n te d A n n n d o m ly h a s th e

s y s te m . T h e p r o p o s e e a lin g (E S A ) a lg o r ith m fr o m th e in te r n e t. T h e e q u ip r o b a b le d is tr ib

d a . W K o u tio

lg o r ith e h a v re a n L n a m o

m e u D S n g

is b a s se d K b u ilt K o re

e d o b y a n

3 6 .1 . I n t r o d u c t io n Th e s p eech r ecog n i ti on s y s tem s a r e tr a i n ed u s i n g th e l ea r n i n g d a ta s et th a t i s col l ected or ex tr a cted fr om a p p r op r i a te d a ta b a n k . Up to n ow, h owever , we h a ve n o cr i ter i on wh eth er th e l ea r n i n g d a ta s et i s p r op er for th e s p eech r ecog n i ti on s y s tem th a t we b u i l d . Wor s e, i t i s h a r d to tr a i n th e s p eech r ecog n i ti on s y s tem a s th e n u m b er of th e tr a i n i n g d a ta i s i n cr ea s i n g . To m a k e th e tr a i n i n g effecti ve, we n eed en ou g h tr a i n i n g d a ta s o th a t th e r ecog n i ti on s y s tem d oes n ot r el y on s om e s p eci fi c wor d s a n d a l p h a b ets . Th e s u i ta b l e tr a i n i n g d a ta a r e r eq u i r ed i n or d er to s et u p th e m od u l e wi th th e h i g h r el i a b i l i ty . We b el i eve th a t a r i g h t tr a i n i n g d a ta s h ou l d b e s u ch th a t ea ch a l p h a b et b e m a n i fes ted eq u i p r ob a b l y . We p r op os e a m eth od of ex tr a cti n g L D S(l ea r n i n g d a ta s et) th a t h a s th e eq u i p r ob a b i l i ty i n th e p a tter n d om a i n wi th a s few el em en ts a s p os s i b l e.

3 6 .2 . D o m a in D e f in it io n f o r L D S E x t r a c t io n T a b l e 1 . Kor ea n a l p h a b et ta b l e

A Kor ea n ch a r a cter i s com p os ed of 19 i n i ti a l s ou n d s , 21 m ed i a l vowel s a n d 27 fi n a l con s on a n t. Her e, Th e fi n a l con s on a n t i s om i s s i b l e(F IL L ), s o we ca n u s e tota l 28 ch a r a cter s a s fi n a l con s on a n ts . Ta b l e 1 s h ows Kor ea n a l p h a b ets . We col l ected th e ca n d i d a te d a ta a t r a n d om i n th e i n ter n et for ex tr a cti n g th e L D S.

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 06-3 10, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

3 6. An Ap p l i ca ti on of E x ten d ed Si m u l a ted An n ea l i n g Al g or i th m

Th e wei g h t of a wor d for ex tr a cti on i s wor d a n d th e en tr op y of Kor ea n a l p h a b occu r r en ce i s m or e often th a n oth er s i n a " m ol ecu l e" of l ea r n i n g d a ta s et a n d ou va l u e of Kor ea n a l p h a b et d i s tr i b u ti on .

r el a ted to th e a eti c d i s tr i b u ti on th e ca n d i d a te d a r a l g or i th m tr i es

3 07

p p ea r a n ce fr eq u en cy of i n th os e wor d s . If a wor ta s et, th en we s el ect i t a to g et th e h i g h es t en tr op

a d s y

3 6 .3 . T h e N u m e r ic a l F o r m u la f o r L D S E x t r a c t io n B e n e f i t ( B ) = α * (∑

μ 1: w μ 2: w N : th M : th E : th

E = −

∑

μ 1i N

+ i

i

∑

μ 1j N j

)− β * E

j

e ig h t o f o c c u r r e n c e fr e q u e n c y e ig h t o f th e le n g th o f w o r d e n u m b e r o f w o r d s h a v in g th e s a m e e n u m b e r e e n tr o p y

P (k ) * L o g

o c c u rre n c e

fr e q u e n c y

o f w o r d s h a v in g th e s a m e le n g th o f le r n in g d a ta s e t in K o r e a n a p lp h a b e t s 10

P (k )

B e n e f i t ( B ) i s th e cr i ter i on th a t we wa n t to m a x i m i z e. L ea r n i n g d a ta s et con s i s ts of th e wor d s wh i ch m a k e th e s et h a ve m a x i m u m B . μ 1 i s th e wei g h t of a wor d . It i s i n p r op or ti on a l to th e a p p ea r a n ce fr eq u en cy i n th e tex t. F or ex a m p l e, i n th e [ h a k g y o (s ch ool )] i s r ep ea ted 2 ti m es a n d th e wor d wh ol e d a ta , i f th e wor d [ h a k s ea n g (s tu d en t)] i s r ep ea ted 1 ti m e, we s el ect th e wor d to i n cl u d e i n th e L D S. μ 2 i s u s ed a s a p a r a m eter to s el ect th e p r op er l en g th of wor d . G en er a l l y , s h or t wor d s a p p ea r r ecu r r en tl y i n th e Kor ea n tex t. If we d i vi d e wor d s etc. a p p ea r s b y th e b l a n k i n th e tex t, th e l etter s l i k e m or e often th a n oth er wor d s . In th e ex p er i m en t, we s et th e a ver a g e l en g th of wor d i s 3 a n d i n th a t ca s e i t h a s th e b i g g es t μ 2 . If th e wor d i s l on g er or s h or ter th a n 3 -l etter -l en g th , i t i s s el ected i n th e ca n d i d a te d a ta s et wi th l ow p r ob a b i l i ty . F i g 1 a n d F i g 2 r ep r es en t th e wei g h ts of μ 1 a n d μ 2 u s ed i n th e ex p er i m en t. TPW

SPS S

T

RP[

1

μ 2

SPW

S

RPY

RPW

RPX RPW

R S

T

U

V

W

X

Y

ƒ’’‡ƒ”ƒ…‡Bˆ”‡“—‡…›B‘ˆB™‘”†

F ig . 1 .

M

RPZ

μ

1

Z

[

SR

S

T

U

V

W

X

Y

Z

n‡‰–ŠB‘ˆB™‘”†

F ig . 2 .

μ

2

s h ows th e n u m b er of wor d s th a t h a ve th e s a m e occu r r en ce-fr eq u en cy a n d N a p p ea r s s h ows th e n u m b er of wor d s th a t h a ve th e s a m e l en g th . If th e wor d 3 ti m es i n th e wh ol e d a ta , th en μ 1 i s to b e 1.75 a s F i g 1. L i k ewi s e, i f th e wor d s

3 08

C.-H. Son g a n d W.D . L ee

wi th μ 1 va l u es eq u a l to 1.75 a p p ea r 100 ti m es i n th e L D S, th en N i s s et to 100. μ 2 of th e wor d s wi th l en g th 3 , s u ch a s i s s et to 1. An d M i s th e n u m b er of wor d s wi th μ 2 va l u e s et to 1 i n th a t ca s e. Th e con s ta n ts , d eter m i n e th e i m p or ta n ce b etween two ter m s . k i s th e i n d ex of Kor ea n a l i n th e wh ol e tex t d a ta . Th er ep h a b ets a n d i s th e p r ob a b i l i ty of Kor ea n a l p h a b et for e E s h ows th e en tr op y of Kor ea n a l p h a b et i n th e s el ected wor d s . As a wh ol e, th e fi n a l r es u l t of L D S con s i s ts of th e wor d s wi th th e a p p r op r i a te l en g th a n d a l s o wi th th e fr eq u en tl y a p p ea r ed wor d s a n d th e s et h a s th e m a x i m u m en tr op y va l u e.

3 6 .4 . T h e A lg o r it h m

fo r E x tr a c tio n o f L D S

Th e p r e-p r oces s i n g i s d on e; 1. D i vi d e th e wh ol e tex t d a ta i n to th e u n i t(wor d ) th a t we i n ten d to r ecog n i z e. 2. R em ove th e r ep ea ted wor d s a n d ca l cu l a te th e va l u es of μ 1, μ 2 3 . ca l cu l a te th e occu r r en ce of ea ch a l p h a b et i n th e wh ol e wor d s . Th e d a ta s et ex tr a cti on a l g or i th m b a s ed on E SA: T ( o ) = i n i ti a l tem p er a tu r e(T ( o ) > = T ( i ) > = T ( f ) ) T ( f ) = fi n a l tem p er a tu r e D = a n n ea l i n g s ch ed u l e p a r a m eter s cl os e to 1. L D S o = a s ta b l e l ea r n i n g d a ta s et, i t s ta r ts wi th th e i n i ti a l L D S. L D S n = a p er tu r b ed L D S o A n n e a lin g ()

L D S n = L D S o + w 1 - w 2 b re a k

{ d o { d o { s w itc h (R a n { c a se 0 : s e le c t L D S n b re a k c a se 1 : s e le c t L D S n b re a k c a se 2 : s e le c t s e le c t

} D e c id e A c c e p tO fL D S n () } w h ile (u n s ta b le s ta te ) T (i) * = D } w h ile (T > F )

d o m [ 0 ,1 ,2 ] )

a w o r d w o u ts id e o f L D S o = L D S o + w

a w o r d w in th e L D S o = L D S o – w

a w o r d w 1 o u ts id e o f L D S o a w o r d w 2 in o f L D S o

} D e c id e A c c e p tO fL { c a lc u la te if (( > 0 .0 e ls e -if e x p (L D S o = L D e ls e R e je c t th e L }

D S n () = B n e fit o f L D S n - B n e fit o f L D S o ) L D S o = L D S n ; /T ( i) ) < R a n d o m [ 0 ,1 ] ] S n ; D S n ;

3 6. An Ap p l i ca ti on of E x ten d ed Si m u l a ted An n ea l i n g Al g or i th m

3 09

Th e L D S ex tr a cti on Al g or i th m i s fou n d ed on E SA[ Wd l ee, 1997] . We ob ta i n th e ca n d i d a te d a ta s et th r ou g h th e p r e-p r oces s a s fol l ows . After m a k i n g ca n d i d a te d a ta s et, we ch oos e s om e a r b i tr a r y wor d s i n th e s et to b u i l d th e i n i ti a l L D S a n d com p u te i ts i n i ti a l b en efi t(B ).

3 6 .5 . E x p e r im e n t a l a n d R e s u lt Th e d a ta u s ed i n th e ex p er i m en t a r e 168 71 wor d s i n tota l a fter r em ovi n g th e r ep ea ted wor d s , col l ected i n th e i n ter n et a t r a n d om . In th e ex p er i m en t, th e b en efi t of . B u t we ca n g et th e d a ta s et d oes n ot d ep en d m u ch on th e va l u es of a n d m a x i m u m b en efi t wh en th e n u m b er of th e wor d s i s a b ou t 4 950. Th e fol l owi n g F i g 3 r ep r es en ts th e ch a n g e of E n er g y (-B en efi t) wi th ti m e wh en 4 219 r a n d om l y ch os en wor d s a r e u s ed to s et u p th e i n i ti a l d a ta s et. F i n a l l y , th e a l g or i th m ex tr a cts i s 100. 4 94 6 wor d s wh en th e va l u e of i s 0.1 a n d In th e fi g u r e 3 , th e en er g y i s d ecr ea s i n g a n d fi n a l l y con ver g es to a va l u e a fter s om e i ter a ti on . Al th ou g h we ch a n g e th e i n i ti a l d a ta s et to i n cl u d e m or e th a n 1000 wor d s or l es s th a n 1000 wor d s , th e fi n a l r es u l t d oes n ot ch a n g e th a t m u ch . K K – ‹ ˆ ‡ 

k–‡ ”ƒ –‹‘  R

‡ d OSR O J _ _ J › OTR ‰ ” ‡  g OUR

S

U

W

Y

[

S S

U S

W S

Y S

[ S

S T

U T

W T

Y T

[ T

S U

U U

W U

Y U

[ U

S V

U V

W V

Y V

[ V

S W

OVR OWR OXR OYR OZR O[R

F i g . 3 . Th e ch a n g e of – B e n e f i t – UW R RR ‡ „ ƒ Š ’ ƒ B ˆ UR R RR ‘ B ” ‡ „ TW R RR  — p

[RRR

S T U

YRRR

–‡ „ ƒŠ Ž’ƒ ˆB B‘” ‡ „ — p

TR R RR SW R RR SR R RR

S T U

ZRRR

XRRR WRRR VRRR URRR TRRR

W R RR

SRRR

R S

T

U

V

W

X

Y

Z

[ S R S S S T S U S V SW S X S Y SZ S[ T R T S T T TU T V TW TX T Y T Z

ƒ Ž’ Š ƒ „ ‡ –B‹ † ‡ š

R S

T

U

V

W

X

Y

Z

[

SR SS ST SU S V SW SX S Y SZ S[ TR TS TT TU TV TW TX TY TZ

ƒ Ž’ Š ƒ „ ‡ –B‹ † ‡ š

F i g . 4 . Al p h a b ets a p p ea r ed i n i n i ti a l d a ta

F i g . 5 . Al p h a b et a fter fi n a l ca l cu l a ti on i n th e L D S

F i g 4 r ep r es en ts th e tota l n u m b er of a l p h a b s en ts i n i ti a l s ou n d a l p h a b ets , l i n e 2 r ep r es a n d th e l i n e3 r ep r es en ts th e fi n a l con s on a s ou n d ) a p p ea r ed m or e th a n 3 0000 ti m es a i n i ti a l d a ta s et. Wh en th e a l g or i th m i s a p p

ets i n th e ca n d i d a te L D S. L i n e 1 r ep r een t th e m ed i a l vowel s s ou n d a l p h a b ets n t a l p h a b ets . In F i g 4 , (th e fi r s t i n i ti a l n d s om e a l p h a b ets d i d n ’ t a p p ea r i n th e l i ed to s u ch a n i n i ti a l d a ta s et, th e oc-

3 10

C.-H. Son g a n d W.D . L ee

cu r r ed l es s th a n 8 000 ti m es a n d th e d a ta b et We ca n g et a tr a i n i n g d a ta s et wi th a r i th m , th er eb y s a vi n g th e ti m e to tr a i n th of wor d s . Th e s y s tem p r otects th e r ecog s p eci fi c s ou n d or s y l l a b l e a n d th u s ca n m

s et b ecom es eq u i p r ob a b l e on ea m i n i m a l n u m b er of wor d s wi th e r ecog n i ti on s y s tem wi th a l a r g n i ti on s y s tem fr om b ei n g tr a i n ed a x i m i z e i ts r el i a b i l i ty .

ch a th i s en u b y

lp h a lg m b s om

a oer e

3 6 .6 . C o n c lu s io n In th e ex p er i m en t, a n d va l u es a r e s et b y tr i a l a n d er r or . To g et th e a n d , we ex p er i m en t on th e p os s i b l e com b i n a ti on s a b ou t 5 ti m es i n th e s a m e con d i ti on . An d th e p r op os ed a l g or i th m i s fou n d to b e s en s i ti ve to th os e va l u es . Som e Kor ea n a l p h a b ets a r e m or e often th a n oth er s i n th e u s u a l Kor ea n tex t, b u t th e a l g or i th m ca n r ed u ce th i s u n b a l a n ce. Th e l ea r n i n g d a ta s et(L D S), ex tr a cted b y th e p r op os ed a l g or i th m h a s th e r eg u l a r d i s tr i b u ti on i n th e d om a i n of Kor ea n a l p h a b ets . Th e p r op os ed a l g or i th m m a k es i m p r ovem en t i n th e r el i a b i l i ty th r ou g h th e r efor m a ti on of th e s p eech r ecog n i ti on s y s tem .

R e fe r e n c e s 3 6.1 3 3 3 3 3

3 3 3

3

H.Cr owd er , a n d M.W.P a d b er g , " Sol vi n g L a r g e-Sca l e Sy m m etr i c Tr a vel l i n g Sa l es m a n P r ob l em to Op ti m a l i ty " , Ma n a g em en t Soci ., 26, 4 95-509, 198 0. 6.2 T.L .Hi l l " Sta ti s ti ca l Th er m od y n a m i cs " , a d d i s on -Wes l ey P u b l i s h i n g Com p a n y , 1960 6.3 B .Wi d om , " Som e top i cs i n th e th eor y of fl u i d s " , J. of Ch em . P h y s . 3 9, 28 08 -28 12, 1963 . 6.4 D .J.Ad a m s , " G r a n d ca n on i ca l en s em b l e Mon te Ca r l o for a l eon a r d -Jon es fl u i d " , Mol . P h y s . 29, 3 07-3 11, 1976. 6.5 N .R .D r a p er , H.Sm i th , " Ap p l i ed R eg r es s i on An a l y s i s " j oh n Wi el y & Son s , In c., 9092, 198 1. 6.6 N .Metr op ol i s , A.R os en b l u s h , M.R os en b l u th , A.Tel l er a n d E .Tel l er , " E q u a ti on of Sta te Ca l cu l a ti on b y F a s t Com p u ti n g Ma ch i n es " , J. of Ch em . P h y s i cs , 21, 108 71092, 1953 . 6.7 S.G em a n , D .G em a n , " Stoch a s ti c r ea l i z a ti on , G i b b s d i s tr i b u ti on s a n d th e b a y es i a n r es tor a ti on of i m a g es " , IE E E Tr a n s ., P AMI-6, 721-74 1, 198 4 . 6.8 S.Ki r k p a tr i ck , C.D .G el l a tt Jr . a n d M.P Vecch i , " Op ti m i z a ti on b y Si m u l a ted An n ea l i n g " , Sci en ce, Vol 220, 671-68 0, 198 3 . 6.9 D .S.Joh n s on , C.R .Ar a g on , L .A.Mcg eoch , C.Sch evon , " Op ti m i z a ti on b y Si m u l a ted An n ea l i n g : a n E x p er i m en ta l E va l u a ti on " , Wor k s h op on Sta ti s ti ca l P h y s i cs i n E n g i n eer i n g a n d B i ol og y , Yor k town Hei g h ts , Ap r i l 198 4 . 6.10 Wd l ee, Ch s on g " E x ten d ed Si m u l a ted An n ea l i n g Al g or i th m B a s ed on G r a n d Ca n on i ca l E n s em b l e" , ICON IP , D u n ed i n , N ewZ ea l a n d , 1997, 11.

37. Generalization of Rough Sets with α-Coverings of the Universe Induced by Conditional Probability Relations Rolly Intan1 , Masao Mukaidono1 , and Y.Y. Yao2 1 2

Meiji University, Kawasaki-shi, Kanagawa-ken, Japan University of Regina, Regina, Saskatchewan, Canada S4S 0A2

Standard rough sets are defined by a partition induced by an equivalence relation representing discernibility of elements. Equivalence relations may not provide a realistic view of relationships between elements in real-world applications. One may use coverings of, or non-equivalence relations on, the universe. In this paper, the notion of weak fuzzy similarity relations, a generalization of fuzzy similarity relations, is used to provide a more realistic description of relationships between elements. A special type of weak fuzzy similarity relations called conditional probability relation is discussed. Generalized rough set approximations are proposed by using α-coverings of the universe induced by conditional probability relations.

37.1 Introduction The theory of rough sets plays essential roles in many applications of data mining and knowledge discovery [37.6]. It offers a mathematical model and tools for discovering hidden patterns in data, recognizing partial or total dependencies in data bases, removing redundant data, and many others [37.4, 37.6]. Rough set theory generalizes classical set theory by offering an alternative formulation of sets with imprecise boundaries. A rough set may be viewed as an approximate representation of a given crisp set in terms of two subsets derived from a partition on the universal set [37.3]. The two subsets are called a lower approximation and an upper approximation. Although rough set theory built on equivalence relation has the advantage of being easy to analyze, it may not be a widely applicable model as equivalence relations may not provide a realistic view of relationships between elements in real world. Coverings of, non-equivalence relations, on the universe may be used to provide a more realistic model of rough sets. A covering of the universe, C= {C1 , ..., Cn }, is a family of subset of non-empty universe U such that U = {Ci | i = 1, ..., n}. The sets in C(x) may describe different types or various degrees of similarity between elements of U . The interpretation and construction of subsets in a covering are some of the fundamental issues of covering based formulation of rough set theory. Crisp and fuzzy binary relations may be used for such purposes. In general, relationships between elements in real-world applications may not necessarily be symmetric or transitive. Recently, conditional probability relations [37.1] was introduced T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 11− 3 15, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

312

R. Intan, M. Mukaidono, and Y.Y. Yao

for representing such non-equivalence relationships between elements. Conditional probability relations may be considered as a generalization of fuzzy similarity relations. The main objective of this paper is to generalize the standard concept of rough sets by coverings of the universe. Conditional probability relations are used in the construction of coverings. Rough set approximations are introduced based on α-coverings of the universe induced by the α-cuts of a conditional probability relation. The proposed rough sets may be considered as generalized fuzzy rough sets [37.7].

37.2 Conditional Probability Relations The concept of conditional probability relations was introduced in the context of fuzzy relational databases [37.1]. It may be considered as a concrete example of weak fuzzy similarity relation,which in turn is a special type of fuzzy binary relation. Definition 37.2.1. A fuzzy similarity relation is a mapping, s : U ×U → [0, 1], such that for x, y, z ∈ U , (a) Reflexivity : s(x, x) = 1, (b) Symmetry : s(x, y) = s(y, x), (c) Max−min transitivity : s(x, z) ≥ max min[s(x, y), s(y, z)]. y∈U

Definition 37.2.2. A weak fuzzy similarity relation is a mapping, s : U × U → [0, 1], such that for x, y, z ∈ U , (a) Reflexivity : s(x, x) = 1, (b) Conditional symmetry : if s(x, y) > 0 then s(y, x) > 0, (c) Conditional transitivity : if s(x, y) ≥ s(y, x) > 0 and s(y, z) ≥ s(z, y) > 0 then s(x, z) ≥ s(z, x). Definition 37.2.3. A conditional probability relation is a mapping, R : U × U → [0, 1], such that for x, y ∈ U , R(x, y) = P(x | y) = P(y → x) =

|x ∩ y| , |y|

where R(x, y) means the degree y supports x or the degree y is similar to x. By definition, a fuzzy similarity relation is regarded as a special case (or type) of weak fuzzy similarity relation, and a conditional probability relation is an example of weak fuzzy similarity relations. The conditional probability relations may be used as a basis of representing degree of similarity relationships between elements in the universe U . In the definition of conditional

37. Generalization of Rough Sets

313

probability relations, the probability values may be estimated based on the semantical relationships between elements by using the epistemological or subjective view of probability theory. When objects in U are represented by sets of features or attributes as in the case of binary information tables, we have a simple procedure for estimating the conditional probability relation as shown in Definition 37.2.3, where | · | denotes the cardinality of a set. The notion of binary information tables can be easily generalized to fuzzy information tables by allowing a number in the unit interval [0, 1] for each cell of the table. The number is the degree to which an element has a particular attribute. Each object is represented as a fuzzy set of attributes. The degree of similarity two objects can be calculated by a conditional probability relation on fuzzy sets [37.1, 37.2]. In this case, |x| = a∈At μx (a), where μx is membership function of x over a set of attribute At, and intersection is defined by minimum. Definition 37.2.4. Let μx and μy be two fuzzy sets over a set of attribute At for two elements x and y of a universe of objects U . A fuzzy conditional probability relation is defined by:  min{μx (a), μy (a)} . R(x, y) = a∈At a∈At μy (a) It can be easily verified that R satisfies properties of a weak fuzzy similarity relation. Additional properties of similarity as defined by conditional probability relations can be found in [37.1].

37.3 Generalized Rough Sets Approximation From weak fuzzy similarity relations and conditional probability relations, coverings of the universe can be defined and interpreted. The standard concept of rough sets can thus be generalized based on coverings of universe. Definition 37.3.1. Let U be a non-empty universe, and R be a conditional probability relation on U . For any element x ∈ U , Rsα (x) and Rpα (x) are defined as the set of elements that support x and the set of elements that are supported by x, respectively, to a degree of at least α ∈ [0, 1], as follows: Rsα (x) = {y ∈ U | R(x, y) ≥ α}, Rpα (x) = {y ∈ U | R(y, x) ≥ α}. The set Rsα (x) can also be interpreted as consisting of elements that are similar to x, while Rpα (x) consisting of elements to which x is similar. By the reflexivity, it follows that we can construct two covering of the universe, {Rsα (x) | x ∈ U } and {Rpα (x) | x ∈ U }. By extending standard rough sets, we obtain two pairs of generalized rough set approximations.

314

R. Intan, M. Mukaidono, and Y.Y. Yao

Definition 37.3.2. For a subset A ⊆ U , we define two pairs of generalized rough set approximations: (i) element-oriented generalization: α Lα e (A) = {x ∈ U | Rs (x) ⊆ A}, α Ue (A) = {x ∈ U | Rsα (x) ∩ A = ∅}. (ii) similarity-class-oriented generalization:  α α Lα c (A) = {Rs (x) | Rs (x) ⊆ A, x ∈ U }, Ucα (A) = {Rsα (x) | Rsα (x) ∩ A = ∅, x ∈ U }. In Definition 37.3.2(i), the lower approximation consists of those elements in U whose similarity classes are contained in A. The upper approximation consists of those elements whose similarity classes overlap with A. In Definition 37.3.2(ii), the lower approximation is the union of all similarity classes that are contained in A. The upper approximation is the union of all similarity classes that overlap with A. Relationships among the these approximations can be represented by: α α α Lα e (A) ⊆ Lc (A) ⊆ A ⊆ Ue (A) ⊆ Uc (A). The difference between lower and upper approximations is the boundary region with respect to A: α α Bndα e (A) = Ue (A) − Le (A),

α α Bndα c (A) = Uc (A) − Lc (A).

Similarly, one can define rough set approximations based on the covering {Rpα (x) | x ∈ U }. α The pair (Lα e , Ue ) may be interpreted as a pair of set-theoretic operators on subset of the universe. It is referred to as rough set approximation operators [37.8]. By combining with other set-theoretic operators such as ¬, ∪, and ∩, we have the following results: α (re0) Lα e (A) = ¬Ue (¬A), α Ue (A) = ¬Lα e (¬A), α (re1) Lα e (A) ⊆ A ⊆ Ue (A), α (∅) = U (∅) = ∅, (re2) Lα e e α (re3) Lα (U ) = U (U ) = U, e e α (A ∩ B) = L (A) ∩ Lα (re4) Lα e e e (B), α α Ue (A ∩ B) ⊆ Ue (A) ∩ Ueα (B),

α α (re5) Lα e (A ∪ B) ⊇ Le (A) ∪ Le (B), α α Ue (A ∪ B) = Ue (A) ∪ Ueα (B), (re6) A = ∅ =⇒ Ue0 (A) = U, (re7) A ⊂ U =⇒ L0e (A) = ∅, (re8) α ≤ β =⇒ [Ueβ (A) ⊆ Ueα (A), β Lα e (A) ⊆ Le (A)], (re9) A ⊆ B =⇒ [Ueα (A) ⊆ Ueα (B), α Lα e (A) ⊆ Le (B)].

(re0) shows that lower and upper approximations are dual operators with respect to set complement ¬. (re2) and (re3) provide two boundary conditions. (re4) and (re5) may be considered as weak distributive and distributive over set intersection and union, respectively. When α = 0, (re6) and (re7) show that lower and upper approximations of a non-empty set A ⊂ U are equal to U and ∅, respectively. (re8) shows that if the value of α is larger then the lower approximation is also bigger, but the upper approximation is smaller. (re9) indicates the consistency of inclusive sets. α Lower and upper approximations of Definition 37.3.2(ii), the pair (Lα c , Uc ), satisfy the following properties:

37. Generalization of Rough Sets α Lα c (A) ⊆ A ⊆ Uc (A), α (∅) = U (∅) = ∅, Lα c c α Lα (U ) = U (U ) = U, c c α (A ∩ B) ⊆ L (A) ∩ Lα Lα c c c (B), α α Uc (A ∩ B) ⊆ Uc (A) ∩ Ucα (B), α α (rc4) Lα c (A ∪ B) ⊇ Lc (A) ∪ Lc (B), α α Uc (A ∪ B) = Uc (A) ∪ Ucα (B),

(rc0) (rc1) (rc2) (rc3)

315

α α (rc5) Lα c (A) = Lc (Lc (A)), α Ucα (A) = Lα (U c c (A)), (rc6) A = ∅ =⇒ Uc0 (A) = U, (rc7) A ⊂ U =⇒ L0c (A) = ∅, (rc8) α ≤ β =⇒ [Ucβ (A) ⊆ Ucα (A), β Lα c (A) ⊆ Lc (A)], (rc9) A ⊆ B =⇒ [Ucα (A) ⊆ Ucα (B), α Lα c (A) ⊆ Lc (B)].

It should be pointed out that they are not a pair of dual operators. Property (rc5) indicates that the results of iterative operations of both lower and upper approximation operators are the same a single application.

37.4 Conclusions In this paper, we introduce the notion of weak fuzzy similarity relations. Two examples of such relations, conditional probability relations and fuzzy conditional probability relations, are suggested for the construction and interpreting coverings of the universe. Based on such coverings, we generalize the standard rough set approximations. Two pairs of lower and upper approximation operators are suggested and studied. Their properties are examined.

References 37.1 Intan, R. and Mukaidono, M., Conditional probability relations in fuzzy relational database, Proceedings of RSCTC’00, pp. 213-222, 2000. 37.2 Intan, R., Mukaidono, M., ‘Fuzzy Functional Dependency and Its Application to Approximate Querying’, Proceedings of IDEAS’00, (2000), pp.47-54. 37.3 Klir, G.J. and Yuan, B., Fuzzy Sets and Fuzzy Logic: Theory and Applications, Prentice Hall, New Jersey, 1995. 37.4 Komorowski, J., Pawlak, Z., Polkowski, L., Skowron, A., ‘Rough Sets: A Tutorial’, (1999). 37.5 Pawlak, Z., Rough sets, International Journal Computation & Information Science, 11, pp. 341-356, 1982. 37.6 Polkowski, L. and Skowron, A. (Eds.), Rough Sets in Knowledge Discovery, I, II, Physica-Verlag, Heidelberg, 1998. 37.7 Yao, Y.Y., Combination of rough and fuzzy sets based on α-level sets, in: Rough Sets and Data Mining: Analysis for Imprecise Data, Lin, T.Y. and Cercone, N. (Eds.), Kluwer Academic Publishers, Boston, pp. 301-321, 1997. 37.8 Yao, Y.Y., A comparative study of fuzzy sets and rough sets, International Journal of Information Science, 109, pp. 227-242, 1998. 37.9 Yao, Y.Y. and Zhang, J.P., Interpreting fuzzy membership functions in the theory of rough sets, Proceedings of RSCTC’00, pp. 50-57, 2000.

38. On Mining Ordering Rules Y.Y. Yao and Ying Sai Department of Computer Science, University of Regina Regina, Saskatchewan, Canada S4S 0A2 [email protected]

Many real world problems deal with ordering of objects instead of classifying objects, although majority of research in machine learning and data mining has been focused on the latter. In this paper, we formulate the problem of mining ordering rules as finding association between orderings of attribute values and the overall ordering of objects. An example of ordering rules may state that “if the value of an object x on an attribute a is ordered ahead of the value of another object y on the same attribute, then x is ordered ahead of y”. For mining ordering rules, the notion of information tables is generalized to ordered information tables by adding order relations on attribute values. Such a table can be transformed into a binary information table, on which any standard data mining algorithm can be used.

38.1 Introduction In real world situations, we may be faced with many problems that are not simply classification [38.1, 38.4]. One such type of problems is the ordering of objects. Two familiar examples of ordering problems are the ranking of universities and the ranking of the consumer products produced by different manufactures. In both examples, we have a set of attributes that are used to describe the objects under consideration, and an overall ranking of objects. Consider the example of ranking consumer products. Attributes may be the price of the products, warranty of the products, and other information. The values of a particular attribute, say the price, naturally induce an ordering of objects. The overall ranking of products may be produced by the market shares of different manufactures. The orderings of objects by attribute values may not necessarily be the same as the overall ordering of objects. The problem of mining ordering rules can be stated as follows. There is a set of objects described by a set of attributes. There is an ordering on values of each attribute, and there is also an overall ordering of objects. The overall ordering may be given by experts or obtained from other information, either dependent or independent of the orderings of objects according to their attribute values. We are interested in mining the association between the overall ordering and the individual orderings induced by different attributes. More specifically, we want to derive ordering rules exemplified by the statement that “if the value of an object x on an attribute a is ordered ahead of the value of another object y on the same attribute, then x is ordered ahead of y”. In this setting, a number of important issues arise. It would be interesting to T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 16− 3 21, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

38. On Mining Ordering Rules

317

know which attributes play more important roles in determining the overall ordering, and which attributes do not contribute at all to the overall ordering. It would also be useful to know which subset of attributes would be sufficient to determine the overall ordering. The dependency information of attributes may also be valuable. For mining ordering rules, we first introduce the notion of ordered information tables as a generalization of information tables. We then transform an ordered information table into a binary information table, on which any standard data mining and machine learning algorithms can be applied. Typically, an ordering rule may not be exact. In order to capture the uncertainty associated with ordering rules, two quantitative measures are used. They are the accuracy and the coverage of the rules [38.5, 38.7]. The former deals with the correctness of the rules, and the latter represents the extent to which the rule covers the positive instances. Ordered information tables are related to ordinal information systems proposed and studied by Iwinski [38.3]. Mining ordering rules has been studied by Greco, Matarazzo and Slowinski [38.2]. Based on these studies, the main objective of the present paper is to precisely define and formulate the problem of mining ordering rules.

38.2 Ordered Information Tables Formally, an ordered information table is defined by: OIT = (U, At, {Va | a ∈ At}, {Ia | a ∈ At}, {a | a ∈ At}), where U is a finite nonempty set of objects, At is a finite nonempty set of attributes, Va is a nonempty set of values for a ∈ At, Ia : U → Va is an information function, a ⊆ Va × Va is an order relation on Va . Each information function Ia is a total function that maps an object of U to exactly one value in Va . An ordered information table can be conveniently given in a tabular form, the rows correspond to objects of the universe, the columns correspond to a set of attributes, and each cell is the value of an object with respect to an attribute. The order relations can be interpreted as additional semantics information about the table. An order relation should satisfy certain conditions. We consider the following two properties [38.6]: Asymmetry : x  y =⇒ ¬(y  x), Negative transitivity : [¬(x  y), ¬(y  z)] =⇒ ¬(x  z).

318

Y.Y. Yao and Y. Sai

An order relation satisfying these properties is called a weak order. An important implication of a weak order is that the following relation, x ∼ y ⇐⇒ [¬(x  y), ¬(y  x)],

(38.1)

is an equivalence relation. For two elements, if x ∼ y we say x and y are indiscernible by . The equivalence relation ∼ induces a partition U/∼ on U , and an order relation on U/∼ can be defined by: [x]∼ ∗ [y]∼ ⇐⇒ x  y,

(38.2)

where [x]∼ is the equivalence class containing x. Moreover, ∗ is a linear order [38.6]. Any two distinct equivalence classes of U/∼ can be compared. It is therefore possible to arrange the elements into levels, with each level consisting of indiscernible elements defined by . For a weak order, ¬(x  y) can be written as y  x or x  y, which means y  x or y ∼ x. For any two elements x and y, we have either x  y or y  x, but not both. We assume that all order relations are weak orders. An order relation on values of an attribute a naturally induces an ordering of objects: x {a} y ⇐⇒ Ia (x) a Ia (y),

(38.3)

where {a} denotes an order relation on U induced by the attribute a. An object x is ranked ahead of another object y if and only if the value of x on the attribute a is ranked ahead of the value of y on a. The relation {a} has exactly the same properties as that of a . For simplicity, we also assume that there is a special attribute, called decision attribute. The ordering of objects by the decision attribute is denoted by o and is called the overall ordering of objects. For a subset of attributes A ⊆ At, we define: x A y ⇐⇒ ∀a ∈ A[Ia (x) a Ia (y)]   Ia (x) a Ia (y) ⇐⇒ {a} . ⇐⇒ a∈A

(38.4)

a∈A

That is, x is ranked ahead of y if and only if x is ranked ahead of y according to all attributes in A.

38.3 Mining Ordering Rules With an ordered information table, we are interested in find ordering rules of the form φ ⇒ ψ, where φ and ψ are expressions regarding ordering of objects based on certain attributes. For an attribute a, we can construct two atomic expressions (a, ) and (a, ). The former indicates that objects are ordered based on  and the latter indicates that objects are ordered based on . A set of expressions can be obtained from atomic expressions through the application of logic connectives ¬, ∧ and ∨. Consider an ordering rule,

38. On Mining Ordering Rules

319

(a, ) ∧ (b, ) ⇒ (c, ). It can be re-expressed as, x {a} y ∧ x {b} y ⇒ x {c} y, and paraphrased as follows. For two arbitrary objects x and y, if x is ranked ahead of y by attribute a, and at the same time, x is not ranked ahead of y by attribute b, then x is ranked ahead of y by attribute c. The meanings of expressions are defined by: (m1). (m2). (m3). (m4). (m5).

m((a, )) = {(x, y) ∈ U × U | x {a} y}, m((a, )) = {(x, y) ∈ U × U | x {a} y}, m(¬φ) = −m(φ), m(φ ∧ ψ) = m(φ) ∩ m(ψ), m(φ ∨ ψ) = m(φ) ∪ m(ψ).

A pair (x, y) ∈ m(φ) is said to satisfy the expression φ. In terms of the meanings of expressions, we can have many conditional probabilistic interpretations for ordering rules [38.7]. We choose to use two measures called accuracy and coverage, which are defined by [38.5]: accuracy(φ ⇒ ψ) =

|m(φ ∧ ψ)| , |m(φ)|

coverage(φ ⇒ ψ) =

|m(φ ∧ ψ)| , |m(ψ)| (38.5)

where | · | denotes the cardinality of a set. While the accuracy reflects the correctness of the rule, the coverage reflects the applicability of the rule. If accuracy(φ ⇒ ψ) = 1, the orderings by φ would determine the orderings by ψ. We thus have a strong association between the two orderings. A smaller value of accuracy indicates a weaker association. An ordering rule with higher coverage suggests that ordering of more pairs of objects can be derived from the rule. The accuracy and coverage are not independent of each other, as both are related to the quantity |m(φ ∧ ψ)|. It is desirable for a rule to be accurate as well as to have a high degree of coverage. In general, one may observe a trade-off between accuracy and coverage. A rule with higher coverage may have a lower accuracy, while a rule with higher accuracy may have a lower coverage. From an ordered information table, we can construct a binary information table. We consider all pairs of objects which are the Cartesian product U ×U . The information function is defined by:  1, x {a} y, Ia (x, y) = (38.6) 0, x {a} y. The value 1 corresponds to the atomic expression (a, ) and the value 0 corresponds to the atomic expression (a, ). Statements in an ordered information

320

Y.Y. Yao and Y. Sai

table can be translated into equivalent statements in the binary information table, and vice versa. For example, a pair (x, y) satisfies the expression (a, ) if and only if it satisfies an expression Ia (x, y) = 1. In other words, the statement x {a} y can be translated into an equivalent statement Ia (x, y) = 1. In the translation process, we will not consider object pairs of the form (x, x), as we are not interested in them. The interpretation of an ordered information table and the translation to a binary information table are crucial for mining ordering rules. Once we obtain the binary information table, any standard machine learning and data mining algorithms can be used to mine ordering rules. One may also use other types of translation methods. For example, we may consider two strict order relations  and ≺, instead of  and . Alternatively, one may translate an ordered information table into a three-valued information table, corresponding to , ≺, and ∼. It is important to realized that the framework presented in this paper can be easily applied with very simple modification.

38.4 Conclusion Ordering of objects is a fundamental issue in human decision making and may play a significant role in the design of intelligent information systems. This problem is considered from the perspective of data mining. The commonly used attribute value approaches are extended by introducing order relations on attribute values. Mining ordering rules is formulated as the process of finding associations between orderings on attribute values and the overall ordering of objects. These ordering rules tell us, or explain, how objects should be ranked according to orderings on their attribute values. Our main contribution is the formulation of the problem of mining ordering rules, and the translation of the problem to existing data mining problems. Consequently, one can directly apply any existing data mining algorithms for mining ordering rules. Depending on the specific problem, one may use different translation methods.

References 38.1 Cohen, W.W., Schapire, R.E. and Singer, Y. Learning to order things, Journal of Artificial Intelligence Research, 10, 243-270, 1999. 38.2 Greco, S., Matarazzo, B., and Slowinski, R. Rough approximation of a preference relation by dominance relations, European Journal of Operational Research 117, 63-83, 1999. 38.3 Iwinski, T.B. Ordinal information system, I, Bulletin of the Polish Academy of Sciences, Mathematics, 36, 467-475, 1988. 38.4 Pawlak, Z., Slowinski, R. Rough set approach to multi-attribute decision analysis, European Journal of Operational Research, 72, 443-359, 1994. 38.5 Tsumoto, S. Automated discovery of plausible rules based on rough sets and rough inclusion, Proceedings of PAKDD’99, 210-219, 1999.

38. On Mining Ordering Rules

321

38.6 Yao, Y.Y. Measuring retrieval effectiveness based on user preference of documents, Journal of the American Society for Information Science, 46, 133-145, 1995. 38.7 Yao, Y.Y. and Zhong, N. An analysis of quantitative measures associated with rules, Proceedings of PAKDD’99, 479-488, 1999.

39. Non-additive Measures by Interval Probability Functions Hideo Tanaka1 , Kazutomi Sugihara2 , and Yutaka Maeda3 1

2

3

Graduate School of Management and Information Science, Toyohashi Sozo College 20-1 Matsushita, Ushikawacho, 440-8511, Toyohashi, JAPAN [email protected], [email protected] Course of Applied Physics, Graduate School of Engineering Osaka University Yamadaoka 2-1, Suita, Osaka, 565-0871, JAPAN [email protected] Department of Industrial Engineering, College of Engineering Osaka Prefecture University Gakuencho 1-1, Sakai, Osaka, 590-8531, JAPAN [email protected]

Probability measures are well-defined ones that satisfy additivity. However, it is slightly tight because of its condition of additivity. Fuzzy measures that do not satisfy additivity have been proposed as the substitute measures. The only belief function involves a density function among them. In this paper, we propose two density functions by extending values of probability functions to interval values, which do not satisfy additivity. According to the definition of interval probability functions, lower and upper probabilities are defined, respectively. A combination rule and a conditional probability can be defined well. The properties of the proposed measure are clarified.

39.1 Introduction Probability theory is well defined for representing uncertainty under the assumption that a probability distribution is always determined from the given information. However this assumption is not satisfied with real situations in many problems. In the case where we can not determine only one probability distribution, it is appropriate that we speculate a set of probability distributions from an uncertain information given by estimators. There are many articles [39.2] [39.4] [39.6] [39.7] [39.8] [39.9] where an uncertain information has been handled by a set of distributions. These measures in the above papers do not satisfy additivity that is an important role in the conventional probabilities. Non-additive measures can be said to be a kind of fuzzy measures [39.12]. Fuzzy measures have been dealt with distribution functions, but density functions are not discussed yet in non-additive measures except for belief functions [39.10]. Belief functions and random sets are different basically from viewpoint of underlying theories. Nevertheless there is the method by which a belief function including the given random set can be obtained [39.11]. T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 22− 3 26, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

39. Non-additive Measures by Interval Probability Functions

323

In this paper, we propose two probability functions by extending values of probabilities to interval values, which do not satisfy additivity. This idea is similar to the concept of intuitionistic fuzzy sets [39.1] that can be said to be fuzzy rough sets [39.5]. According to the definition of interval probabilities, lower and upper probabilities are defined, respectively. A combination rule and a conditional probability can be defined well. The properties of the proposed measure are clarified.

39.2 Interval Probability Functions In this paper, interval probability functions denoted as IPF are proposed by two density functions. IPF is an extension of probability values to interval probability values. Definition 1. The set of two functions denoted as (h∗ , h∗ ) is called IPF if and only if (a) ∀x ∈ X, h∗ (x ) ≥ h∗ (x ) ≥ 0  h∗ (x) + (h∗ (x ) − h∗ (x )) ≤ 1 (b) x∈X

(c)



h∗ (x) − (h∗ (x ) − h∗ (x )) ≥ 1.

x∈X

The above (b) and (c) can be rewritten as  (b ) h∗ (x) + max (h∗ (x ) − h∗ (x )) ≤ 1  x

x∈X

(c )



h∗ (x) − max (h∗ (x ) − h∗ (x )) ≥ 1.  x

x∈X

Theorem 1. There exists a probability function h (x) that satisfies  h∗ (x) ≤ h (x) ≤ h∗ (x), h (x) = 1. (39.1) x∈X

Two distribution functions can be defined by IPF as follows. Let the lower and upper functions be denoted as LB(·) and U B(·) respectively. Definition 2. LB(·) and U B(·) can be defined as     h (x) LB(A) = min  h

x∈A

 U B(A) = max  h



(39.2)

 

h (x)

x∈A

where h∗ (x) ≤ h (x) ≤ h∗ (x).

(39.3)

324

H. Tanaka, K. Sugihara, and Y. Maeda

From Definition 2, the following theorem holds. Theorem 2. For ∀A ⊆ X, we have ⎛ ⎞   LB(A) = h∗ (x) ∨ ⎝1 − h∗ (x)⎠ x∈A

U B(A) =



⎛ h∗ (x) ∨ ⎝1 −

x∈A

x∈A



(39.4)

⎞ h∗ (x)⎠ .

(39.5)

x∈A

Theorem 3. The functions LB and U B are superadditive and subadditive, respectively. The proof is omitted because of the limited space. It follows clearly from Theorem 2 that the following dual relation holds. LB(A) = 1 − U B(A).

(39.6)

Let us consider the properties of IPF. Property 1. There is only one element such that its value of IPF is positive if and only if ∀ x ∈ (X − {x1 }), h∗ (x) = 0, h∗ (x1 ) = h∗ (x1 ) = 1.

(39.7)

Property 2. There are only two elements such that these values of IPF are positive if and only if ∀ x ∈ (X − {x1 , x2 }), h∗ (x) = 0, h∗ (x1 ) + h∗ (x2 ) = h∗ (x1 ) + h∗ (x2 ) = 1.

(39.8)

Property 3. There is no element such that an interval value is positive (h∗ (x) − h∗ (x)) > 0) if and only if it is a probability function. Property 4. There is no case such that only one element has an interval value. Property 5. There are only two elements such that interval values are positive (h∗ (x) − h∗ (x)) > 0) if and only if ∀ x ∈ (X − {x1 , x2 }), h∗ (x) = h∗ (x), h∗ (x1 ) + h∗ (x2 ) = h∗ (x1 ) + h∗ (x2 )   =1− h∗ (x) = 1 − x∈(X−{x1 ,x2 })

h∗ (x).

x∈(X−{x1 ,x2 })

These properties can be easily proved from the definition of IPF.

(39.9)

39. Non-additive Measures by Interval Probability Functions

325

39.3 Combination and Conditional Rules for IPF Let us consider a combination rule to combine two interval probability functions into one probability function. Definition 3. Let two interval density functions be denoted as (h1∗ (x), h∗1 (x)) and (h2∗ (x), h∗2 (x)). Then the combination rule is defined as h12∗ = h1∗ ∧ h2∗ h∗12 = h∗1 ∨ h∗2

(39.10) (39.11)

It is verified that the combined function (h12∗ (x), h∗12 (x)) is also IPF. This combination rule is proposed from viewpoint of possibility, although Dempster’s combination rule on belief functions [39.10] is defined from viewpoint of necessity. In belief measures, the combination rule entails the conditional rule, but in IPF the conditional rule is defined independently as follows. Definition 4. The lower and upper functions conditioned by B ⊂ X are defined as LB(AB) (39.12) LB(A|B) = LB(AB) + U B(B − AB) U B(AB) U B(A|B) = (39.13) U B(AB) + LB(B − AB) where U B(B) = 0 and for 00 we set LB(A|B) = 1 and U B(A|B) = 0. From the dual relation, we can see easily that LB(A|X) = LB(A) and U B(A|X) = U B(A). Using Definition 4, we can obtain two density functions as follows: h1∗ (x) = LB({x}|B) h∗1 (x) = U B({x}|B)

(39.14) (39.15)

where h1∗ (x) = h∗1 = 0 for x ∈ B. These two functions can be rewritten as follows: h1∗ (x) = LB({x}|B) = h (x) h0∗ (x) 0∗  ∨ h0∗ (x) + x ∈B−{x} h∗0 (x ) h0∗ (x) + x ∈B h0∗ (x )

(39.16)

h∗1 (x) = U B({x}|B) = h∗ (x) h∗0 (x) 0  ∨ ∗ ∗ ∗  h0∗ (x) + x ∈B−{x} h0 (x ) h0∗ (x) + x ∈B h0∗ (x )

(39.17)

where (h0∗ , h∗0 ) is a given IPF. Theorem 4. Two probability functions (h1∗ , h∗1 ) obtained by the above equations satisfy the definition of IPF.

326

H. Tanaka, K. Sugihara, and Y. Maeda

Here the proof of Theorem 4 is skipped. Conditional probability functions are IPF. The lower and upper functions based on IPF (h1∗ , h∗1 ) defined the above are denoted as LB1 (AB) and and U B1 (AB) respectively. Then, we have the following relation. LB1 (AB) ≤ LB1 (A|B) ≤ U B1 (A|B) ≤ U B1 (AB).

(39.18)

This means that the lower and upper functions obtained from probability functions induced by the conditional rule are wider than ones directly calculated by the conditional rule are.

39.4 Concluding Remarks IPF is useful to obtain interval weights in AHP [39.12]. The definition of IPF can be regarded as an extension of normalization of conventional probabilities. This research work is a first step for interval probability functions, but there are many problems with respect to IPF for future study.

References 39.1 K.T. Atanassov (1986): Intuitionistic fuzzy sets, Int. J. of Fuzzy Sets and Systems, 20, 87-96. 39.2 L.M. De Campos, M.T. Lamata and S. Moral (1990): The concept of conditional fuzzy measure, Int. J. of Intelligent Systems, 5, 58-67. 39.3 L.M. De Campos, J.F. Huets and S. Moral (1994): Probability interval; A tool for uncertain reasoning, Int. J. of Uncertainty, Fuzziness and KnowledgeBased Systems, 2, 167-196. 39.4 C. Choquet (1953): Theory of capacities, Ann. Inst. Fourier, 5, 131-295. 39.5 D. Coker (1998): Fuzzy rough sets are intuitionistic L-fuzzy sets, Int. J. of Fuzzy Sets and Systems, 96, 381-383. 39.6 A.P. Dempster (1967): Upper and lower probabilities induced by a multivalued mapping, Ann. Math. Stat., 38, 325-339. 39.7 D. Dubois and H. Prade (1986): A set-theoritic view of belief functions, Int. J. of General Systems, 12, 193-226. 39.8 J.F. Lemmer and H.E. Kyburg (1991): Conditions for the existence of belief functions corresponding to intervals of belief, Proc. 9th. National Conference on AI, 488-493. 39.9 Y. Pan and G.J. Klir (1997): Baysian inference based on interval valued prior distributions and likelihood, J. of Intelligent and Fuzzy Systems, 5, 193-203. 39.10 G. Shafer (1976): The Mathematical Theory of Evidence, Princeton Univ. Press. 39.11 M. Sugeno (1977): Fuzzy measures and fuzzy integrals; A survey, In M.M. Gupta, G.N. Saridis and B.R. Gaines(eds.), Fuzzy Automata and Decision Processes, 89-102. 39.12 K. Sugihara and H. Tanaka: Interval Evaluation in the Analytic Hierarchy Process by Possibility Analysis, J. of Computational Intelligence (to appear).

40. Susceptibility to Consensus of Conflict Profiles Ngoc Thanh Nguyen Department of Information Systems, Wroclaw University of Technology, Poland [email protected]

By a conflict profile we understand a set of data versions representing different opinions on some matter, generated by agents functioning in some sites of a distributed system. In purpose to solve this conflict the management system should determine a proper version of data for this matter. The final data version is called a consensus of given conflict profile. The main subject of this paper consists of consideration of existence and reasonableness of potential consensus. In other words, we consider problems related to consensus susceptibility of conflict profiles.

40.1 Introduction Consensus theory [1],[2] is useful in conflict solving. The resource of conflicts in distributed systems arises as the result of the autonomy feature of systems sites [3]. The simplest conflict takes place when two bodies have different opinions on the same subject. In work [5] Pawlak specifies the following elements of an one-value conflict: a set of agents, a set of issues, and a set of opinions of these agents on these issues. The agents and the issues are related with one another in some social or political context. Information tables [6] should be useful for representing this kind of conflicts. In this paper we define a consensus system which represents multi-value conflicts. In this system we distinguish conflict profiles containing versions of data which are generated by different participants of a conflict and refer to a conflict subject. Next consensus for conflict profiles is defined and two problems of susceptibility to consensus for profiles are considered.

40.2 Conflict Profiles For representing potential conflicts we use a finite set  A of attributes and a set V of attribute elementary values, where V = a∈A Va (Va is the do ) denote the set of subsets of set Va and main of attribute a). Let (V a    (VB ) = b∈B (Vb ). Let B ⊆ A, a tuple rB of type B is a function rB : B → (VB ) where (∀b ∈ B)(rb ⊆ Vb ). A tuple is elementary if all attribute values are empty sets or 1-element sets. Empty tuple is denoted by symbol φ. The set of all tuples of type B is denoted by T Y P (B) and T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 27− 3 3 2, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

328

N.T. Nguyen

the set all elementary tuples of type B is denoted by E − T Y P (B). We assume that some real world is commonly considered by agents which are placed in sites of a distributed system. The interest of the agents consists of events which occur (or have to occur) in the world. The task of the agents is based on determining the values of event attributes (an event is described by an elementary tuple of some type). The consensus system is defined as a triple Consensus Sys = (A, X, P), where: A is a finite set of attributes, which includes a special attribute Agent; values of attribute a are subsets of Va ; values  of attribute Agent are 1-element sets, which identify the agents; X = { (Va ) : a ∈ A} is a finite set of consensus carriers; P is a finite set of relations on carriers from X, each relation is of some type A (for A ⊆ A and Agent ∈ A). Relations belonging to set P are classified in such way that each of them includes relations representing similar events. For identifying relations belonging to given group the symbols ”+ ” and ”− ” should be used as the upper index. If P is the name of a group, then relation P + is called a positive relation (contains positive knowledge) and P − is the negative relation (contains negative knowledge). The structures of the consensus carriers is defined as a distance function between tuples of the same type. This function can be defined on the basis of one of distance functions δ P and ρP between sets of elementary values [4].  Definition 40.2.1. For 2 tuples r and A the distance function ϕ  r of type 1  ϕ(ra , ra ) where ϕ ∈ {ρP , δ P }. assigns a number ∂(r, r ) = card(A) a∈A

Consensus is considered within a consensus situation, defined as follows: Definition 40.2.2. A consensus situation is a pair {P + , P − }, A → B! where A, B ⊆ A, A ∩ B = ∅ and for every tuple r ∈ P + ∪ P − there should be held rA = φ. The first element of a consensus situation includes the domain from which consensus should be chosen, and the second element presents the subjects of consensus (i.e. set Subject(s) ⊆ E − T Y P (A)) and the content of consensus, such that for a subject e there should be assigned only one tuple of type B. For each subject e 2 conflict profiles prof ile(e)+ , prof ile(e)− ⊆ T Y P (A ∪ B) should be determined. Definition 40.2.3. Consensus on subject e ∈ Subject(s) is a set of 2 tuples {C(s, e)+ , C(s, e)− } of type A ∪ B which fulfill the following conditions: − a) C(s,  e)+ A = C(s, e)A = e, b) ∂(rB , C(s, e)+ B ) and r∈prof ile(e)+

 r∈prof ile(e)−

∂(rB , C(s, e)− B ) are minimal,

− c) C(s, e)+ B ∩ C(s, e)B = φ.

Any tuples C(s, e)+ and C(s, e)− satisfying conditions a)-b) are called consensuses of profiles prof ile(e)+ and prof ile(e)− respectively.

40. Susceptibility to Consensus of Conflict Profiles

329

40.3 Susceptibility to Consensus In this section we investigate two problems referring to susceptibility to consensus for conflict profiles. For given situation s = {P + , P − }, A → B! there may exit two following problems: 1. For given subject e ∈ Subject(s) when any tuples C(s, e)+ , C(s, e)− ∈ T Y P (A ∪ B) satisfying conditions a)-b) specified in Definition 4 (that is being consensuses of profiles prof ile(e)+ and prof ile(e)− respectively) may create a consensus, that means when the last condition C(s, e)+ B ∩ = φ may be fulfilled? C(s, e)− B 2. If consensus of a profile exists, is it good enough for this profile? For explaining the first problem we give the following example: Example 40.3.1. In the meteorological system [3] let us consider the following situation s = Rain+ , Rain− , Region → T ime! where determined from relations Rain+ and Rain− profiles prof ile(e)+ and prof ile(e)− for e = Region : r1 ! have the following form: prof ile(e)− prof ile(e)+ Agent Region Time Agent Region Time a1 r1 3a.m.-5a.m. a1 r1 2a.m.-6a.m. a2 r1 4a.m.-6a.m. a2 r1 4a.m.-5a.m. a3 r1 4a.m.-5a.m. a3 r1 6a.m.-8a.m. Using distance function ρP we obtain the following tuples C(s, e)+ and C(s, e)− of type {Region, T ime}, which fulfill conditions a)-b) of Definition 4: C(s, e)+ = Region : r1 , T ime : 4a.m. − 5a.m.!, C(s, e)− = Region : r1 , T ime : 4a.m. − 6a.m.!. Let us note that these tuples do not satisfy the − condition c) of Definition 4, because C(s, e)+ {T ime} ∩ C(s, e){T ime} = φ, thus they can not create a consensus for subject e. The second problem is the main subject of this section. It often happens that for a given conflict it is possible to determine consensus. The question is: is the chosen consensus good enough and can it be acceptable as the solution of given conflict situation? In other words, is the conflict situation susceptible to (good) consensus? We will consider the susceptibility to consensus for conflict profiles. Before defining the notion of susceptibility to consensus below we present an example. Example 40.3.2. Let a space (U, ∂) be defined as follows: U = {a, b} where a and b are tuples of some type, and distance function ∂ is given as: For x, y ∈ U ∂(x, y) = 0 if x = y and ∂(x, y) = 1 otherwise. Let X be a profile being a set with repetitions, where X = {50 · a, 50 · b}. Assume that X represents the result of some voting, in which 100 agents take part, each of them gives one vote (for a or b). There are 50 votes for a and 50 votes for b. It is easy to note that for profile X the consensus should be equal to a or b, but it intuitively seems that none of them is a good consensus, because there is lack of a compromise in this conflict situation. Let us consider now another

330

N.T. Nguyen

profile X  = {50 · a, 51 · b}. For this profile the only consensus should be b and it seems to be a good consensus, that means this profile is susceptible to consensus. The above example shows that although consensus may always be chosen for a conflict profile, it does not have to be a good one. We define below the notion of profile’ susceptibility to consensus. Let X ∈ {prof ile(e)+ , prof ile(e)− } for e ∈ Subject(s), card(X) = n and 

n  ˆ B , yB ) ∂(x

ˆ ∂(X) = ∂ˆmin (X) =

x,y∈X

ˆ X) = , ∂(x,

n(n + 1) ˆ X), ∂ˆmax (X) = min ∂(x,

x∈T Y P (B)

∂(xB , yB )

y∈X

n max

, ˆ X). ∂(x,

x∈T Y P (B)

ˆ X) = ∂(y, ˆ X) Profile X is regular if for each x, y ∈ X equality ∂(x, follows, profile X is irregular if there exist two its elements x and y such that ˆ X) = ∂(y, ˆ X) . ∂(x, ˆ Definition 40.3.1. Profile X is susceptible to consensus iff ∂(X) ≥ ∂ˆmin (X). 2·50·50 50 ˆ For the profiles defined in Example 2 we have ∂(X) = 100·101 = 101 < 50 ˆ ∂min (X) = 100 . Thus profile X should not be susceptible to consensus. It is agreed with intuition because neither a nor b should be a ”good consensus” ˆ  ) = 2·50·51 = 50 = ∂ˆmin (X  ). Thus profile for this profile. However ∂(X 101·102 101  X should be susceptible to consensus. According to intuition b should be a ”good” consensus because it dominates a in profile X  . Below we present some properties of consensus susceptibility. Theorem 40.3.1. If X is a regular profile and card(X) > 1 then X is not susceptible to consensus. Profile X in Example 2 is a regular one, therefore it is not susceptible to consensus. Let symbol ∪˙ represent the sum operation on sets with repetitions, we have the following: Theorem 40.3.2. Let X and X  be such conflict profiles that card(X) > 1, ˙ for some x ∈ X and X is regular, then profile X  should be X  = X ∪{x} susceptible to consensus. Notice that profiles X and X  in Example 2 satisfy the conditions in Theorem 2, so, as stated, X  should be susceptible to consensus. Theorem 2 shows also that if profile X is regular then its extending by some element of itself gives a profile which should be susceptible to consensus. The practical sense of this theorem is that if in given conflict situation none of votes dominates and in the second voting extended by one voter who gives his vote for one of the previous ones, then the new profile should be susceptible to consensus. For given conflict profile X, where X ∈ {prof ile(e)+ , prof ile(e)− }, which elements are tuples of type B let Occ(X, x) for x ∈ E − T Y P (B) denote the number of occurrences of elementary tuple x in tuples belonging to X. Let card(X) = n and

40. Susceptibility to Consensus of Conflict Profiles M =

 y∈E−T Y P (B)

2Occ(X, y)(n − Occ(X, y))py ;

X1 = {x ∈ E − T Y P (B) : Occ(X, x) =

n }; 2

X2 = {x ∈ E − T Y P (B) : 0 < Occ(X, x) < X3 = {x ∈ E − T Y P (B) :

n 2

M1 = n }; 2

M2 =

< Occ(X, x) < n}; M3 =

 y∈X 1 y∈X 2

331

n p ; 2 y

Occ(X, y)py ; (n − Occ(X, y))py

y∈X3

for py = 1 if function ρP is used and py = d(y) if function δ P is used, where d(y) is the cost of adding (moving) elementary tuple to (from) profile X. Theorem 40.3.3. If distance functions δ P and ρP are used for determining consensus then the following dependencies are true: a) If n is an odd number then profile X is always susceptible to consensus, b) If n is an even number then profile X is susceptible to consensus if and only if M1 + M2 + M3 ≤ M/(n + 1). Theorem 3 allows to state if a given profile is susceptible to consensus or not without determining the consensus. It has been pointed out that if the number of agents taking part in the conflict is odd then the profile is always susceptible to consensus, and if this number is even then some condition must be satisfied.

40.4 Conclusions In this paper some results of investigation on the problems related to specifying conditions which allow to find out if a conflict profile is susceptible to consensus, are presented. The future work should concern the first problem specified in Section 3. Its solution should allow us to find out if a conflict situation is consensus-oriented or not. Another interesting aspect of the consensus model is introducing probabilistic elements to conflict content to extend the possibilities for agents for their opinion representation. In this case the tools which enable to join rough set theory and probabilistic calculus [7] should be useful.

References 40.1 Arrow, K.J.: Social Choice and Individual Values. Wiley New York (1963) 40.2 Barthelemy, J.P., Janowitz M.F.: A Formal Theory of Consensus. SIAM J. Discrete Math. 4 (1991) 305-322. 40.3 Nguyen, N.T.: Using Consensus Methods for Solving Conflicts of Data in Distributed Systems. Lecture Notes on Computer Science 1963. Springer– Verlag (2000) 409-417. 40.4 Nguyen, N.T.: Conflict Profiles’ Susceptibility to Consensus in Consensus Systems. Bulletin of Int. Rough Sets Society 5, No. 1/2 (2001) 217-224 40.5 Pawlak, Z.: An Inquiry into Anatomy of Conflicts. Journal of Information Sciences 108 (1998) 65-78.

332

N.T. Nguyen

40.6 Skowron, A., Rauszer, C.: The Discernibility Matrices and Functions in Information Systems. In: E. Slowi´ nski (ed.): Intelligent Decision Support, Handbook of Applications and Advances of the Rough Sets Theory, Kluwer Academic Publishers (1992) 331-362. 40.7 Tsumoto, S., Tanaka, H.: Introduction of Probabilistic Rules based on Rough Set Theory. Lecture Notes on Artificial Intelligence 744. Springer–Verlag (1993) 441-448.

41. Analysis of Image Sequences for the Unmanned Aerial Vehicle∗ Hung Son Nguyen, Andrzej Skowron, and Marcin S. Szczuka Institute of Mathematics, Warsaw University Banacha 2, 02–097, Warsaw, Poland {son,skowron,szczuka}@mimuw.edu.pl

A method for extracting relevant information from image sequence data is presented. The image sequences, being output of video system of the Unmanned Aerial Vehicle, are analysed with use of EM-clustering techniques and Rough Set based methods. The possibilities of construction of an automated system for recognition/identification of cars on the road, on the basis of colour-related data are discussed.

41.1 Introduction The issue of constructing and controlling an autonomous Unmanned Aerial Vehicle (UAV) is a multi-fold one. The idea of constructing such a vehicle (helicopter) for the purposes of traffic control drives the WITAS project (see [41.8]). Apart of difficulties in construction of proper hardware the problem of establishing software is a challenging one. The UAV is supposed to recognise the road situation underneath on the basis of sensor readings and make the decision about acts that are to be performed. The issue of constructing adaptive, intelligent and versatile system for identification of situation was addressed in [41.5]. In the paper we focus on one of the subtasks necessary for the entire system to work – the problem of discerning between objects that are visible to the UAV. The most crucial information for UAV is provided by its video systems. We have to be able to provide UAV control system with information about car colors and so on. Such information may allow for making the identification that is core for operations performed by UAV, such as tracking a single vehicle over some time. In the paper we address only a part of issues that have to be resolved. The particular task we are dealing with is identification of techniques that may be used for the purpose of discerning and/or classifying objects from image sequence data. Given a series of images gathered by UAV’s video system we have to extract the valuable information about cars present in the image. The key is to have compact set of features that at the same time are robust the image data may be heavily distorted. The unwanted effects coming from changes in UAV’s position, lighting conditions, scaling, rotation and weather conditions have to be compensated. ∗

Supported by the Wallenberg Foundation and grant KBN 8T11C02519.

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 3 3 − 3 3 8 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

334

H.S. Nguyen, A. Skowron, and M.S. Szczuka

41.2 Data Description At the current stage we are dealing with two sequences of images consisting of 100 frames each. They represent two situations on the road, each about 4 second long. Every frame is a 24 bit .tiff image with resolution 726 × 512 pixels. The image sequences have been manually interpreted. Altogether 18 objects representing cars on the road have been identified The object instance (colour blob) is represented with 30 attributes. number (identifier) assigned to an object, two numerical attributes representing X and Y coordinates (within an image) of the center of colour blob (object) and 27 attributes representing coordinates in the RGB colour space for 9 pixels being a 3 × 3 matrix surrounding the center of colour blob. For each of 18 identified object we have 100 instances, one for each image in sequence (1800 samples in total).

41.3 The Task The overall problem of situation identification on the basis of image (and possibly some other) data is very compound. In the first stage, described in this paper we would like to find the answers to the following questions: 1. Is the existing amount of information (27 colour-related attributes) sufficient for construction of classification support system that is able to distinguish between 18 pre-identified objects? 2. Is it possible to transform the existing 27 dimensional attribute space to the form better supporting car colour classification tasks? 3. Is it possible to learn the basic concepts allowing for establishment of prototypes rules of classification provided we have part of the sequence, say first 50 images, and then classify objects for the rest of sequence properly?

41.4 The Method Initially, an attempt to perform car (colour blob) dissemination and/or classification with use of typical methods from the Rough Set armory (see [41.4]) have been made. Unfortunately, it turned out that the data is too vague and distorted for the typical tools like RSES ([41.7]) or Rosetta ([41.6]). We came to the conclusion that some method for extraction of more relevant features from the raw data is needed. Therefore we turned our attention at unsupervised learning methods that allow for identification of characteristic features of objects in the corpus. The main intention is to eliminate unwanted effects caused by changes in object RGB colours as the object (car) moves between zones of different light. The particular approach we apply uses clustering and simple time series analysis.

41. Analysis of Image Sequences for the Unmanned Aerial Vehicle

335

First, we perform clustering treating all 1800 measurements as points in 27 dimensional space (9 points×3 RGB coordinates). To do the clustering we utilise Expectation Maximisation (EM) method. EM-clustering is an iterative, unsupervised clustering method aimed at establishment of possibly small number of not intersecting clusters constructed with assumptions about normal distribution of objects. For details about EM clustering see [41.1] and [41.3]. After the clustering have been found we recall the information about sequential character of our data. Namely, we analyse the sequences of cluster assignments for each of 18 cars. Going frame-by-frame we check to which clusters the object belong in scope of this frame. In this way for each of 18 cars we get a vector of 100 cluster assignments. Such vectors may be compared and on the basis of differences between them we may discern one car from the others.

41.5 Results The clustering was applied to the entire data. As a result of several experiments we got 15 to 18 clusters on the average. For all objects the assignment to cluster was very characteristic. In most cases it was possible to distinguish 2-3 clusters to which the samples corresponding to the single cases were assigned. These 2-3 clusters contained more than 80% of car on the average. Moreover, it was possible to correlate the change of cluster assignments with changes in lighting of car on the road. As the car enters the area of shadow, the visual perception of its colour is changing and so its cluster assignment. This effect is very welcome from our point of view since it makes clear evidence of cluster relevance. On the basis of clustering new features were constructed for the objects. For each object (car) Ci (i = 1, ..., 18) we construct new attributes na1 , ..., nac where c is the number of the clusters derived. The value of attribute naj for the car Ci is the number of occurrences of an object representing i-th car in j-th cluster. So, if the value of attribute na1 for car C1 is 20 then we know that an object corresponding to this car was assigned to first cluster 20 times out of hundred. This new set of attributes undergone further analysis. By applying Rough Set based techniques it was possible to find out that attributes derived from clusters are sufficient for discernibilty. Namely, it was possible, with the use of RSES software (see [41.7]), to calculate a set of if..then.. decision rules classifying (discerning) the cars. In this way we got a simple set of rules such that there was exactly one rule for each of 18 cars. Since clustering have led us to so promising results in terms of ability for object dissemination, we tried to exploit its potential to the limit. Since the clustering process takes some time in case of 1800 objects and 27 numerical attributes we were looking for the way to make it simpler. Reduction of computational effort is in our case very important since major part of recognition

336

H.S. Nguyen, A. Skowron, and M.S. Szczuka

process has to performed on-line, during UAV operation. We found out that the clustering-based approach is quite powerful. We performed an experiment using reduced information about colour blobs. Instead of 27 attributes representing three RGB coordinates of 9 points (3×3 matrix) we take only three. These three are averages over 9 points for Red, Green and Blue coordinate values respectively. For this reduced set of features we obtained a clustering and it was still possible to have good discrernibility between objects. Moreover, the time needed for computation was reduced several times. The results presented above address the question about amount of useful information that can be retrieved from image sequences. The other question on our task list was the one about potential abilities for construction of classification system. Initial experiments aimed at construction of classification method based on inductive learning of concepts were performed. We wanted to check what are the possibilities to create a system that will be able to classify previously unseen objects as being similar to the prototypes learned during presentation of training sample. For this purpose we first split our set of examples into halves. One half, used for training, contains first 50 samples for each car i.e. frames 1 to 50 from both image sequences. The remaining 50 frames from each sequence form the dataset used for testing. On the basis of training set we establish clustering-based features and decision rules using these features. Then we take a sample from the testing set and label them with the car numbers. In the experiments we use simplified version of cluster-based attributes presented above. Instead of attributes na1 , ..., nac for training samples we take binary attributes ma1 , ..., mac . Attribute ma1 for a given sample is equal to 1 iff na1 > 0 for this sample, and 0 otherwise. Since we have to check abilities of classification system we start first with the learning phase. Learning of classification (decision) rules is done on the basis of 18 samples of 50 frames each. So the learning data consists of 18 objects, each object described by c attribute values, where c is the number of clusters. First attempt was performed for testing samples consisting of entire 50 remaining frames. By matching those examples against previously created clusters, producing cluster-based attributes and the assigning decisions (car numbers) to the samples we got the result for training sample. In this particular experiment we got a perfect accuracy (100%). Unfortunately, taking 50 frames requires approximately two seconds which is too long for real-time application. Therefore, we would like to be able to reduce the number of frames in testing sample to no more than 15-20 and still retain good classification ratio. To do that we process our test data and produce testing samples with use of moving window. We set a size of the window to be some integer not greater than 50. Then from 50 frames we produce the testing sample by taking as many sequences of the size of window as possible and calculate cluster-related

41. Analysis of Image Sequences for the Unmanned Aerial Vehicle

337

attributes ma1 , ..., mac for them. For instance, if we set the size of the window to be 15 then we will get 35 samples for each car. First of these samples will contain frames from 51 to 66 while the last will consist of frames 86 to 100. So, altogether for 18 cars we will get 830 testing instances. The key is now to find the size of the window to at the same time small enough to allow on-line classification and big enough to have good quality of this classification. From several attempts we have learned that with the methods of attribute generation and decision rule derivation depicted above, we are able to get perfect accuracy of classification for testing sample if the size of the window exceeding 17. For the window size less than 17 the accuracy decreases, being 89% and 78% for the windows of size 16 and 15, respectively. It is worth mentioning that these experiments are, at the moment of writing, only initially finished. We expect to improve the results by allowing more information to be passed to classifier e.g. by using the original attributes na1 , ..., nac instead of simplified ma1 , ..., mac .

41.6 Conclusions The method for extracting information from image sequences was presented. It is based on combination of unsupervised clustering with Rough Set based approach. From the initial experiment we may see that this approach has a significant potential and may be further developed into complete solution. The proposed method have to be tuned to fit the requirements for co-operation with other components of UAV’s control system as well as expectations about robustness, versatility and speed of operation. The natural next step is the application of developed solutions to other sets of image data. We expect that some further evolution of the methods will be necessary, since many problems may arise. We believe that with more data we will be able to generalise our approach using tools such as more compound time series analysis.

References 41.1 Ambroise C., Dang M., Govaert G. (1997), Clustering of Spatial Data by the EM Algorithm.In: A. Soares et al. (eds), geoENV I-Geostatistics for Environmental Applications, Kluwer, Dordrecht, pp. 493–504. 41.2 Backer E., Kandel A.(1995), Computer-Assisted Reasoning in Cluster Analysis, Prentice Hall, New York. 41.3 Celeux G., Govaert G. (1992) A classification EM algorithm for clustering and two stochastic versions, Computational Statistics and Data Analysis, 14, pp. 315 – 332. 41.4 Komorowski J., Pawlak Z., Polkowski L., Skowron A.(1999). Rough sets: A tutorial. In: S.K. Pal and A. Skowron (eds.), Rough fuzzy hybridization: A new trend in decision-making, Springer-Verlag, Singapore, pp. 3–98.

338

H.S. Nguyen, A. Skowron, and M.S. Szczuka

41.5 Nguyen H.S., Skowron A., Szczuka M. (2000), Situation Identification by UAV, Proceedings of CS&P 2000, Informatik Berichte, Humboldt-Univerit¨ at zu Berlin 41.6 The Rosetta Homepage – www.idi.ntnu.no/∼aleks/rosetta 41.7 The RSES Homepage – logic.mimuw.edu.pl/∼rses 41.8 The WITAS project Homepage –www.ida.liu.se/ext/witas/

42. The Variable Precision Rough Set Inductive Logic Programming Model and Web Usage Graphs V. Uma Maheswari, Arul Siromoney, and K.M. Mehata Anna University, Chennai 600 025, India [email protected]

42.1 Introduction Inductive Logic Programming [42.1] is the research area formed at the intersection of logic programming and machine learning. Rough set theory [42.2, 42.3] defines an indiscernibility relation, where certain subsets of examples cannot be distinguished. The gRS–ILP model [42.4] introduces a rough setting in Inductive Logic Programming and describes the situation where the background knowledge, declarative bias and evidence are such that it is not possible to induce any logic program from them that is able to distinguish between certain positive and negative examples. Any induced logic program will either cover both the positive and the negative examples in the group, or not cover the group at all, with both the positive and the negative examples in this group being left out. The Variable Precision Rough Set (VPRS) model [42.5] is a generalized model of rough sets that inherits all basic mathematical properties of the original rough set model but allows for a controlled degree of misclassification. The Variable Precision Rough Set Inductive Logic Programming (VPRSILP) model [42.6] extends the gRS–ILP model using features of the VPRS model. This paper applies the VPRSILP model to graphs, and presents the results of an illustrative experiment on web usage graphs.

42.2 The VPRSILP Model and Web Usage Graphs The generic Rough Set Inductive Logic Programming (gRS–ILP) model introduces the basic definition of elementary sets in ILP [42.4, 42.7]. A parameter β, a real number in the range (0.5, 1], is used in the VPRSILP model [42.6] as a threshold in elementary sets that have both positive and negative examples, to decide if that elementary set can be classified as positive or negative. A standard ILP algorithm GOLEM [42.8] is modified in [42.6] to fit this model. The formal definitions and the modified algorithm are omitted here due to space constraints.

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 3 9− 3 4 3 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

340

V.U. Maheswari, A. Siromoney, and K.M. Mehata

42.2.1 A Simple–Graph–VPRSILP–ESD System Let U be a universe of examples. Let a graph Gx be associated with every x in U . Let GU = (NU , EU ) be a graph associated with the universe such that Gx , for every x ∈ U , is a subgraph of GU . NU is the set of nodes and EU is the set of links of GU . Definition 42.2.1. We define a simple–graph–VPRSILP–ESD system as a 2–tuple S = (S  , GU ), where: (1) GU is a directed graph, and (2) S  = (E, B, L, β) is a VPRSILP–ESD system [42.6] such that (i) E is the universe of examples consisting of a unary predicate, say p. Each example p(x) has a directed graph Gx associated with it which is a subgraph of GU (ii) B is the background knowledge consisting of ground unit clauses, using the following predicate: edge (of arity 3) where for any p(x) ∈ E : edge(x, sourcenode, destnode) ∈ B ⇒ the graph associated with the example p(x) has an edge from the sourcenode to destnode. (iii) L is the declarative bias L = Lpi ∧ Lrd ∧ Leu , (defined in [42.7]) (iv) β is a real number in the range (0.5, 1]. Our aim is to find a hypothesis H such that P = B ∧ H ∈ Pβ (S), the β–restricted program space [42.6]. 42.2.2 Web Usage Graphs We now consider an example of a simple–graph–VPRSILP–ESD system using Web usage graphs. Let us consider a particular set of Web pages and links between them. Each node in NU corresponds to one of these Web pages and each link in EU corresponds to one of these links. A single session x when a user enters any Web page in NU till the user finally leaves the set of Web pages in NU is a subgraph of GU = (NU , EU ), denoted by Gx = (Nx , Ex ). The universe U is considered to be the set of all such sessions. A session x ∈ U is a positive example (x ∈ X) or a negative example (x ∈ (U − X)) based on some concept of interest (X). The notion of Posgraph and Neggraph (that cumulatively capture all known positive and negative sessions) is introduced in [42.9]. Using Posgraph (Neggraph), edges that are distinctly present in the positive (negative) sessions are obtained. These edges are considered to be important and used in predicting unknown sessions. Let GU be the weighted directed graph representing the Web pages and links under consideration. Let E = E + ∪ E − , E + = {p(e1), p(e2)}, E − = {p(e3)}. Let B = {edge(e1, //m.org/a, //m.org/b), edge(e1, //m.org/c, //m.org/d), edge(e2, //m.org/c, //m.org/d), edge(e3, //m.org/b, //m.org/c)}.

42. The Variable Precision Rough Set

341

S = (S  , GU ), where S  = (E, B, Lpi ∧ Lrd ∧ Leu , β), is a simple–graph– VPRSILP–ESD system. One induced hypothesis H, such that P = H ∧ B ∈ Pβ (S), for β = 0.5, is of the form p(X):-edge(X, //m.org/a, //m.org/b),edge(X, //m.org/c, //m.org/d).

It is seen that for P = B ∧ H, P " p(e1), P " p(e2), P " p(e3).

42.3 Experimental Illustration The dataset used in our experiment is taken from the website http://www.cs.washington.edu/research/adaptive/download.html and is the data set used in [42.10, 42.11]. The data pertains to web access logs at the site http://machines.hyperreal.org during the months of September and October 1997. Each day of the month has a separate file. Each file records all the requests for Web pages made to the Web server on that particular day. Sessions with less than 3 edges or more than 499 edges were not considered. The dataset (U ) is divided into positive example sessions (X) and negative example sessions (U − X). As an illustration, all sessions that had an access from www.paia.com are treated as positive examples and all sessions that had access from www.synthzone.com as negative examples. The data is first preprocessed to determine a set of useful edges based on the number of positive and negative sessions traversing the edges. The universal graph GU consists of these useful edges. Each elementary set corresponds to the set of examples whose session graphs have the same set of edges in GU . The training set is used to determine the number of positive and negative examples in each elementary set. In the modified GOLEM algorithm [42.6] all elementary sets covered by any rule fall within the β–positive region. A β value of 0.5 is used in this experiment. The two counters in each elementary set are used to calculate the conditional probability, and hence to determine whether the elementary set is in the β–positive region or β–negative region. The modified GOLEM algorithm is implemented with the following changes. (1)For each session, the corresponding elementary set is determined based on which of the useful edges are traversed in that session. (2)The maximal common subgraph between example sessions is used instead of rlgg. (3) Every example is used rather than a random subset. (4) The innermost loop is not implemented, since every example is being considered. Ten fold cross validation was done by using days ending with 0, 1, 2, . . . 9 as the ten sets. The experiment consists of two separate runs. The first run uses the original positive and negative examples, whereas the second run uses the original negative examples as the positive examples, and the original positive examples as the negative examples. In other words, the positive and negative examples are inverted in the second run. The results of the ten fold cross validation in the original run (original positive and negative examples) are tabulated below.

342

V.U. Maheswari, A. Siromoney, and K.M. Mehata

Serial 0 1 2 3 4 5 6 7 8 9 Average

Positives Correct Wrong 20 33 24 19 18 32 24 47 20 42 34 27 30 32 18 34 19 35 29 34 23.6 33.5

Negatives Correct Wrong 294 0 134 2 215 0 285 0 219 1 209 0 294 1 192 0 205 0 320 3 236.7 0.7

The results of the ten fold cross validation in the inverted run are tabulated below. The original positive and negative examples are inverted and are used as negative and positive examples, respectively. Serial 0 1 2 3 4 5 6 7 8 9 Average

Positives Correct Wrong 68 226 33 103 56 159 83 202 47 173 62 147 66 229 58 134 55 150 63 260 59.1 178.3

Negatives Correct Wrong 52 1 42 1 50 0 71 0 62 0 60 1 62 0 51 1 51 3 62 1 56.3 0.8

The average results in the two runs are tabulated below. The following table is the average result of the ten–fold cross-validation on the original positive and negative examples. Actually Positive Actually Negative

Pred. Pos 23.6 0.7

Pred. Neg 33.5 236.7

The following table is the average result of the ten–fold cross-validation on the inverted positive and negative examples. It is to be noted that the positives and negatives reported in the table are the original positives and negatives (i.e. reinverted from those used in the actual inverted run). Actually Positive Actually Negative

Pred. Pos 56.3 178.3

Pred. Neg 0.8 59.1

42. The Variable Precision Rough Set

343

It is seen from the tables, that in the original run, if a test case is predicted positive, it has 97.1% chance of being positive; and in the inverted run, if a test case is predicted as original negative, it has 98.7% chance of being an original negative. (This high degree of accuracy of prediction applies to the 41.3% of the positive test cases that are predicted positive and the 24.9% of the negative test cases that are predicted negative.)

42.4 Conclusions The VPRSILP model is applied to Web usage graphs. An illustrative experiment on the prediction of Web usage sessions is presented. Possibilities for further work include the application of the VPRSILP model to other ILP algorithms and to other application areas.

References 42.1 S. Muggleton. Inductive logic programming. New Generation Computing, 8(4):295–318, 1991. 42.2 Z. Pawlak. Rough sets. International Journal of Computer and Information Sciences, 11(5):341–356, 1982. 42.3 Z. Pawlak. Rough Sets — Theoretical Aspects of Reasoning about Data. Kluwer Academic Publishers, Dordrecht, The Netherlands, 1991. 42.4 A. Siromoney and K. Inoue. The generic Rough Set Inductive Logic Programming (gRS-ILP) model. In T.Y. Lin et al, editor, Granular Computing and Data Mining. Physica Verlag, 1999. 42.5 W. Ziarko. Variable precision rough set model. Journal of Computer and System Sciences, 46(1):39–59, 1993. 42.6 V. Uma Maheswari, Arul Siromoney, K. M. Mehata, and K. Inoue. The Variable Precision Rough Set Inductive Logic Programming Model and Strings. Computational Intelligence, 2001. Accepted for publication. 42.7 A. Siromoney and K. Inoue. Elementary sets and declarative biases in a restricted gRS–ILP model. Informatica, 24:125–135, 2000. 42.8 S. Muggleton and C. Feng. Efficient induction in logic programs. In S. Muggleton, editor, Inductive Logic Programming, pages 281–298. Academic Press, 1992. 42.9 V. Uma Maheswari, Arul Siromoney, and K. M. Mehata. Mining web usage graphs. In Knowledge Based Computer Systems (KBCS2000), pages 186–192. Allied Publishers Limited, 2000. 42.10 Jos´e Borges and Mark Levene. Mining association rules in hypertext databases. In Proc. Fourth Int. Conf. on Knowledge Discovery and Data Mining, pages 149–153, 1998. 42.11 Mike Perkowitz and Oren Etzioni. Towards adaptive web sites: Conceptual framework and case study. In The Eighth Int. World Wide Web Conference, Toronto, Canada, May 1999.

43. Optimistic Priority Weights with an Interval Comparison Matrix Tomoe Entani1 , Hidetomo Ichihashi1 and Hideo Tanaka2 1

2

Graduate School of Engineering, Osaka Prefecture University, 1–1 Gakuen-cho, Sakai, Osaka, 599–8531 Japan [email protected] Graduate School of Management and Information, Toyohashi Sozo College, 20–1 Matsushita, Ushikawa-cho, Toyohashi, Aichi, 440–8511

AHP is proposed to give the importance grade with respect to many items. The comparison value is used to be crisp, however, it is easy for a decision maker to give it as an interval. The interval comparison values can reflect uncertainty due to human judgement. In this paper, the interval importance grade is obatained from an interval comparison matrix so as to include the decision maker’s judgement. To choose the crisp importance grades and the crisp efficinency in the decision maker’s judgement, we use DEA, which is an evaluation method from the optimistic viewpoint.

43.1 Introduction AHP (Analytic Hierarchical Process) is proposed to determine the importance grades of each item [43.1]. AHP is a method to deal with the importance grades with respect to many items. In conventional AHP, the crisp importance grade of each item can be obtained by solving eigenvector problem with a crisp comparison matrix. Since a decision maker’s judgement is uncertain and it is easier for him/her to give it as an interval value than to give it as a crisp value, we extend the crisp comparison values to intervals. Based on the idea that a comparison matrix is inconsistent due to human judgements, the model that gives the importance grade as an interval is proposed [43.2]. We take another way to obtain the interval importance grades based on eigenvector method and interval regression analysis[43.3]. When a decision maker gives comparison matrices for input and output items, the interval importance grades of input and output items are obtained respectively. The obtained interval importance grades can be considered as the acceptable importance grades for a decision maker. We choose the most optimistic importance grades for the analyzed object in the interval by DEA (Data Envelopment Analysis) [43.4][43.5]. DEA is a well-known method to evaluate DMUs (Decision Making Units) from the optimistic viewpoint. The weights in DEA and the importance grades through AHP are similar then DEA is used to choose the most optimistic importance grades of input and output items in the decision maker’s acceptable ranges[43.6]. Our aim is to choose the importance grades in a possible ranges which are estimated from a decision maker’s judgement. T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 4 4 − 3 4 8 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

43. Optimistic Priority Weights

345

43.2 Interval AHP with Interval Comparison Matrix When a decision maker compares a pair of items for all possible pairs with n items, I1 , ..., In , we can obtain a comparison matrix A as follows. A decision maker’s judgement is usually uncertain. Therefore, it is more suitable to give the comparison values as intervals. ⎛ ⎜ A=⎝

L

1 .. . an1 , U an1



L

··· aij , U aij ···

L 

⎞ a1n , U a1n ⎟ .. ⎠ . 1

where the element of matrix A, [L aij ,U aij ], shows the importance grade of Ii obtained by comparing with Ij , the diagonal elements are equal to 1, that is L aii =U aii = 1 and the reciprocal property is satisfied, that is L aij = 1/U aji . Then, we estimate the importance grade of item i, as an interval denoted as Wi , that is determined by its center wci and its radius di as follows. Wi = [L wi , U wi ] = [wci − di , wci + di ] In order to determine interval importance grades, we have two problems where one is to obtain the center and the other is to obtain the radius. The center is obtained by eigenvector method with the obtained comparison matrix A. Since the elements of A are intervals, their centers are used. The eigenvector problem is formulated as follows. Aw = λw

(43.1) (wc1 , . . .

, wcn )

Solving (43.1), the eigenvector for the principal eigenvalue λmax is obtained as the center of theinterval importance grade of each item. n c∗ The center wc∗ i is normalized to be i=1 w i = 1. The radius is obtained based on interval regression analysis, which is to find the estimated intervals to include the original data. In our problem, aij is approximated as an interval ratio such that the following relation holds. ! c∗ " wi −di wc∗ Wi i +di [L aij ,U aij ] ⊆ W = , (43.2) c∗ c∗ w +dj w −dj j j

j

where W i /W j is defined as the maximum range. The interval importance grades are determined to include the interval by (43.1), the radius comparison values. Using the obtained centers wc∗ i should be minimized subject to the constraint conditions that the relation (43.2) for all elements should be satisfied. min λ wc∗ i − di ≤ L aij , s.t. c∗ w j + dj di ≤ λ, (∀i)

U

aij ≤

wc∗ i + di , (∀(i, j)) wc∗ j − dj

(43.3)

The interval importance grade shows the acceptable range for a decision maker.

346

T. Entani, H. Ichihashi, and H. Tanaka

43.3 Choice of Optimistic Weights and Efficiency by DEA 43.3.1 DEA with Normalized Data In DEA the maximum ratio of output data to input data is assumed as the efficiency which is calculated from the optimistic viewpoint for each DMU. The basic DEA model is formulated as follows. ∗

t θE o = max u y o u s.t. v t xo = 1 −v t X + ut Y ≤ 0 u, v ≥ 0

(43.4)

where the decision variables are the weight vectors u and v, X ∈ #m×n and Y ∈ #k×n are input and output matrices consisting of all input and output vectors that are all positive and the number of DMUs is n. In the conventional DEA as in (43.4), it is difficult to compare importance of input and output items to their weights, because the weights largely depend on the scales of the original data X and Y . Then we normalize the given input and output data based on DM Uo so that the input and output weights represent the importance grades of the items. The normalized input and output denoted as x ˆjp and yˆjr are obtained as follows. x ˆ jp = xjp /xop , yˆjr = y jr /y or The problem to obtain the efficiency with the normalized input and output are formulated as follows. ∗

θE u1 + · · · + u ˆk ) o = max(ˆ ˆ u s.t. vˆ1 + · · · + vˆm = 1 ˆ +u ˆ t Yˆ ≤ 0 −ˆ vt X ˆ, v ˆ≥0 u

(43.5)

ˆ and Yˆ are all the normalized data and u ˆ and v ˆ are the decision where X variables. The efficiency from the normalized input and output is equal to that from the original data by conventional DEA. The obtained weight represents the importance grade itself. Then we can use DEA with the normalized data to choose the optimistic weight in the interval importance grade obtained by a decision maker through interval AHP. 43.3.2 Optimistic Importance Grades in Interval Importance Grades A decision maker gives comparison values for all pairs of input and output items matricesfor input and output items whose elements  and Utheincomparison  L out are L ain are obtained. aij ,U aout ij , aij and ij

43. Optimistic Priority Weights

347

By the proposed interval AHP in 43.2, the importance grades of input and output items are denoted as follows.     Wpin = L wpin ,U wpin , Wrout = L wrout ,U wrout The optimistic or substitutional weights and efficiency are obtained by considering the interval importance grades through interval AHP as the weight constraints in DEA with the normalized data. By DEA, we can determine the optimistic weights for DM Uo in the possible ranges. The constraint conditions for the input and output weights are as follows, considering the difference between sums of centers of the interval importance grades and weights. L

wout ≤ r

u ˆr ≤ U wout r , (ˆ u1 + ... + u ˆk )

L

win ˆp ≤ U win p ≤v p

(43.6)

The problem to choose the most optimistic weights for DM Uo in the decision maker’s judgement is formulated by adding (43.6) to (43.5) as the constraint conditions. Any optimal solutions are within the inteval importance grades that are given by a decision maker based on his/her evaluation. As the character of DEA the optimal weights are obtained from the most optimistic viewpoint for DM Uo . Therefore both of a decision maker and DMUs are satisfied with the obtained evaluations.

43.4 Numerical Example 1-input and 4-output data of example DMUs(A,...,J) are shown in Table 43.1. The interval comparison matrix given by a decision maker and the interval Table 43.1. Data with 1-input and 4-output and efficiencies

A B C D E F G H I J

x1

y1

y2

y3

y4

DEA

proposed model

1 1 1 1 1 1 1 1 1 1

1 2 2 3 3 4 4 5 6 7

8 3 6 3 7 2 5 2 2 1

1 4 6 5 4 3 5 1 7 2

3 4 1 5 2 1 3 6 1 5

1.000 0.850 1.000 1.000 1.000 0.706 1.000 1.000 1.000 1.000

0.925 0.703 0.670 0.753 1.000 0.376 0.954 0.548 0.381 0.294

importance grades by (43.1) and (43.3) are shown in Table 43.2. Interval importance grades reflect inconsisntecy in the given interval comparison matrix.

348

T. Entani, H. Ichihashi, and H. Tanaka

Table 43.2. Comparison matrix and importance grades of output items

y1 y2 y3 y4

y1

y2

y3

y4

centers

importance grades

1 [3,6] [1/7,1/3] [2,6]

[1/6,1/3] 1 [1/8,1/6] [1/4,1/2]

[3,7] [6,8] 1 [3,9]

[1/6,1/2] [2,4] [1/3,1/9] 1

0.135 0.522 0.049 0.294

[0.071,0.200] [0.391,0.652] [0.029,0.070] [0.163,0.424]

Within the given interval importance grades DMUs are evaluated and their efficiencies obtained by the proposed model (43.6) with (43.5) are shown in Table 43.1. The efficiency through the proposed model can be obtained from the optimistic viewpoint within a decision maker’s acceptable importance grades. Therefore, the efficiencies in the proposed model are smaller than those in conventional DEA.

43.5 Concluding Remarks In this paper, we dealt with an interval comparison matrix that contains a decision maker’s uncertain judgements and obtained the interval importance grade of each item through interval AHP. Then, using DEA, we chose the most optimistic weights for DM Uo within the interval importance grades obtained by a decision maker. A decision maker’s evaluation and a DMU’s opinion are taken into consideration by interval AHP and DEA respectively.

References 43.1 Saaty,T.L. (1980): The Analytic Hierarchy Process. McGraw-Hill 43.2 Sugihara,K., Maeda,Y. and Tanaka,H. (1999): Interval Evaluation by AHP with Rough Set Concept. New Directions in Rough Sets, Data Mining and Granular-Soft Computing, Lecture Note in Artificial Intelligence 1711, Springer. 375–381 43.3 Tanaka,H. and Guo,P. (1999): Possibilistic Data Analysis for Operation Research. Physica-Verlag, A Springer Verlag Company 43.4 Charnes,A. Cooper,W.W. and Rhodes,E. (1978): Measuring the Efficiency of Decision Making Units. European Journal of Operational Research, 429–444 43.5 Tone,K.(1993) : Mesurement and Improvement of Efficiency by DEA. Nikkagiren (Japanese) 43.6 Ozawa,M. Yamaguchi,T. and Fukukawa,T. (1993): The Modified Assurance Region of DEA with Interval AHP. Communication of the Operations Reserch Society of Japan, 471-476 (Japanese)

4 4 . R o u g h S e t T h e o r y in C o n flic t A n a ly s is 5DIDá'HMD1 DQG'RPLQLN O ]DN2 1 2

Tok a r s k i eg o 4 /14 , 4 0-74 9 Ka towi ce, P ol a n d , r d @ a l ta .p l In s ti tu te of Ma th em a ti cs , Wa r s a w Un i ver s i ty B a n a ch a 2, 02-097 Wa r s a w, P ol a n d

4 4 .1 I n t r o d u c t io n Th e i m p or ta n ce of m u l ti -a g en ts s y s tem s , m od el s of a g en ts ’ i n ter a cti on i s i n cr ea s i n g n owa d a y s a s d i s tr i b u ted s y s tem s of com p u ter s s ta r ted to p l a y a s i g n i fi ca n t r ol e i n s oci ety . An i n ter a cti on occu r s wh en two or m or e a g en ts , wh i ch h a ve to a ct i n or d er to a tta i n th ei r ob j ecti ves , a r e b r ou g h t i n to a d y n a m i c r el a ti on s h i p . Th i s r el a ti on s h i p i s th e con s eq u en ce of th e l i m i ted r es ou r ces wh i ch a r e a va i l a b l e to th em i n a s i tu a ti on . If th e n u m b er of r es ou r ces i s i n s u ffi ci en t to a tta i n a g en ts ’ g oa l s i t often com es i n to th e con fl i cts . Th i s ca n h a p p en i n a l m os t a l l i n d u s tr i a l a cti vi ti es r eq u i r i n g d i s tr i b u ted a p p r oa ch , s u ch a s n etwor k con tr ol , th e d es i g n a n d m a n u fa ctu r e of i n d u s tr i a l p r od u cts or th e d i s tr i b u ted r eg u l a ti on of a u ton om ou s r ob ots . However , d i s tr i b u ted s y s tem s i s on l y on e fr om m a n y d i ffer en t a r ea s wh er e a con fl i ct ca n a r i s e a n d wh er e i t i s wor th to a p p l y com p u ter a i d ed con fl i ct a n a l y s i s . Ju s t to m en ti on s om e h u m a n a cti vi ti es l i k e b u s i n es s , g over n m en t, p ol i ti ca l or m i l i ta r y op er a ti on s , l a b ou r -m a n a g em en t n eg oti a ti on s etc. etc. In th e p a p er , we ex p l a i n th e n a tu r e of con fl i ct a n d we d efi n e th e con fl i ct s i tu a ti on m od el i n a wa y to en ca p s u l a te th e con fl i ct com p on en ts i n a cl ea r m a n n er . We p r op os e s om e m eth od s to s ol ve th e m os t fu n d a m en ta l p r ob l em s r el a ted to con fl i cts . P a w la k M o d e l Th e m od el i n tr od u ced i n th i s p a p er i s a n en h a n cem en t of th e m od el p r op os ed b y P a wl a k i n p a p er s e.g . [ 4 4 .6, 4 4 .8 ] . In th e P a wl a k m od el , s om e i s s u es a r e ch os en , a n d th e a g en ts a r e a s k ed to s p eci fy th ei r vi ews : a r e th ey fa vou r a b l e, n eu tr a l or a g a i n s t. Th u s th e a n a l y s i s a r e n a tu r a l l y r es tr i cted to ou ter m os t con cl u s i on s l i k e fi n d i n g th e m os t con fl i cti n g a ttr i b u tes or th e coa l i ti on s of a g en ts i f m or e th a n two ta k e p a r t i n th e con fl i ct [ 4 4 .8 ] . In th e r ea l wor l d , vi ews on th e i s s u es to vote a r e con s eq u en ces of th e d eci s i on ta k en , b a s ed on th e l oca l i s s u es , th e cu r r en t s ta te a n d s om e b a ck g r ou n d k n owl ed g e u s i n g s om e s tr a teg y . Th er efor e, th e P a wl a k m od el i s en h a n ced h er e b y a d d i n g to th e m od el s om e l oca l a s p ects of con fl i cts .

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 4 9-3 53 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

R 'HMDDQG' O ]DN

3 50

4 4 .2 C o n f lic t M o d e l Th e i n for m a ti on a b ou t th e l o c a l s t a t e s U a g of a n a g en t a g ca n b e p r es en ted i n th e for m of a n i n for m a ti on ta b l e, cr ea ti n g th e a g en t a g ’ s i n for m a ti on s y s tem I a g = (U a g , A a g ), wh er e a : U a g → V a for a n y a ∈ A a g a n d V a i s th e va l u e s et of a ttr i b u te a . We a s s u m e:

V

a g

=

UV

a ∈ A

a

An y l oca l s ta te s ∈ U

a g

i s ex p l i ci tl y d es cr i b ed b y i ts i n f o r m a -

a g

t i o n v e c t o r I n f A a g (s ), wh er e I n f A a g (s )= { (a , a (s )): a ∈ A a g } . Th e s et I n f A a g (s ): s ∈ U a g i s d en oted b y I N F A a g a n d i t i s ca l l ed th e i n f o r m a t i o n v e c t o r s e t o f a g . We a s s u m e th a t s ets { Aa g } a r e p a i r wi s e d i s j oi n t, i .e., A a g ∩ A a g ’ = ∅ for a g ≠ a g ’ . Th i s con d i ti on em p h a s i z es th a t a n y a g en t i s d es cr i b i n g th e s i tu a ti on i n i ts own wa y . R el a ti on s h i p s a m on g a ttr i b u tes of d i ffer en t a g en ts wi l l b e d efi n ed b y con s tr a i n ts a s s h own i n s ecti on 0. L o c a l S e t o f G o a ls (S im ila r ity o f S ta te s ) E ver y a g en t eva l u a tes th e l oca l s ta tes . Th e s u b j e c t i v e e v a l u a t i o n cor r es p on d s to a n or d er (or p a r ti a l or d er ) of th e s ta tes of th e a g en t i n for m a ti on ta b l e. We a s s u m e th a t th e fu n cti on e a g ca l l ed th e t a r g e t f u n c t i o n , a s s i g n s a n eva l u a ti on s cor e to ea ch s ta te; l et for ex a m p l e e a g : U a g → [ 0, 1] . Th e s ta tes wi th s cor e 1 a r e m os tl y p r efer r ed b y th e a g en t a s ta r g et s ta tes , wh i l e th e s ta tes wi th s cor e 0 a r e n ot a ccep ta b l e. Ma x i m a l el em en ts (d eter m i n ed b y a n p a r ti a l or d er ) ca n b e i n ter p r eted a s th os e, wh i ch a r e ta r g ets of th e a g en t, i .e., th e a g en t wa n ts to r ea ch th em e.g . i n a n eg oti a ti on p r oces s . Mor e p r eci s el y th e a g en t a g ’ s s e t o f g o a l s (t a r g e t s ) d en oted b y T (a g ) i s d efi n ed a s th e s et of ta r g et s ta tes of a g , wh i ch m ea n s T a g = { s ∈ U a g : e a g (s )> μ a g } , a n d μ a g i s th e a c c e p t a n c e l e v e l , ch os en b y th e a g en t a g – i t i s s u b j ecti ve wh i ch eva l u a ti on l evel i s a ccep ta b l e b y th e a g en t. Th e s ta te eva l u a ti on ca n a l s o h el p u s to fi n d th e s ta te s i m i l a r i ty [ 4 4 .4 ] . F or a n y ε > 0 a n d s ∈ U a g , we d efi n e ε -n e i g h b o u r h o o d of s b y : τ a g , ε (s )= { s ' ∈ Ua g : | ea g (s ) ea g (s ' )| ≤ ε } . Th e fa m i l y { τ a g , ε (s )} s ∈ U a g d efi n es a tol er a n ce r el a ti on τ a g , ε (s ) i n U a g × U a g b y s τ a g , ε s ’ i ff s ’ ∈ τ a g , ε (s ). L o c a l C o n flic t Th e a g en t a g i s i n th e ε -l o c a l c o n f l i c t i n a s ta te s ε -n ei g h b ou r h ood of s ’ , for a n y s ’ fr om th e s et of a th r es h ol d . L oca l con fl i cts for a n a g en t a g a r i s e fr om eva l u a ti on of th e cu r r en t s ta te b y a g . It ca n b e ex p r es d oes n ot b el on g to th e ε -en vi r on s of th e s et of g oa l s wh er e τ

a g ,ε

(s ’ )= { s ’ ’ : s ’ ’ τ

a g ,ε

s ’} .

i ff g -ta th s ed T ag

s d oes n ot b el on r g ets wh er e ε i s e l ow l evel of s u d i ffer en tl y th a t th i .e.: s ∉ U τ a s ’∈ T

a g

g to th e a g i ven b j ecti ve e s ta te s g , ε (s ’ ) ,

4 4 . R ou g h Set Th eor y i n Con fl i ct An a l y s i s

3 51

S itu a tio n L et u s con s i d er a s et A g con s i s ti n g of n a g en ts a g 1, ..., a g n . A s i t u a t i o n of A g i s a n y el em en t of th e Ca r tes i a n p r od u ct S ( A g ) = s et

of

a ll

p os s i b l e

IN F * (a g ) = { f : A

a g

→

i n for m a ti on

UV

a ∈ A

a

∏

n

I N F * (a g i ) , wh er e I N F * (a g i ) i s th e i= 1

vector s

( a g ) : f (a ) ∈ V

of a

a g en t

a g i,

(a g ) for a ∈ A

a g

d efi n ed b y : } . Th e s i t u a t i o n

a g

S cor r es p on d i n g to a g l ob a l s ta te (I n f A a g 1 (s 1 ), ..., I n f A a g n (s n )).

s = (s 1 , ..., s n )∈ U

a g 1

× ...× U

a g n

i s d efi n ed b y

C o n s tr a in ts Con s tr a i n ts a r e d es cr i b ed b y s om e d ep en d en ci es a m on g l oca l s ta tes of a g en ts . Wi th ou t a n y d ep en d en ci es , a n y a g en t cou l d ta k e th e s ta te fr eel y a n d th er e i s n o con fl i ct a t a l l . D ep en d en ci es com e fr om th e b ou n d on th e n u m b er of r es ou r ces (a n y k i n d of a r es ou r ce m a y b e con s i d er ed , e.g . wa ter on G ol a n Hi l l s s ee [ 4 4 .8 ] or a n i n ter n a ti on a l p os i ti on [ 4 4 .5] , ever y th i n g th a t i s es s en ti a l for a g en ts ). Con s tr a i n i n g r el a ti on s a r e i n tr od u ced to ex p r es s wh i ch l oca l s ta tes of a g en ts ca n coex i s t i n th e (g l ob a l ) s i tu a ti on . Mor e p r eci s el y , c o n s t r a i n t s a r e u s ed to d efi n e a s u b s et S (A g ) of g l ob a l s i tu a ti on s . Con s tr a i n ts r es tr i ct th e s et of p os s i b l e s i tu a ti on s to a d m i s s i b l e s i tu a ti on s s a ti s fy i n g con s tr a i n ts . S itu a tio n s E v a lu a tio n Us u a l l y a g en ts ten d to a tta i n th e b es t s ta tes wi th ou t ta k i n g ca r e a b ou t th e g l o b a l g o o d . However , th e n eg oti a tor s ex p er i en ce s h ows th a t th e r ea l , s ta b l e con s en s u s ca n on l y b e fou n d wh en th e g l ob a l g ood i s con s i d er ed . Th u s th e o b j e c t i v e e v a l u a t i o n o f s i t u a t i o n s i s i n tr od u ced – a n ex p er t (a n a r b i ter ) j u d g em en t. F or ex a m p l e, th e Un i ted N a ti on Or g a n i s a ti on ca n b e th ou g h t a s a n ex p er t i n th e m i l i ta r y con fl i cts . We a s s u m e th er e i s a fu n cti on q : S A g → [ 0, 1] , ca l l ed th e q u a l i t y f u n c t i o n , wh i ch a s s i g n s a s cor e to ea ch s i tu a ti on (th i s s cor e i s a s s u m e to b e g i ven b y a n ex p er t). Th e s et of s i tu a ti on s s a ti s fy i n g a g i ven l evel of q u a l i ty t i s d efi n ed b y : S c o r e A g (t ) = { S ∈ S ( A g ) : q ( S ) ≥ t }

S y ste m

w ith C o n s tr a in ts

Th e m u l ti -a g en t s y s tem , wi th l oca l s ta tes for ea ch a g en t d efi n ed a n d th e g l ob a l s i tu a ti on s s a ti s fy i n g con s tr a i n ts , wi l l b e ca l l ed t h e s y s t e m w i t h c o n s t r a i n t s . We d en ote ou r s y s tem wi th con s tr a i n ts b y M A g .

3 52

R 'HMDDQG' O ]DN

4 4 .3 A n a ly s is Th e i n tr od u ced a b ove con fl i ct m od el g i ves u s p os s i b i l i ty , fi r s t to u n d er s ta n d a n d , th en , to a n a l y s e d i ffer en t k i n d s of con fl i cts . P a r ti cu l a r l y , th e m os t fu n d a m en ta l p r ob l em ca n b e wi d el y i n ves ti g a ted , th a t i s , th e p os s i b i l i ty to a ch i eve th e con s en s u s . B eca u s e of th e l a ck of s p a ce on l y th e con s en s u s p r ob l em on l oca l p r efer en ces i s d es cr i b ed i n th i s p a p er . We p r op os e B ool ea n r ea s on i n g [ 4 4 .1] a n d R ou g h Set m eth od ol og y [ 4 4 .7] for a l l a n a l y s i s . Th e m a i n i d ea of B ool ea n r ea s on i n g i s to en cod e th e op ti m i s a ti on p r ob l em , b y cor r es p on d i n g B ool ea n fu n cti on f π i n s u ch a wa y th a t a n y p r i m e i m p l i ca n t of f π s ta tes a s ol u ti on of π . Th e el em en ta r y B ool ea n for m u l a i s u s u a l l y ob ta i n ed h er e b y tr a n s for m i n g th e i n for m a ti on ta b l e i n to th e d eci s i on ta b l e, g en er a ti n g r u l es (m i n i m a l wi th r es p ect of n u m b er of a ttr i b u tes on l eft s i d e) a n d d eter m i n i n g th e d es cr i p ti on of d eci s i on cl a s s 4 4 .9. F r om th e el em en ta r y for m u l a s th e fi n a l for m u l a d es cr i b i n g th e p r ob l em i s s h a p ed . Un for tu n a tel y ca l cu l a ti n g p r i m e i m p l i ca n ts of s u ch for m u l a s i s u s u a l l y a h a r d com p u ta ti on a l p r ob l em [ 4 4 .4 ] . Th er efor e d ep en d i n g on th e for m u l a , s om e s i m p l e s tr a teg i es or even tu a l l y q u i te com p l ex h eu r i s ti cs m u s t b e u s ed to r es ol ve th e p r ob l em i n r ea l ti m e. C o n s e n s u s P r o b le m

o n L o c a l a n d G lo b a l L e v e l

In th i s p oi n t a con fl i ct a n a l y s i s i s p r op os ed wh er e l oca l i n for m a ti on ta b l es a n d th e s et of l oca l g oa l s a r e ta k en i n to con s i d er a ti on . I N P U T Th e s y s tem wi th con s tr a i n ts M A g d efi n ed i n Secti on 0. t - a n a ccep ta b l e th r es h ol d of th e ob j ecti ve g l ob a l con fl i ct for A g . O U T P U T Al l s i tu a ti on s wi th th e ob j ecti ve eva l u a ti on r ed u ced to d eg r ee a t m os t t , a n d wi th ou t l oca l con fl i ct for a n y a g en t. (i t i s r eq u i r ed th a t a n y n ew s i tu a ti on i s con s tr u cted i n th e wa y th a t a l l l oca l s ta tes i n th i s s i tu a ti on a r e fa vou r a b l e for th e a g en ts ). A L G O R I T H M Th e a l g or i th m i s b a s ed on ver i fi ca ti on of g l ob a l s i tu a ti on s fr om S c o r e A g (t ) wi th th e l oca l s et of g oa l s of a g en ts a n d con s tr a i n ts . Th e p r ob l em i s d es cr i b ed b y th e for t a g ∧ f C ∧ f ϕ , wh er e t a g d es cr i b es th e s et of g oa l s of th e a g en t a g , m u la f: f =

∧

a g ∈ A g

a n d f C d es cr i b es S c o r e A g (t ) a n d f ϕ th e con s tr a i n ts . Th e for m u l a f C ∧ f ϕ r ep r es en ti n g a l l a d m i s s i b l e s i tu a ti on s wi th ou t th e g l ob a l con fl i ct r eg a r d i n g th e th r es h ol d t .

4 4 .4 C o n c lu s io n s We h a ve p r es en ted a n d d i s cu s s ed th e ex ten s i on of th e P a wl a k con fl i ct m od el . Th e u n d er s ta n d i n g of th e u n d er l y i n g l oca l s ta tes a s wel l a s con s tr a i n ts i n th e g i ven s i tu a ti on i s th e b a s i s for a n y a n a l y s i s of ou r wor l d . Th e l oca l g oa l s a n d th e eva l u a -

4 4 . R ou g h Set Th eor y i n Con fl i ct An a l y s i s

ti on of th e g l ob a l fl i ct a n d ca n s u g g Th e fu n d a m en B ool ea n r ea s on i n p r es en ted p r ob l em fl i ct ex a m p l e – s m od el .

3 53

s i tu a ti on a r e ob s er ved a s fa ctor s d efi n i n g th e s tr en g th of th e con es t th e wa y to r ea ch th e con s en s u s . ta l con s en s u s p r ob l em h a s b een a n a l y s ed i n th e p a p er . Th en , g a n d r ou g h s et th eor y h a s b een s u cces s fu l l y a p p l i ed for s ol vi n g . Th e l a ck of s p a ce n ot a l l owed th e a u th or s to p r es en t a n y con ee [ 4 4 .2] for s om e ex em p l a r con fl i ct a n a l y s i s wi th i n p r op os ed

R e fe r e n c e s 4 4 .1 4 4 .2

B r own , F . N ., 1990, “ B ool ea n R ea s on i n g ” . Kl u wer , D or d r ech t. D ej a , R . 2000, “ Con fl i ct An a l y s i s ” , R ou g h Set Meth od s a n d Ap p l i ca ti on s ; N ew D evel op m en ts . In : L . P ol k ows k i , et a l . (ed s .), Stu d i es i n F u z z i n es s a n d Soft Com p u ti n g , P h y s i ca -Ver l a g , p p .4 91-520. 4 4 .3 F r a s er , N .M.; Hi p el , K.W., 198 4 , “ Con fl i ct An a l y s i s : Mod el s a n d R es ol u ti on s ” , N or th -Hol l a n d , N ew Yor k . 4 4 .4 Howa r d , N ., 1975, “ Meta g a m e a n a l y s i s of b u s i n es s p r ob l em s ” , IN F OR 13 , p p . 4 8 67. 4 4 .5 Kom or ows k i , J.; P a wl a k , Z .; P ol k ows k i , L .; Sk owr on , A., 1999, “ R ou g h s ets : A tu tor i a l .” i n : S.K. P a l a n d A. Sk owr on (ed s .), R ou g h fu z z y h y b r i d i z a ti on : A n ew tr en d i n d eci s i on m a k i n g , Sp r i n g er -Ver l a g , Si n g a p or e, p p . 3 -98 . 4 4 .6 1 FNL= ³1HJRWLDWLRQVLQ EXVLQHVV´3URIHVVLRQDO 6FKRRO RI %XVLQHVV (Gi ti on . (Th e b ook i n P ol i s h ), Kr a k ow. 4 4 .7 P a wl a k , Z ., 198 4 , “ On Con fl i cts ” , In t. J. of Ma n -Ma ch i n e Stu d i es , 21, p p . 127-13 4 . 4 4 .8 P a wl a k , Z ., 1991, “ R ou g h Sets – Th eor eti ca l As p ects of R ea s on i n g a b ou t D a ta ” , Kl u wer Aca d em i c P u b l i s h er s , D or d r ech t. 4 4 .9 P a wl a k , Z ., 1998 , “ An In q u i r y i n to An a tom y of Con fl i cts ” , Jou r n a l of In for m a ti on Sci en ces 109 p p . 65-78 . 4 4 .10 P ol k ows k i , L .; Sk owr on , A., (E d s .) 1998 , “ R ou g h Sets i n Kn owl ed g e D i s cover y ” (two p a r ts ), P h y s i ca -Ver l a g , Hei d el b er g . 4 4 .11 R os en h ei m , J.S.; Z l otk i n , G ., 1994 , “ D es i g n i n g Con ven ti on s for Au tom a ted N eg oti a ti on ” , AI Ma g a z i n e 15(3 ) p p . 29-4 6. Am er i ca n As s oci a ti on for Ar ti fi ci a l In tel l i g en ce. 4 4 .12 Sa n d h ol m , T.; L es s er , V., 1997, “ Coa l i ti on s a m on g Com p u ta ti on a l l y B ou n d ed Ag en ts ” , Ar ti fi ci a l In tel l i g en ce 94 (1), p p . 99-13 7, Sp eci a l i s s u e on E con om i c P r i n ci p l es of Mu l ti a g en t Sy s tem s .

45. Dealing with Imperfect Data by RS-ILP Chunnian Liu1 and Ning Zhong2 1 2

School of Computer Science, Beijing Polytechnic University, China [email protected] Dept. of Information Eng., Maebashi Institute of Technology, Japan [email protected]

Rough Set theory and Granular Computing (GrC) have a great impact on the study of intelligent information systems. This paper investigates the feasibility of applying Rough Set theory and Granular Computing (GrC) to deal with imperfect data in Inductive Logic Programming (ILP). We propose a hybrid approach, RS-ILP, to deal with some kinds of imperfect data which occur in real-world applications.

45.1 Introduction Inductive Logic Programming (ILP, see [45.2, 45.7, 45.8]) can be regarded as a new method in machine learning with the advantages of more expressive power and ease of using background knowledge. If databases are involved, ILP is also relevant to Knowledge Discovery and Data Mining (KDD, see [45.1, 45.3]). In a simplified form, the normal problem setting of ILP is as follows: Given: – The target predicate p. – The positive examples E + and the negative examples E − (two sets of ground atoms of p). – Background knowledge B (a finite set of definite clauses). To find: – Hypothesis H (the defining clauses of p) which is correct with respect to E + and E − , i.e. 1. H ∪ B is complete with respect to E + (that is: For all e ∈ E + , H ∪ B implies e ). We also say that H ∪ B covers all positive examples. 2. H ∪ B is consistent with respect to E − (that is: For no e ∈ E − , H ∪ B implies e ). We also say that H ∪ B rejects any negative examples. To make the ILP problem meaningful, we assume the following prior conditions: 1. B is not complete with respect to E + (Otherwise there will be no learning task at all, because the background knowledge itself is the solution). 2. B ∪ E + is consistent with respect to E − (Otherwise there will be no solution to the learning task). T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 54 − 3 58 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

45. Dealing with Imperfect Data by RS-ILP

355

In the above normal problem setting for ILP, everything is assumed correct and perfect. But in large, real-world empirical learning, data are not always perfect. In contrary, uncertainty, incompleteness, vagueness, impreciseness, etc. are frequently observed in the input to ILP – the training examples and/or background knowledge. Imperfect input, in addition to improper bias setting, will induce imperfect hypotheses. Thus ILP has to deal with imperfect data. In this aspect, the theory, measurement, techniques and experiences are much less mature for ILP than in the traditional attribute-value learning methods (compare [45.12], for example). We observe that many problems concerning imperfect input or too strong bias in ILP have a common feature. In these situations, while it is impossible to differentiate distinct objects, we may consider granules – sets of objects drawn together by similarity, indistinguishability, or functionality. The emerging theory of Granular Computing (GrC) (see [45.15, 45.16, 45.14]) grasps the essential concept – granules, and makes use of them in general problem solving. In this paper we concentrate on a particular form of GrC, Pawlak’s Rough Set theory [45.9, 45.10, 45.11], investigating its potentials in dealing with imperfect data of ILP. The main idea is that, when we use granules instead of individual objects, we are actually relaxing the strict requirements in the standard normal problem setting for ILP. In the following sections, we will discuss some kinds of imperfect data in ILP and propose a hybrid approach, RS-ILP, as a solution using Rough Set theory, to deal with such imperfect data.

45.2 Imperfect Data in ILP We discuss here two kinds of imperfect data encountered in ILP as examples. – Incomplete background knowledge. Background knowledge B is essential in ILP learning, and the ease of using background knowledge is one of the major advantages of ILP over traditional attribute-value learning methods. However, if B lacks essential predicates (or essential clauses of some predicates), it is possible that no non-trivial hypothesis H can be induced. (Note that E + itself can be always regarded as a hypothesis, but it is trivial). In some cases, even a large amount of positive examples are given, some examples are not generalized by hypotheses if some background knowledge is missing. This has been a big topic in the research area of ILP. – Missing classification. This means that some examples have unknown classification values (i.e., we do not know if an example belongs to E + or E − ). Here we have a set of classified training instances E + ∪ E − and a set of unclassified instances E ? . If the classified set is small and the unclassified set E ? is ignored, we

356

C. Liu and N. Zhong

are facing with the problem of too sparse data to induce reliable hypothesis H. But here we have got a set of additional examples E ? though we don’t know their classification. The challenge is how to utilize our knowledge about E ? to induce more reliable hypotheses. One approach is to combine learning and conceptual clustering techniques (see [45.2]): a conceptual clustering algorithm is applied to the set of all known examples, climbing the hierarchytree, using the classified examples to identify class descriptions forming H. We have proposed several rough problem settings of ILP (RS-ILP for short) to deal with such imperfect data. The key idea is to relax the requirement in the normal problem setting that H should be “correct with respect to E + and E − ”, so that rough but useful hypotheses can be induced. Some of them will be discussed in the following sections.

45.3 RS-ILP for Missing Classification If E + ∪ E − is a small set, we cannot expect that the induced hypothesis H will have high prediction accuracy. Sometimes we may have an additional set of examples E ? that are unclassified (that is, we do not know whether these examples belong to E + or E − ). Can we utilize E ? to increase the prediction accuracy? We propose the following rough problem setting for this purpose: Given: – The target predicate p (the set of all ground atoms of p is U ). – An equivalence relation R on U (we have the approximation space A = (U, R)). – A set of positive examples E + ⊆ U and A set of negative examples E − ⊆ U. – A set of unclassified examples E ? ⊆ U. – Background knowledge B. Considering the following rough sets: 1. E +? = E + ∪ {e? ∈ E ? |∃e∈E + eRe? }; 2. E −? = E − ∪ {e? ∈ E ? |∃e∈E − eRe? }. To find: – Hypothesis H ? (the defining clauses of p) which is correct with respect to E +? and E −? . That is, 1. H ? ∪ B covers all examples of E +? ; 2. H ? ∪ B rejects all examples of E −? . In such rough problem setting, we use equivalence relation R to “enlarge” the training set (by distributing some examples from E ? to E + and E − ). Different R will produce different hypothesis H ? . It is reasonable to expect

45. Dealing with Imperfect Data by RS-ILP

357

that the more unclassified examples are added to E + , the more general hypothesis will be induced; the more unclassified examples are added to E − , the more specific hypothesis will be induced.

45.4 RS-ILP for Too Strong Bias Declarative bias (restrictions on the hypothesis space and/or on the search strategies) is necessary in any inductive learning (so in ILP). Clearly, there is a trade-off between the tractability of search, which is improved by a small search space, and the availibility of a correct hypothesis, which is improved by a large search space. Particularly, if the bias is too strong, we may miss some useful solutions or have no solution at all. Most ILP systems provide mechanisms for the user to specify bias, and allow the user to change bias (weakening the restrictions when the current ILP session fails). This strategy is called bias shift. Here we investigate this problem from another point of view. Supposing that the training set E + ∪ E − and the background knowledge B are perfect, but if we restrict the hypotheses to non-recursive clauses (a bias often imposed in some ILP systems), we still could not find any meaningful hypothesis in the normal problem setting of ILP. However, relaxing the requirement in the normal problem setting of ILP that H should be “correct with respect to E + and E − ”, in order to find a “rough” solution that is within the language defined by the bias.

45.5 Concluding Remarks This paper addressed the problem of imperfect data handling in Inductive Logic Programming (ILP) using some ideas, concepts and methods of Rough Set theory and GrC. We presented a hybrid approach, RS-ILP, to deal with some kinds of imperfect data which occur in large real-world applications. Although some part of this work is still in the initial shape, we believe that the general ideas presented here may give rise to more concrete results in future research. Future work in this direction includes finding more concrete formalisms and methods to deal with other kinds of imperfect data, and giving quantitative measures associated with hypotheses induced of RS-ILP. Acknowledgements. C. Liu’s work is supported by the Natural Science Foundation of China (NSFC), Beijing Municipal Natural Science Foundation (BMNSF) and Chinese 863 High-Tech Program.

358

C. Liu and N. Zhong

References 45.1 S. Dzeroski, “Inductive Logic Programming and Knowledge Discovery in Databases”, Advances in KDD, MIT Press, 117-151, 1996. 45.2 N. Lavrac, S. Dzeroski, and I. Bratko “Handling Imperfect Data in Inductive Logic Programming”, in L.de Raedt (Eds), Advances in Inductive Logic Programming, IOS Press, 48-64, 1996. 45.3 C. Liu, N. Zhong, and S. Ohsuga, “Constraint ILP and its Application to KDD”, Proc. of IJCAI-97 Workshop on Frontiers of ILP, Nagoya, Japan, 103-104, 1997. 45.4 C. Liu and N. Zhong, “Rough Problem Settings for Inductive Logic Programming”, Zhong, N., Skowron, A., and Ohsuga, S. (eds.) New Directions in Rough Sets, Data Mining, and Granular-Soft Computing, Springer LNAI 1711, 168-177, 1999. 45.5 T.M. Mitchell, Machine Learning, McGraw-Hill, 1997. 45.6 S. Moyle and S. Muggleton, “Learning Programs in the Event Calculus”, in Proc. 7th International Workshop on ILP, 205-212, 1997. 45.7 S. Muggleton, “Inductive Logic Programming”, New Generation Computing, 8(4):295-317, 1991. 45.8 S. Muggleton (Eds), Inductive Logic Programming, Academic Press, 1992 45.9 Z. Pawlak, “Rough Sets”, International Journal of Computer and Information Science, Vol.11, 341-356, 1982. 45.10 Z. Pawlak, Rough Sets: Theoretical Aspects of Reasoning about Data, Kluwer Academic Publishers, Boston, 1991. 45.11 Z. Pawlak, “Granularity of knowledge, indiscernibility and rough sets”, Proc. 1998 IEEE International Conference on Fuzzy Systems, 106-110, 1998. 45.12 Y.Y. Yao and N. Zhong, “An Analysis of Quantitative Measures Associated with Rules”, N. Zhong and L. Zhou (Eds), Methodologies for Knowledge Discovery and Data Mining, Springer LNAI 1574, 479-488, 1999. 45.13 Y.Y. Yao, “Granular Computing using Neighborhood Systems”, Roy, R., Furuhashi, T., and Chawdhry, P.K. (eds.) Advances in Soft Computing: Engineering Design and Manufacturing, Springer 539-553, 1999 45.14 Y.Y. Yao, Granular Computing: Basic Issues and Possible Solutions, Proc. JCIS 2000, invited session on Granular Computing and Data Mining, Vol.1, 186-189, 2000. 45.15 L.A. Zadeh, “Fuzzy Sets and Information Granularity”, Gupta, N., Ragade, R. and Yager, R. (Eds.) Advances in Fuzzy Set Theory and Applications, North-Holland, Amsterdam, 3-18, 1979. 45.16 L.A. Zadeh, “Toward a Theory of Fuzzy Information Granulation and Its Centrality in Human Reasoning and Fuzzy Logic”, Fuzzy Sets and Systems, Vol.19, 111-127, 1997. 45.17 N. Zhong, J. Dong, and S. Ohsuga, “Data Mining: A Probabilistic Rough Set Approach”, L. Polkowski and A. Skowron (Eds) Rough Sets in Knowledge Discovery, Physica-Verlag, 127-146, 1998.

46. Extracting Patterns Using Information Granules: A Brief Introduction Andrzej Skowron1 , Jaroslaw Stepaniuk2 , and James F. Peters3 1 2 3

Institute of Mathematics, Warsaw University, 02-097 Banacha 2, Warsaw, Poland, [email protected] Institute of Computer Science, Bialystok University of Technology, Wiejska 45A, 15-351 Bialystok, Poland, [email protected] Computer Engineering, University of Manitoba, Winnipeg, MB R3T 5V6 Canada, [email protected]

The paper realizes a step in developing a foundation for approximate reasoning from experimental data to conclusions in natural language. Granule decomposition strategies based on background knowledge are outlined.

46.1 Introduction Information granulation belongs to a collection of intensively studied topics in soft computing (see, e.g., [46.19], [46.20], [46.21]). One of the recently emerging approaches to deal with information granulation is based on information granule calculi (see, e.g., [46.10], [46.12], [46.15], [46.13])developed on the basis of the rough set [46.6] and rough mereological approaches (see, e.g., [46.9], [46.10], [46.12]). The development of such calculi is important for making progress in many areas like object identification by autonomous systems (see, e.g., [46.1], [46.18]), web mining (see, e.g., [46.4]), approximate reasoning based on information granules (see, e.g., [46.15], [46.7]) or spatial reasoning (see, e.g., [46.2], [46.8]). In particular, reasoning methods using background knowledge as well as knowledge extracted from experimental data (e.g., sensor measurements) represented by concept approximations [46.1] are important for making progress in such areas. Schemes of approximate reasoning (AR-schemes, for short) are derived from parameterized productions [46.11], [46.13]. The productions, specifying properties of operations on information granules, are assumed to be extracted from experimental data and background knowledge. The problem of ARschemes deriving is closely related to perception (see, e.g., [46.21]). In the paper we outline some methods for decompostion of information granules.

46.2 Granule Decomposition In this section, we discuss briefly a granule decomposition problem. This is one of the basic problems in synthesis of approximate schemes of reasoning

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 59− 3 63 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

360

A. Skowron, J. Stepaniuk, and J.F. Peters

from experimental data. We restrict our considerations to the case of information granule decomposition supported by background knowledge. Some other decomposition methods are presented in [46.9], [46.5]. Assume that a knowledge base consists of a fact expressing that if two objects belong to concepts C1 and C2 , then the object constructed out of them by means of a given operation f belongs to the concept C provided that the two objects satisfy some constraints. However, we can only approximate these concepts on the basis of available data. Using a (generalized) rough set approach [46.14] one can assume that an inclusion measure vp for p ∈ [0, 1] is given making it possible to estimate the degree of inclusion of data patterns P at, P at1 , and P at2 from languages L, L1 , and L2 in the concepts C, C1 , and C2 , respectively. Patterns included to a satisfactory degree p in a concept are classified as belonging to its lower approximation while those included to a degree less than a preset threshold q ≤ p are classified as belonging to its complement. Information granule decomposition supported by background knowledge is accomplished by searching for patterns P at of high quality (e.g., supported by a large number of objects) and included in a satisfactory degree in the target concept C. These patterns are obtained by performing a given operation f on some input patterns P at1 and P at2 (from languages L1 and L2 , respectively) sufficiently included in C1 and C2 , respectively. One can develop a searching method for such patterns P at based on tuning of inclusion degrees p1 , p2 of input patterns P at1 , P at2 in C1 , C2 , respectively, to obtain patterns P at (constructed from P at1 , P at1 by means of a given operation f ) included in C in a satisfactory degree p and of acceptable quality (e.g., supported by the number of objects larger than a given threshold). Assume degrees p1 , p2 are given. There are two basic steps of searching procedures for relevant pairs of patterns (P at1 , P at2 ) : (i) searching in languages L1 and L2 for sets of patterns included in degree at least p1 and p2 in concepts C1 and C2 , respectively, (ii) selecting from sets of patterns generated in step (i) satisfactory pairs of patterns. We would like to add some general remarks on the above steps. One can see that our method is based on a decomposition of degree p into degrees p1 and p2 under some constraints. In Step 2, we search for a relevant constraint relation R between patterns. By Sem(P at) we denote the meaning of P at in, e.g., a given information system. The goal is to extract the following approximate rule of reasoning: if R(Sem(P at1 ), Sem(P at2 )) ∧ νp1 (Sem(P at1 ), C1 ) ∧ νp2 (Sem(P at2 ), C2 ) then νp (f (Sem(P at1 ) × Sem(P at2 )), C) ∧ Qualityt (f (Sem(P at1 ) × Sem(P at2 ))) where p is a given inclusion degree, t - a threshold of pattern quality measure Qualityt , f - operation on objects (patterns), P at- target pattern, C, C1 , C2 -given concepts, R, p1 , p2 are expected to be extracted from data

46. Extracting Patterns Using Information Granules

361

and (P at1 , P at2 ) is satisfying R (in the case we consider R is represented by a finite set of pattern pairs). One can also consider soft constraint relations Rr where r ∈ [0, 1] is a degree of truth to which the constraint relation holds. Two sets P1 , P2 are returned as the result of the first step. They consist of pairs (pattern, degree) where pattern is included in C1 , C2 , respectively in degree at least degree. These two sets are used to learn the relevant relation R. We outline two methods. The first method is based on an experimental decision table (U, A, d) [46.6] where U is a set of pairs of discovered patterns in the first step; A = {deg1 , deg2 } consists of two attributes such that degi ((P at1 , P at2 )) is equal to the degree to which P ati is at least included in Ci for i = 1, 2; the decision d has value p to which the granule composed by means of operation f from (P at1 , P at2 ) is at least included in C. From this decision table the decision rules of a special form are induced: if deg1 ≥ p1 ∧ deg2 ≥ p2 then d ≥ p where (p1 , p2 ) is a minimal degree pair such that if p1 ≥ p1 and p2 ≥ p2 then the decision rule obtained from the above rule by replacing p1 , p2 instead of p1 , p2 , respectively, is also true in the considered decision table. A version of such a method has been proposed in [46.9]. The relation R consists of the set of all pairs (P at1 , P at2 ) of patterns with components included in C1 , C2 , respectively in degrees p1 ≥ p1 , p2 ≥ p2 where p1 , p2 appear on the left hand side of some of the generated decision rules. The second method is based on another experimental decision table (U, A, d) where objects are triplets (x, y, f (x, y)) composed out of objects x, y and the result of f on arguments x, y; attributes from A describe features of arguments of objects and the decision d is equal to the degree to which the elementary granule corresponding to the description of f (x, y) by means of attributes is at least included in C. This table is extended by adding new features being characteristic functions aP ati of patterns P ati discovered in the first step. Next the attributes from A are deleted and from the resulting decision table the decision rules of a special form are induced: if aP at1 = 1 ∧ aP at2 = 1 then d ≥ p where if P at1 , P at2 are included in C1 , C2 , in degree at least p1 , p2 , respectively and P at1 , P at2 are included in C1 , C2 in degree p1 ≥ p1 and p2 ≥ p2 , respectively then a decision rule obtained from the above rule by replacing P at1 , P at2 instead of P at1 , P at2 is also true in the considered decision system. The decision rules describe constraints specifying the constraint relation R. Certainly, in searching procedures one should also consider constraints for the pattern quality. The searching methods discussed in this section return local granule decomposition schemes. These local schemes can be composed using techniques discussed in [46.10]. The received schemes of granule construction (which can be also treated as approximate reasoning schemes) have also the following

362

A. Skowron, J. Stepaniuk, and J.F. Peters

stability (robustness) property: if the input granules are sufficiently close to input concepts then the output granule is sufficiently included in the target concept provided this property is preserved locally [46.10].

Conclusions We have discussed methods for decompostion of information granules as a way to extract from data productions used to derive AR-schemes. Searching for relevant patterns for information granule decomposition can be based on methods for tuning parameters of rough set approximations of fuzzy cuts or concepts defined by differences between cuts [46.13], [46.16], i.e., by using so called rough-fuzzy granules. In this case, pattern languages consist of parameterized expressions describing the rough set approximations of parts of fuzzy concepts being fuzzy cuts or differences between cuts. Hence, an interesting research direction related to the development of new hybrid roughfuzzy methods arises aiming at developing algorithmic methods for rough set approximations of such parts of fuzzy sets relevant for information granule decomposition. In our further study we plan to implement the proposed strategies and test them on mentioned above real life data. This will recquire: (i) to develop ontologies for considered applications, (ii) further development of methods for extracting productions from data on the basis of decomposition, and (iii) synthesis methdos for AR-schemes from productions. These methods will make it possible to reason by means of sensor measurements along inference schemes over ontologies (i.e., inference schemes over some standards) by means of attached to them AR-schemes discovered from backround knowledge (including ontologies) and experimental data. Acknowledgements. The research of Andrzej Skowron has been supported by the State Committee for Scientific Research of the Republic of Poland (KBN) research grant 8 T11C 025 19 and partially by the Wallenberg Foundation grant. The research of Jaroslaw Stepaniuk has been supported by the State Committee for Scientific Research of the Republic of Poland (KBN) research grant 8 T11C 025 19. The research of James Peters has been supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) research grant 185986.

References 46.1 Doherty, P., L  ukaszewicz, W., Skowron A., Szalas, A.: Combining rough and crisp knowledge in deductive databases (submitted). 46.2 D¨ untsch, I. (Ed.): Spatial Reasoning, Fundamenta Informaticae 45(12)(2001) (special issue)

46. Extracting Patterns Using Information Granules

363

46.3 Hirano, S., Inuiguchi, M., Tsumoto, S. (Eds.): Proc. of the JSAI International Workshop on Rough Set Theory and Granular Computing (RSTGC’01), May 20-22, 2001, Matsue, Shimane, Japan, Bulletin of International Rough Set Society 5(1-2) (2001) 46.4 Kargupta H., Chan Ph.: Advances in Distributed and Parallel Knowledge Discovery, AIII Press/MIT Press, Cambridge (2001) 46.5 Nguyen, H.S., Skowron, A., Stepaniuk, J.: Granular computing: A rough set approach, Computational Intelligence (2001) (to appear) 46.6 Pawlak, Z.: Rough Sets. Theoretical Aspects of Reasoning about Data, Kluwer Academic Publishers, Dordrecht (1991) 46.7 Peters, J.F., Ramanna, S., Skowron, A., Stepaniuk, J., Suraj, Z., Borkowsky, M.: Sensor fusion: A rough granular approach, Proc. of Int. Fuzzy Systems Association World Congress (IFSA’01), Vancouver, July (2001) (to appear) 46.8 Peters, J.F., Skowron, A. Stepaniuk, J.: Rough granules in spatial reasoning, Proc. of Int. Fuzzy Systems Association World Congress (IFSA’01), Vancouver, July (2001) (to appear) 46.9 Polkowski, L., Skowron, A.: Rough mereological approach to knowledgebased distributed AI, in: (Eds.) J.K. Lee, J. Liebowitz, and J.M. Chae, Critical Technology, Proc. of the Third World Congress on Expert Systems, February 5-9, Seoul, Korea, Cognizant Communication Corporation, New York (1996) 774–781 46.10 Polkowski, L., Skowron, A.: Towards adaptive calculus of granules, in: [46.20] 30 201–227 46.11 Polkowski, L., Skowron, A.: Grammar systems for distributed synthesis of approximate solutions extracted from experience, (Eds.) Paun, G., Salomaa, A., Grammar Systems for Multiagent Systems, Gordon and Breach Science Publishers, Amsterdam (1999) 316–333 46.12 Polkowski L., Skowron A.: Rough mereological calculi of granules: A rough set approach to computation, Computational Intelligence (to appear) 46.13 Skowron, A.: Toward intelligent systems: Calculi of information granules, in: [46.3] 9–30 46.14 Skowron, A., Stepaniuk, J.: Tolerance approximation spaces Fundamenta Informaticae 27(2-3) 245–253 46.15 Skowron, A., Stepaniuk, J.: Information granules: Towards foundations of granular computing, International Journal of Intelligent Systems 16(1) (2001) 57–86 46.16 Skowron, A., Stepaniuk, J., Peters,J.F.: Extracting patters using information granules, in: [46.3] 135–142 46.17 Stone, P.: Layered Learning in Multiagent Systems: A Winning Approach to Robotic Soccer, MIT Press, Cambridge (2000) 46.18 WITAS project web page: http://www.ida.liu.se/ext/witas/eng.html 46.19 Zadeh, L.A.: Fuzzy logic = Computing with words, IEEE Trans. on Fuzzy Systems 4 (1996) 103–111 46.20 Zadeh, L.A., Kacprzyk, J. (Eds.): Computing with Words in Information/Intelligent Systems 1–2, Studies in Fuzziness and Soft Computing 3031, Physica–Verlag, Heidelberg (1999) 46.21 Zadeh, L.A.: A new direction in AI: Toward a computational theory of perceptions, AI Magazine 22(1) (2001) 73–84

47. Classification Models Based on Approximate Bayesian Networks ´ ˛zak Dominik Sle Polish-Japanese Institute of Information Technology Koszykowa 86, 02-008 Warsaw, Poland [email protected]

Approximate Bayesian networks are applied to construction of the new case classification schemes. Main topics of their extraction from empirical data are discussed.

47.1 Introduction A Bayesian network (BN) is a directed acyclic graph (DAG) designed to represent knowledge about probabilistic conditional independence statements between features ([47.4]). One can model data by extraction of approximate BNs with possibly low number of edges, but still approximately preserving information entropy of data (cf. [47.9]). This idea agrees with a common principle of tending to possibly short descriptions of models, what is assumed to provide the best knowledge generalization abilities ([47.2, 47.5, 47.6, 47.7]). We show how methodology based on approximate Bayesian networks can be applied to the new case classification problem. We introduce the Bayesianlike decision model, which classifies new cases along the structure of BN with decision attribute as a root.

47.2 Frequencies in Data It is assumed that data can be represented as an information system A = (U, A), where each attribute a ∈ A is identified with function a : U → Va , for Va denoting the set of values on a. Let us write A = a1 , . . . , an ! according to some ordering over the set of attributes. For any B ⊆ A, function B : U → VBU labels objects u ∈ U with vectors B(u) = ai1 (u), . . . , aim (u)!, where values of successive attributes aij ∈ B, j = 1, . . . , m. The set VBU = {B(u) : u ∈ U } gathers all vectors of values on B, which occur in A. Reasoning about data can be stated, e.g., as the classification problem concerning a distinguished decision to be predicted under information provided over the rest of attributes. For this purpose, one represents data as a decision table A = (U, A ∪ {d}), d ∈ / A. To express conditions→decision dependencies, one can use frequencies of occurrence of vd ∈ Vd conditioned by wB ∈ VBU , of the form T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 64 − 3 69, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

47. Classification Models Based on Approximate Bayesian Networks

PA (vd /wB ) =

|{u ∈ U : B(u) = wB ∧ d(u) = vd }| |{u ∈ U : B(u) = wB }|

365

(47.1)

Then, for a given α ∈ [0, 1], α-inexact decision rule B = wB ⇒α d = vd is satisfied, iff PA (vd /wB ) ≥ α, i.e., iff for at least α · 100% of objects u ∈ U such that B(u) = wB we have also d(u) = vd . The strength of the rule is provided by quantity PA (wB ) = |{u ∈ U : B(u) = wB }| / |U |. It corresponds to the chance that an object u ∈ U will be recognized, i.e., it will satisfy the left side of the rule. Frequencies were introduced to rough sets as rough membership functions ([47.3]). The rough set principle of reduction of possibly large amount of redundant information takes in their context the following form: Definition 47.2.1. Given A = (U, A ∪ {d}), we say that B ⊆ A preserves frequency of d iff for each u ∈ U we have PA (d(u)/B(u)) = PA (d(u)/A(u)). If, additionally, there is no proper subset of B satisfying such a condition, then B is called a frequency decision reduct. Several alternative definitions of a frequency-based reduct were proposed within the rough set framework (cf. [47.5, 47.6]). One can mention about the following aspects of adaptation of frequencies to the rough set methodology: Remark 47.2.1. If we treat PA as the empirical probability for the product space over the set of random variables A ∪ {d}, then preserving frequency of d by B means that d is independent on A \ B conditioned by B. So, each frequency decision reduct is actually a Markov boundary of d within A ([47.4]). Remark 47.2.2. Frequency distribution provides the basis for expressing inexact dependencies in various ways. For instance, the set approximations or generalized decision functions developed directly within rough sets ([47.2]) can be derived from PA (cf. [47.8]).

47.3 Approximate Independence Condition for preserving frequency turns out to be too rigorous with respect to possible noises or fluctuations in real life data. This is the general problem of dealing with probabilistic conditional independence (PCI) while analyzing empirical data. In [47.9] the information entropy-based approximation of PCI was proposed. Definition 47.3.1. Let A = (U, A) and X, Y ⊆ A be given. Entropy of X conditioned by Y is defined by HA (X/Y ) = −

1  log2 PA (X(u)/Y (u)) |U | u∈U

(47.2)

366

´ ˛zak D. Sle

Definition 47.3.2. For ε ∈ [0, 1), A = (U, A), X, Y, Z ⊆ A, we say that X is ε-approximately independent on Z conditioned by Y (we will denote such ε (X/Y /Z)), iff a fact by IA HA (X/Y ) + log2 (1 − ε) ≤ HA (X/Y ∪ Z)

(47.3)

If A takes the form of A = (U, A ∪ {d}), we say that B ⊆ A ε-approximately ε preserves frequency of d, iff IA (d/B/A \ B) holds. If, additionally, there is no proper subset of B satisfying such a condition, then B is called an εapproximate frequency decision reduct. Proposition 47.3.1. The notions of a frequency decision reduct and a 0approximate frequency decision reduct are equivalent. According to Remark 47.2.1, ε-approximate frequency decision reducts can be treated as ε-approximate Markov boundaries of d. By tuning ε ∈ [0, 1), we can search for smaller boundaries ”ε-almost” preserving entropy-based information about decision. Theorem 47.3.1. Let ε ∈ [0, 1) be given. The problem of finding minimal ε-approximate frequency decision reduct is NP-hard. One can deal with the above problem by adaptation of techniques developed in [47.5, 47.6], devoted to searching for decision reducts of various types.

47.4 Bayesian Classification One of the aims of searching for approximate reducts is to improve the ability of classification of new cases. Any B ⊆ A corresponds to the bunch of possibly inexact rules B = B(u) ⇒PA (d(u)/B(u)) d = d(u) indexed by successive objects u ∈ U . If B is an ε-approximate frequency decision reduct, then elements of the above bunch imply particular decision classes in a way ”ε-close” to decision rules based on the whole A. If B is substantially smaller than A, then the rules generated by B are shorter and stronger. Thus, they usually recognize new cases more effectively. The classification process can also correspond to the rules with decision situated at their left sides. This is the case for the Bayesian methods (cf. [47.1, 47.4, 47.7]). A new case with values equal to those of some object u ∈ U can be classified as, e.g., having decision value v = arg maxvd ∈Vd PA (A(u)/vd ), i.e. the value on d for which the observed vector on A occurs the most frequently in A. To improve the ability of the new case recognition, one can set up an ordering A = a1 , . . . , an ! and note that PA (A(u)/d(u)) =

n  i=1

PA (ai (u)/d(u), a1 (u), . . . , ai−1 (u))

(47.4)

47. Classification Models Based on Approximate Bayesian Networks

367

Proposition 47.4.1. Let A = (U, A ∪ {d}) with ordering A = a1 , . . . , an ! be given. Assume that for each i = 1, . . . , n, a frequency decision reduct Bi for table Ai = (U, {d, a1 , . . . , ai−1 } ∪ {ai }) is provided. For each u ∈ U , we have  arg max PA (A(u)/vd ) = arg max PA (ai (u)/vd , Bi \ {d}(u)) vd ∈Vd

vd ∈Vd

i: d∈Bi

(47.5) According to the above equality, there is no need to consider probabilities corresponding to subsets Bi not including d, since they are independent on the choice of decision value. We thus obtain a new formula for the new case classification, which is comparable to the previous one over vectors occurring in data. In case of combinations not included in VAU , it remains to trust into the generalization abilities of the classification scheme based on the right side of (47.5). Obviously, these abilities could be still improved by considering subsets Bi as approximate decision reducts. Then, however, one must remember that outcomes of classification based on the right side of (47.5) would be just ”ε-close” to those obtained by application of the left one.

47.5 Approximate Bayesian Networks The ordered frequency models turn out to be closely related to the notion of a Bayesian network ([47.4]) – a tool for the graphical representation of knowledge about probabilistic independence statements, by using the structure of a directed acyclic graph (DAG). In its approximate form, the notion of a Bayesian network can be defined as follows: − → Definition 47.5.1. For given  and A = (U, A), DAG D = (A, E ) is called an ε-approximate Bayesian network (an ε-BN, in short), iff ε (X/Y /Z)] ∀X,Y,Z⊆A [ X/Y /Z!D ⇒ IA

(47.6)

where by X/Y /Z!D we mean that X is d-separated from Z by Y , i.e, that any path between any elements of X and Z comes through (1) a serial or diverging connection covered by some node in Y , or (2) a converging connection not in Y , with no directed path towards any node in Y . − → Theorem 47.5.1. Given , A = (U, A) and DAG D = (A, E ), let us define the entropy of D by  − → − → HA (a/{b ∈ A : b, a! ∈ E }) (47.7) HA ( E ) = a∈A

− → If inequality HA ( E ) + log2 (1 − ε) ≤ HA (A) holds, then D is an ε-BN for A.

368

´ ˛zak D. Sle

One can consider Bayesian nets for decision tables as well. Actually, the construction of the product at the right side of (47.5), based on reducts calculated along a given ordering, corresponds to the structure of BN over A∪{d}, with the root in d. Theorem 47.5.1, applied to decision tables, results with a conclusion that similar classification schemes may be worth considering also as based on approximate reducts. Definition 47.5.2. Let  and A = (U, A ∪ {d}), A = a1 , . . . , an !, be given. We say that B = B1 , . . . , Bn ! is an ε-approximate ordered frequency model for A, iff there are thresholds ε1 , . . . , εn ∈ [0, 1) satisfying inequality (1 − ε1 ) · · · · · (1 − εn ) ≥ 1 − ε, such that Bi is an εi -approximate frequency decision reduct for Ai = (U, {d, a1 , . . . , ai−1 } ∪ {ai }), for each i = 1, . . . , n. Proposition 47.5.1. Let  and A = (U, A∪{d}) be given. Any ε-approximate − → ordered frequency model B = B1 , . . . , Bn ! induces the ε-BN D = (A∪{d}, E ) − → n defined by putting E = i=1 { b, ai ! : b ∈ Bi }.

47.6 Conclusions We presented the Bayesian-like classification model based on approximate frequency decision reducts, extracted from training data with respect to an ordering over conditional attributes. It turned out to have much in common with modeling data with approximate Bayesian networks introduced in [47.9]. We believe that presented methodology will provide new possibilities of application of Bayesian networks to the real life data analysis. Acknowledgements. The work is supported by the grant of Polish National Committee for Scientific Research (KBN) No. 8T11C02417.

References 47.1 Pawlak, Z.: Decision rules, Bayes’ rule and rough sets. In: Proc. of RSFDGrC’99, Yamaguchi, Japan, LNAI 1711 (1999) pp. 1–9. 47.2 Pawlak, Z.: Rough sets – Theoretical aspects of reasoning about data. Kluwer Academic Publishers (1991). 47.3 Pawlak, Z., Skowron, A.: Rough membership functions. In: Advances in the Dempster Shafer Theory of Evidence, John Wiley & Sons (1994) pp. 251–271. 47.4 Pearl, J.: Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann (1988). 47.5 Polkowski, L., Skowron, A. (eds.): Rough Sets in Knowledge Discovery. Physica Verlag (1998) parts 1, 2. 47.6 Polkowski, L., Tsumoto, S., Lin, T.Y. (eds.): Rough Sets in Soft Computing and Knowledge Discovery: New Developments. Physica Verlag (2000). 47.7 Rissanen, J.: Modeling by the shortest data description. Authomatica, 14 (1978) pp. 465–471.

47. Classification Models Based on Approximate Bayesian Networks

369

´ ˛zak, D.: Normalized decision functions and measures for inconsistent deci47.8 Sle sion tables analysis. Fundamenta Informaticae, 44/3 (2000) pp. 291–319. ´ ˛zak, D.: Foundations of Entropy-Based Bayesian Networks: Theoretical Re47.9 Sle sults & Rough Set Based Extraction from Data. In: Proc. of IPMU’00, July 3–7, Madrid, Spain (2000) vol. 1, pp. 248–255.

48. Identifying Adaptable Components – A Rough Sets Style Approach Yoshiyuki Shinkawa1 and Masao J. Matsumoto2 1

2

IBM Japan, Ltd, Systems Laboratory , 1-1, Nakase, Mihama-ku,Chiba-shi, Chiba Japan [email protected] The University of Tsukuba, Graduate School of Systems Management, 3-29-1, Otsuka, Bunkyo-ku,Tokyo Japan

This paper presents a formal approach to identifying partial adaptability of software components. First we discuss the partial adaptability of components with the same arity (or interface) as a requirement. Then we extend the approach to the components with the different arities. Rough Set Theory (RST)-like method is used to identify algebraic equivalency between the components and the requirements, on which the adaptability is based.

48.1 Introduction Component based development imposes several new difficulties on us, in spite of many advantages. One of the critical difficulties is that there are no appropriate ways to identify adaptable components, since there are no comprehensive measures to evaluate the adaptation of software components. While the most previous component based approaches focused on full adaptability of the components, this paper discusses partial adaptability and component collaboration.

48.2 Defining Adaptation of Software Components There are many aspects in the requirements to software, however most essential and imperative one is functional adaptation, which implies each software component performs desired data transformation [48.5]. Since requirements and software components deal with many types of data in order to define functionality of them, the above transformation rules are expressed in the form of S-sorted functions which compose (many-sorted) Σ algebra [48.1]. Σ algebra provides an interpretation for the signature Σ = (S, Ω), where S is a set of sorts, and Ω is an S ∗ × S sorted set of operation names. S ∗ is the set of finite sequences of elements of S. A Σ algebra is an algebra (A, F ), where 1. A = {Aσ |σ ∈ S} (a set of carriers) and fA : Aσ1 × · · · × Aσn → Aσ . 2. F = {fA |f ∈ Ωσ1 ...σn ,σ } T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 70− 3 74 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

48. Identifying Adaptable Components

371

S-sorted function fA is said to have arity σ1 . . . σn and result sort σ. Operational equivalency between two Σ algebras A and B is evaluated by Σ homomorphism , which is defined as a family of functions η = {ησ |σ ∈ S, ησ : Aσ → Bσ } such that     ∀f ∈ Ωσ1 ...σn ,σ , ∀ai ∈ Aσi [ησ fA (a1 , . . . , an ) = fB ησ1 (a1 ), . . . , ησn (an ) ]

where A=(A, F ) and B=(B, G) are Σ algebras. fA and fB are the functions in F and G respectively, or elements of A and B if n = 0. Each requirement compose a Σ algebra including only one function, and so does each component. If the domain of definition in such function is finite, or countable, each function can be represented in the form of a decision table [48.4]. Assuming a requirement and a component fσ1 ...σn ,σ : D (⊆ Aσ1 × · · · × Aσn ) −→ Aσ gσ1 ...σn ,σ : E (⊆ Bσ1 × · · · × Bσn ) −→ Bσ are given, those functions f and g are represented by decision tables as shown in Table 48.2. In those tables ith row or kth row of the table f or g means f (ui1 , . . . , uin ) = vi or g(xk1 , . . . , xkn ) = yk respectively. Table 48.1. A decision table of the requirement f and the component g U (r) 1 .. . i .. .

The table of f Aσ1 . . . Aσj . . . Aσn u11 . . . u1j . . . u1n .. .. .. . . . ui1 . . . uij . . . uin .. .. .. . . .

U (c) 1 .. . k .. .

Aσ v1 .. . vi .. .

The table of g Bσ1 . . . Bσj . . . Bσn x11 . . . x1j . . . x1n .. .. .. . . . xk1 . . . xkj . . . xkn .. .. .. . . .

Bσ y1 .. . yk .. .

48.3 Identifying One-to-One Component Adaptation When a requirement is expressed in the form of S-sorted function with the arity σ1 . . . σn and the result sort σ, the adaptable components to it must have the same arity and result sort, if the adaptation is evaluated by Σ homomorphism. The carriers Aσj and Bσj can be regarded as attributes which can classify U (r) and U (c) by them [48.3]. In order to identify Σ homomorphism from the requirement to the component, we examine all the possible sets of mappings {ησj : Aσj −→ Bσj } (j = 0, 1, . . . , n) and there are the permutations injections.

qj Ppj

kinds of mappings as ησj , if they are

372

Y. Shinkawa and M.J. Matsumoto

If {ησj } is Σ homomorphism, the formula     ∀i ∈ U (r) [g ησ1 (ui1 ), . . . , ησn (uin ) = ησ f (ui1 , . . . , uin ) ]

(48.1)

holds. This formula is equivalent to the following formula ∀i ∈ U (r) ∃k ∈ U (c)

[xkj = ησj (uij ) and yk = ησ (vi )]

(48.2)

when f and g are expressed in the forms of Table 48.2, since they include all the possible data transformation by f and g respectively. The formula (48.2) implies that each row in Table f is mapped into a single row in Table g by Σ homomorphism {ησj }. When the above Σ homomorphism does not exist, we have to reduce the requirement f in the following way in order to define Σ homomorphism for partial adaptation. $ # 1. Let E = {ησj | j = 0, 1, . . . , n} be a set of all the possible sets of mapping {ησj : Aσj −→ Bσj } (j = 0, 1, . . . , n) where ησ0 = ησ . 2. For each {ησj | j = 0, 1, . . . , n}, classify U (r) for the requirement f into (r) (r) UA ({ησj }) and UN ({ησj }), where (r) UA ({ησj }) = {i | ∃k ∈ U (c) [ησj (uij ) = xkj , and ησ (vi ) = yk ]} and (r) (r) UN ({ησj }) = U (r) − UA 3. Select the set of mapping {ησ∗j | j = 0, 1, . . . , n} ∈ E (r)

which makes the cardinality of UA maximum, that is,  (r)   (r)  ∀{ησj } ∈ E [card UA ({ησ∗j }) ≥ card UA ({ησj }) ] holds. (r)

UA ({ησj }) represents the maximum adaptation of the component g to the (r) (r) requirement f . We denote UA ({ησj }) = UA (g). By extracting the rows (r) belonging to UA (g) from Table f , we can define the new function f ∗ ⊆ f . The Σ algebra B = (B, G) is evidently Σ homomorphic to the Σ algebra A∗ = (A∗ , F ∗ ), where A∗ = {A∗σ1 , . . . , A∗σn , A∗σ }, F ∗ = {f ∗ }, (r) (r) A∗σj = {xij | i ∈ UA } and A∗σ = {yi | i ∈ UA }. The function f ∗ is re∗ ferred to as the restriction of f into A . After examining all the functions in G, that is, the set of S-sorted functions with the same arity and result sort as the requirement f , if

(r)

f ∗ , or equivalently U (r) = UA (g) f= g∈G

g∈G

holds, the requirement f is satisfiable by #the set of the functions G$ = {g}, using the set of Σ homomorphism E ∗ = {ησ∗j | (j = 0, 1, . . . , n)} . When implementing the requirement f by the above set of components G = {g}, we need the knowledge on which part of the function f (or the domain of definition of f ) is satisfied by each component g. Obviously the less components require the less knowledge, and it is desirable from practical viewpoint.

48. Identifying Adaptable Components

373

Identifying the minimum set of components is a kind of the set cover problem with identical costs, and a near optimal solution can be found by the greedy method [48.2]. This method is expressed in the following way in our situation.  (r)  (r) 1. Select g1 ∈ G which make card UA (g1 ) maximum. Denote this UA (g1 ) ∗ as U1 . (r) 2. Select g2 ∈ G (g2 = g1 ) which make card(UA (g2 )∩UA∗ ) maximum, where (r) U means the complementary set of U . Denote this UA (g2 ) as U2∗ . 3. Repeat the above procedure. Each time we identify gi ∈ G which make i−1     card UA∗ (gi ) ∩ ( Uk∗ ) maximum. k=1

4. Terminate the iteration if we identify gn ∈ G which satisfies Un∗ ∪ · · · ∪ U1∗ = U (r) G ∗ = {g1 , . . . , gn } is the minimum set of components that satisfies the requirement f .

48.4 Identifying One-to-Many Component Adaptation Even though a requirement can not be satisfied by a set of components by the above way, there could be a possibility to satisfy it by collaboration of several components. Assuming there is a pair of components g1 , g2 ! for a requirement f , which satisfy fσ1 ...σn ,σ : D −→ Aσ (D ⊆ Aσ1 × · · · × Aσn ) g1 σ1 ...σm ,ρ : E1 −→ Bρ (E1 ⊆ Bσ1 × · · · × Bσm )

(48.3)

g2 σm+1 ...σn ,σ : E2 −→ Bσ (E2 ⊆ Bσm+1 × · · · × Bσn )

(48.5)





∃m (m < m ≤ n) [Bσm = Bρ ]

(48.4) (48.6)

Since the order of the sorts in the requirement f , that is the arity of f , is arbitrary and does not affect the data transformation rule of f , we can reorder the arity of f in order to satisfy the above conditions. When g1 and g2 are represented in the form of decision tables, we can connect g1 and g2 in order to compose the new function with the same arity and result sort as f in the following way, supposing f is represented in the form of Table 48.2. 1. Let ησj and ησ be mappings ησj : Aσj −→ Bσj (1 ≥ j ≥ n) and ησ : Aσ −→ Bσ 2. Connect two rows xk1 , . . . , xkm ! and xk ,m+1 , . . . , xk n !, which satisfy xkj = ησj (uij ) (j = 1, . . . , m), xk j  = ησj (uij  ) (j  = m + 1, . . . , n), and yj = xk m 3. Define the new decision table which is composed of the rows xk1 , . . . , xkm , xk ,m+1 , . . . , xk n , zk !

374

Y. Shinkawa and M.J. Matsumoto

There are multiple {ησj } as discussed in the previous section, therefore there could be multiple decision tables composed through the above way. We denote the table with maximum number of rows as g ∗ , which represents a new S-sorted function. We obtain a subset of f corresponding to g ∗ , which is derived from Table f by extracting the rows indexed by i selected in the step 2 in the above procedure. We denote this subset as f ∗ . The g ∗ is Σ homomorphic to the f ∗ as we discussed in Section 48.3. By examining all the possible pair of the above g1 and g2 , we can identify the set of Σ homomorphic function pairs { f ∗ , g ∗ !}. If

f∗ = f g1 ,g2 

holds, the requirement f can be satisfied by the set of S-sorted functions { g1 , g2 !}. We can identify the minimum set of the above pair { g1 , g2 !} in the similar way to Section 48.3. This approach can be extend to the S-sorted function tuple g1 , . . . , gl ! similarly.

48.5 Conclusions A formal approach to identifying adaptable software components to requirements is proposed in this paper. The adaptation is evaluated by Σ homomorphism between the requirements and the components. Unlike the previous approaches, we define partial adaptation based on decision tables which represent the requirements and the components. We also defined two forms of adaptation, that is, one-to-one adaptation and one-to-many adaptation.

References 48.1 Astesiano, E., Kreowski, H.J., Br¨ uckner, B.K. (1999): Algebraic Foundation of System Specification. (IFIP State-of-the-Art Reports) Springer, Berlin Heidelberg New York 48.2 Fujito, T (2000): Approximation Algorithms for Submodular Set Cover with Applications. IEICE Trans. on Inf. and Syst. Vol.E83-D. No.3. 480–487 48.3 Pawlak, Z. (1992): Rough Sets : Theoretical Aspects of Reasoning About Data. Kluwer, Dordrecht Boston London 48.4 Shinkawa, Y. Matsumoto, M.J. (2000): Knowledge-Based Software Composition Using Rough Set Theory. IEICE Trans. on Inf. and Syst. Vol.E83-D. No.4. 691–700 48.5 Shinkawa, Y. Matsumoto, M.J. (2001): Identifying the Structure of Business Processes for Comprehensive Enterprise Modeling. IEICE Trans. on Inf. and Syst. Vol.E83-D. No.4. 691–700

49. Rough Measures and Integrals: A Brief Introduction Zdzislaw Pawlak1 , James F. Peters2 , Andrzej Skowron3 , Z. Suraj4 , S. Ramanna2 , and M. Borkowski2 1 2 3 4

Polish Academy of Sciences, Baltycka 5, 44-000 Gliwice, Poland Computer Engineering, Univ. of Manitoba, Winnipeg, MB R3T 5V6 Canada Institute of Mathematics, Warsaw Univ., Banacha 2, 02-097 Warsaw, Poland Univ. of Information Technology and Management, H. Sucharskiego 2, 35-225 Rzesz´ ow, Poland

This paper introduces a measure defined in the context of rough sets. Rough set theory provides a variety of set functions that can be studied relative to various measure spaces. In particular, the rough membership function is considered. The particular rough membership function given in this paper is a non-negative set function that is additive. It is an example of a rough measure. The idea of a rough integral is revisited in the context of the discrete Choquet integral that is defined relative to a rough measure. This rough integral computes a form of ordered, weighted ”average” of the values of a measurable function. Rough integrals are useful in culling from a collection of active sensors those sensors with the greatest relevance in a problem-solving effort such as classification of a ”perceived” phenomenon in the environment of an agent.

49.1 Introduction This paper introduces a measure defined in the context of rough sets [49.3]. In this paper, we investigate measures defined on a family ℘(X) of all subsets of a finite set X, i.e. on the powerset of X. A fundamental paradigm in rough set theory is set approximation. Hence, there is interest in discovering a family of measures useful in set approximation. By way of practical application, an approach to fusion of homogeneous sensors deemed relevant in a classification effort is considered (see, e.g., [49.6]). Application of rough integrals has also been considered recently relative to sensor signal classification by intelligent agents [49.8] and by web agents [49.9]. This research also has significance in the context of granular computing [49.10]. This paper is organized as follows. Section 49.2 presents a brief introduction to classical additive set functions. Basic concepts of rough set theory are presented in Section 49.3. The discrete Choquet integral is defined relative to a rough measure in Section 49.4. A brief introduction to sensor relevance is given in Section 49.5. T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 75− 3 79, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

376

Z. Pawlak et al.

49.2 Classical Additive Set Functions This section gives a brief introduction to one form of additive set functions in measure theory. Let card(X) denote the cardinality of a finite set X (i.e., the number of elements of set X). Definition 49.2.1. Let X be a finite, non-empty set. A function λ : ℘(X) → # where # is the set of all real numbers is called a set function on X. Definition 49.2.2. Let X be a finite, non-empty set and let λ be a set function on X. The function λ is said to be additive on X iff λ(A ∪ B) = λ(A) + λ(B) for every A, B ∈ ℘(X) such that A ∩ B = ∅ (i.e., A and B are disjoint subsets of X). Definition 49.2.3. Let X be a finite, non-empty set and let λ be a set function on X. A function λ is called to be non-negative on X iff λ(Y ) ≥ 0 for any Y ∈ ℘(X). Definition 49.2.4. Let X be a set and let λ be a set function on X. A function λ is called to be monotonic on X iff A ⊆ B implies that λ(A) ≤ λ(B) for every A, B ∈ ℘(X). A brief introduction to the basic concepts in rough set theory (including the introduction of an additive rough measure) is briefly given in this section.

49.3 Basic Concepts of Rough Sets Rough set theory offers a systematic approach to set approximation [49.2]. To begin, let S = (U, A) be an information system where U is a non-empty, finite set of objects and A is a non-empty, finite set of attributes, where a : U → Va for every a ∈ A. For each B ⊆ A, there is associated an equivalence relation IndA (B) such that IndA (B) = {(x, x ) ∈ U 2 | ∀a ∈ B.a(x) = a(x )} If (x, x ) ∈ IndA (B), we say that objects x and x are indiscernible from each other relative to attributes from B. The notation [x]B denotes equivalence classes of IndA (B). Definition 49.3.1. Let S = (U, A) be an information system, B ⊆ A, u ∈ U and let [u]B be an equivalence class of an object u ∈ U of IndA (B). The set function B μB u : ℘ (U ) → [0, 1], where μu (X) =

card (X ∩ [u]B ) card ([u]B )

for any X ∈ ℘(U ) is called a rough membership function (rmf ).

(49.1)

49. Rough Measures and Integrals

377

The form of rough membership function in Def. 49.3.1 is slightly different from the classical definition where the argument of the rough membership function is an object x and the set X is fixed [49.3]. Definition 49.3.2. Let u ∈ U . A non-negative and additive set function ρu : ℘(X) → [0, ∞) defined by ρu (Y ) = ρ (Y ∩ [u]B ) for Y ∈ ℘(X), where ρ : ℘(X) → [0, ∞) is called a rough measure relative to U/IndA (B) and u on the indiscernibility space (X, ℘(X), U/IndA (B)). The rough membership function μB u : ℘(X) → [0, 1] is a non-negative set function [49.4]. Proposition 49.3.1. (Pawlak et al. [49.4]) The rough membership function μB u as defined in Definition 49.3.1 ( formula (49.1)) is additive on U . Proposition 49.3.2. (X, ℘(X), U/IndA (B), {μB u }u∈U ) is a rough measure space over X and B. Other rough measures based on upper {lower} approximations are possible but consideration of these other measures is outside the scope of this paper.

49.4 Rough Integrals Rough integrals of discrete functions were introduced in [49.5]. In this section, we consider a variation of the Lebesgue integral, the discrete Choquet integral defined relative to a rough measure. In what follows, let X = {x1 , . . . , xn } be a finite, non-empty set with n elements. The elements of X are indexed from 1 to n. The notation X(i) denotes the set {x(i) , x(i+1) , . . . , x(n) } where i ≥ 1 and n = card(X). The subscript (i) is called a permutation index because the indices on elements of X(i) are chosen after a reordering of the elements of X. This reordering is ”induced” by an external mechanism. Definition 49.4.1. Let ρ be a rough measure on X where the elements of X are denoted by x1 , . . . , xn . The discrete Choquet integral of f : X → #+ with respect to the rough measure ρ is defined by n  (f (x(i) ) − f (x(i−1) ))ρ(X(i) ) f dρ = i=1

where •(i) specifies that indices have been permuted so that 0 ≤ f (x(i) ) ≤ · · · ≤ f (x(n) ), X(i) := {x(i) , . . . , x(n) }, and f (x(0) ) = 0. This definition of the Choquet integral is based on a formulation in Grabisch [49.1], and applied in [49.2], [49.7]. The rough measure ρ(X(i) ) value serves as a ”weight” of a coalition (or combination) of objects in set X(i) relative to f (x(i) ). It should be observed that in general the Choquet integral has the effect of ”averaging” the values of a measurable function. This averaging closely resembles the well-known Ordered Weighted Average (OWA) operator [49.11].

378

Z. Pawlak et al.

Proposition 49.4.1. Let 0 < s ≤ r. If a(x) ∈ [s, r] for all x ∈ Xa , then % a dμeu ∈ (0, r] where u ∈ U .

49.5 Relevance of a Sensor In this section, we briefly consider the measurement of the relevance of a sensor using a rough integral. A sensor is considered relevant in a classification % effort in the case where a dμeu for a sensor a is close enough to some threshold in a target interval of sensor values. Assume that a denotes a sensor that responds to stimuli with measurements that govern the actions of an agent. Let {a} = B ⊆ A where a : U → [0, 0.5] where each sample sensor value a(x) is rounded to two decimal places. Let (Y, U − Y ) be a partition defined by an expert and let [u]e denote a set in this partition containing u for a selected u ∈ U . We further assume the elements of [u]e are selected relative to an interval (u − ε, u + ε) for a selected ε ≥ 0. We assume a decision system (Xa , a, e) is given for any considered sensor a such that Xa ⊆ U, a : Xa → #+ and e is an expert decision restricted to Xa defining a partition (Y ∩Xa , (U −Y )∩Xa ) of Xa . Moreover, we assume that Xa ∩ [u]e = ∅. The set [u]e is used to classify sensors and is given the name ”classifier”. Let u ¯ denote the average value in the classifier [u]e , and let δ ∈ [0, 1]. Then, for example, the selection R of the most relevant sensors in a set of sensors is found using & &  ' & & e & & u)& ≤ δ R = ai ∈ B : & ai μu − a(¯ % In effect, the integral ai dμeu serves as a filter inasmuch as it ”filters” out all sensors with integral values not close enough to a(¯ u).

49.6 Conclusion Rough set theory provides a variety of set functions that can be studied relative to various measure spaces. In particular, the rough membership function is considered. The particular rough membership function given in this paper is a non-negative set function which is additive and, hence, is an example of a rough measure. We are interested in identifying those sensors considered relevant in a problem-solving effort. The rough integral introduced in this paper serves as a means of distinguishing relevant and non-relevant sensors in a classification effort. Acknowledgment. The research of Sheela Ramanna and James Peters has been supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) research grant 194376 and research grant 185986, respectively. These authors also wish to thank Michel Grabisch for introducing

49. Rough Measures and Integrals

379

us to capacity theory and the Choquet integral. The research of Maciej Borkowski has been supported by a grant from Manitoba Hydro. The research of Andrzej Skowron has been supported by grant 8 T11C 025 19 from the State Committee for Scientific Research (KBN) and grant from the Wallenberg Foundation. Zbigniew Suraj has been supported by grant 8 T11C 025 19 from KBN and by NSERC grant STP201862.

References 49.1 M. Grabisch, Alternative expressions of the discrete Choquet integral. In: Proc. 7th Int. Fuzzy Systems Association World Congress (IFSA’97), Prague, 25-29 June 1997, 472-477. 49.2 M. Grabisch, T. Murofushi, M. Sugeno (Eds.), Fuzzy Measures and Integrals: Theory and Applications, Berlin, Physica-Verlag, 2000. 49.3 Z. Pawlak, Rough Sets: Theoretical Aspects of Reasoning About Data, Boston, MA, Kluwer Academic Publishers, 1991. 49.4 Z. Pawlak, A. Skowron, Rough membership functions. In: R. Yager, M. Fedrizzi, J. Kacprzyk (Eds.), Advances in the Dempster-Shafer Theory of Evidence, NY, John Wiley & Sons, 1994, 251-271. 49.5 Z. Pawlak, On rough derivatives, rough integrals, and rough differential equations. ICS Research Report 41/95, Institute of Computer Science, Nowowiejska 15/19, 00-665 Warsaw, Poland, 1995. 49.6 J.F. Peters, S. Ramanna, A. Skowron, J. Stepaniuk, Z. Suraj, M. Borkowski, Sensor fusion: A rough granular approach. In: Proc. of the International Fuzzy Systems Association World Congress (IFSA’01), Vancouver, July 2001 [to appear]. 49.7 J.F. Peters, S. Ramanna, L. Han, The Choquet integral in a rough software cost estimation system. In: M. Grabisch, T. Murofushi, M. Sugeno (Eds.), Fuzzy Measures and Integrals: Theory and Applications. (Springer-Verlag, Heidelberg, Germany, 2000) 392-414. 49.8 J.F. Peters, S. Ramanna, A. Skowron, M. Borkowski, Approximate sensor fusion in a navigation agent. In: Proc. Intelligent Agents Technology, Japan, October 2001 [submitted]. 49.9 J.F. Peters, S. Ramanna, A. Skowron, M. Borkowski, Wireless agent guidance of remote mobile robots: Rough integral Approach to Sensor Signal Analysis. In: Proc. Web Intelligence, Japan, October 2001 [submitted]. 49.10 A. Skowron, Toward Intelligent Systems: Calculi of Information Granules. In: Proc. RSTGC’01, Bull. Int. Rough Set Society 5(1/2), 2001, 9-30. 49.11 R.Yager, On ordered weighted averaging aggregation operators in multicriteria decision making, IEEE Trans. on System, Man and Cybernetics 18 (1988) 183-190.

50. Association Rules in Semantically Rich Relations: Granular Computing Approach T.Y. Lin1 and Eric Louie2 1

2

Department of Mathematics and Computer Science San Jose State University, San Jose, California 95192-0103 [email protected] IBM Almaden Research Center 650 Harry Road, San Jose, CA 95120 [email protected]

In ”real world” databases, attribute domains are more than Cantor sets; the additional semantics defined, in this paper, is assumed to be carried by a binary relation. Association rules in such databases are investigated. In this paper, we show that the cost of checking the additional semantics is rather small. Some experiments are reported.

50.1 Introduction In relation theory, all attribute domains are assumed to be Cantor sets. However, in practice, they are ”real world sets,” that is, there are interact among themselves. The question is: Can such interactions be modeled mathematically? In first order logic, the real world is modeled by a Cantor set with relational structure. We follow this approach; as a first step we consider simplest case, that is, the relational structure is defined by a binary relation. Such ”real world sets;” have been called binary neighborhood system spaces, or BNS-spaces [50.4], [50.3]. This paper report the study of association rules in such semantically rich relations.

50.2 Relational Models and Rough Granular Structures A relation is a knowledge representation that maps each entity to a tuple of attribute values. Table 50.1 illustrates the knowledge representation of the universe V = {v1 , v2 , v3 , v4 , v5 }. In this view, an attribute can be regarded as a projection that maps entities to attribute values, for example in Table 50.1, the CITY attribute is the map, f : V −→ Dom(CIT Y ), which assigns, at every tuple, the element in the first column to the element in the last column. The family of complete inverse image f −1 (y) forms a T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 8 0− 3 8 4 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

50. Association Rules in Semantically Rich Relations

381

partition (equivalence relation). So each column (attribute) defines an equivalence relation. So Table 50.1 gives rise to 4 named equivalence relations. Pawlak called the pair V and a finite family of equivalence relations a knowledge base. Since knowledge bases often have different meaning, we will call it rough granular structure, or rough structure, which is a special form of binary granular structure [50.3]. Table 50.1. Information Table of Suppliers; arrows and parentheses will be suppresed V v1 v2 v3 v4 v5

→ → → → →

(S# (S1 (S2 (S3 (S4 (S5

SNAME Smith Jones Blake Clark Adams

Status TWENTY TEN TEN TWENTY THIRTY

City) C1 ) C2 ) C2 ) C1 ) C3 )

50.3 Databases with Additional Semantics Relational theory assumes everything is a Cantor set. In other words, the interactions among real world objects (entities or attribute values respectively) are ”forgotten” in the relational modeling. In practical database processing, additional semantics in attribute domain are often processed. For example, in numerical attributes, the order of numbers is often used in SQL statements by human operators. Therefore these additional semantics implicitly exist in the stored database. To capture such additional semantics in data mining, we need a data model which is semantically richer than relational. What would be the ”correct” mathematical structure of real world objects? We will follow the first order logic; the universe and attribute domains are assumed to have relational structures. As a first step, we will confine ourselves to the simplest kind of relational structure, namely, binary relations. In Table 50.2, we give an example of a binary relation defining ”near”semantics on CITY. Note that a binary relation B define a binary neighborhood Bp = {x | x B p} at every p-called a binary granular structure. A relation with such additional semantics defines a binary granular structure on the universe V , that is the universe is equipped with a finite family of named binary relations.

382

T.Y. Lin and E. Louie

Table 50.2. ”Near”-Binary Relation CIT Y C1 C1 C2 C2 C2 C3 C3

CIT Y C1 C2 C2 C1 C3 C3 C2

Table 50.3. Binary Granular Structure; Relation with additional semantics the center ∗ ∗ v1 , v 4 v2 , v 3 v5 v1 , v 4 v2 , v 3 v5

Elementary Granule encoded label S#(∗) SN AM E(∗) STATUS(10010) STATUS(01100) STATUS(00001) CIT Y (11110) CIT Y (11111) CIT Y (01101)

Attribute value meaningful name ∗ ∗ TWENTY TEN THIRTY C1 C2 C3

50.4 Mining Real World or Its Representations What is data mining? The common answer is essentially ”to find the pattern in data.” This is not entirely accurate; we would like to amend the notion as follows: The goal of data mining is to find patterns in Real World, represented by the given data. For convenience, we will denote the real world by RW and the data (knowledge representation) by KR. For example, we will not be interested in a discovered rule, say, ”all data are represented by 5 characters.” Because this is a pattern of KR, not RW. To show that a discovered pattern in a KR is, indeed, a pattern of RW. We need to show that the patterns is INVARIANT under attribute transformations. In other words, the pattern also exits in other knowledge representations. However, we can take the following alternate approach: Find the patterns in the mathematical structure, RW, of Real World. For relational data base, the mathematical structure of RW is the rough granular structure (or knowledge base, if we use Pawlak’s terminology); see Table 50.3. If we conduct the data mining in such a structure, it is RW mining; no attribute transformations are needed In this paper, we extend this approach to ”real world” databases.

50. Association Rules in Semantically Rich Relations

383

50.5 Clustered Association Rules-Mining Semantically Machine oriented model uses granules as its attribute values, so any logical formula is translated to set theoretical formula of granules. However, we should note that attribute values are semantically related, so in processing any logical formula based on attribute values, it is important that one checks the continuity (namely, see if it respects the semantics). We will call any pattern or rule that respects the semantics a clustered pattern or clustered rule. Let c and d be two attribute values in a relation Definition Clustered Association rules 1. A pair (c, d) in a given relation is one-way (c −→ d) continuous (or clustered) if every point x in the elementary neighborhood Bc there is at least one y in Bd such that (x, y) is in the given relation. 2. A pair (c, d) in a given relation is a two way continuous (or clustered) if (c −→ d) and (d −→ c) are both continuous. 3. Clustered association rule: A pair (c, d) is an association rule iff the pair is an association rule and two way continuous. 4. Soft association rule: A pair (c, d) is a soft association rule, if Card (N EIGH(c) ∩ N EIGH(d)) ≥ threshhold. [50.1], [50.2] Here is some of our experimental results: see Table 50.4. Data characteristics and meaning of comments: (1) Rows 100000; (2) Columns 16; (3) Support 500 items; (4) Main Memory size 10 mega bytes; (5)”Generated 56345 2-combinations” means ”56345 candidates of length 2 are generated”;(6) ”4973 2-large granule” means ”4973 candidates meet the support threshhold”; (7)”Continuous 4973 2-associaton rules” means ”4973 continuous association rules of length 2 are checked.” From the table, it is clear the cost of checking continuity is small.

50.6 Conclusion The advantage of data mining by granular computing are: 1. it is fast in mining classical relations, granular computing is faster than Apriori [50.5],[50.6] because the “database scan” are replaced by bit operations. 2. the use of granular computing is extended to ”real world” databases (semantically richer relations); its cost is small. Moreover, such extra semantics can be used to analyze unexpected rules [50.7].

384

T.Y. Lin and E. Louie

Table 50.4. Granular and neighborhood method: Length 0 1

Cand 0 351

2

56345

Supp

Rules

351

351

4973 4973 3

38109 3841 3841

4

3851 641 641

5

100 11 11

6

0

98756

0

0

9817

9817

Delta 0.000s 2.333s 0.000s 0.130s 20.009s 2.023s 425.902s 13.650s 0.090s 205.065s 1.552s 0.010s 1.272s 0.030s 0.000s 0.000s 0.000s 0.000s 672.066s

Comment Start Generated 351 1-combinations Continuous 1-association rules Generated 56345 2-combinations 4973 2-large granule Continuous 4973 2-association rules Generated 38109 3-combinations 3841 3-large granule Continuous 3841 3-association rules Generated 3851 4-combinations 641 4-large granule Continuous 641 4-association rules Generated 100 5-combinations. 11 5-large granule Continuous 11 5-association rules Generated 0 6-combinations. Continuous 0 6-association rules Discover Complete Totals

References 50.1 T. Y. Lin, “Data Mining and Machine Oriented Modeling: A Granular Computing Approach,” Journal of Applied Intelligence, Kluwer, Vol. 13,No 2, September/October,2000, pp.113-124. 50.2 T. Y. Lin, ”Data Mining: Granular Computing Approach.” In: Methodologies for Knowledge Discovery and Data Mining, Lecture Notes in Artificial Intelligence 1574, Third Pacific-Asia Conference, Beijing, April 26-28, 1999, 24-33. 50.3 T. Y. Lin, ”Granular Computing on Binary Relations I: Data Mining and Neighborhood Systems.” In: Rough Sets In Knowledge Discovery, A. Skoworn and L. Polkowski (eds), Springer-Verlag, 1998, 107-121. 50.4 T. Y. Lin,”Neighborhood Systems and Relational Database”. Abstract, Proceedings of CSC ’88, February, 1988, pp. 725. 50.5 Eric Louie and T.Y. Lin, “Finding Association Rules using Fast Bit Computation: Machine-Oriented Modeling.” In: Proceeding of 12th International Symposium ISMIS2000, Charlotte, North Carolina, Oct 11-14, 2000. Lecture Notes in AI 1932. 486-494. 50.6 T. Y. Lin and E. Louie, ”A Data Mining Approach using Machine Oriented Modeling: Finding Association Rules using Canonical Names.”. In: Proceeding of 14th Annual International Symposium Aerospace/Defense Sensing, Simulation, and Controls , SPIE Vol 4057, Orlando, April 24-28, 2000, pp.148-154 50.7 Balaji Padmanabhan and Alexander Tuzhilin “Finding Unexpected Patterns in Data.“ In: Data Mining and Granular Computing T. Y. Lin, Y.Y. Yao and L. Zadeh (eds), Physica-Verlag, to appear.

51. A Note on Filtration and Granular Reasoning Tetsuya Murai1 , Michinori Nakata2 , and Yoshiharu Sato1 1

2

Division of Systems and Information Engineering, Graduate School of Engineering, Hokkaido University, Kita-ku, Sapporo 060-8628, JAPAN {murahiko, ysato}@main.eng.hokudai.ac.jp Department of Environment and Information Management for Social Welfare, Josai International University, Togane, Chiba 283-855, JAPAN [email protected]

The filtration method in modal logic is considered to develop a way of formulating an aspect of granular reasoning, which, roughly speaking, means human reasoning based on granularity. The method, however, originates in purely logical problems like decidability. Then, for our purpose, an extended concept of relative filtration is newly introduced using lower and upper approximations in rough set theory. An example of reasoning process using the relative filtration is illustrated.

51.1 Introduction This paper aims to provide a small step for formulating an aspect of granular reasoning. What, however, is granular reasoning? Although, as far as the authors know, there seems to have been no consensus of what it means as a technical term, it would indicate some mechanism for reasoning using rough set theory (Pawlak[51.3]) and granular computing (Lin[51.2]). Our point of departure is the filtration method in modal logic(Chellas [51.1]). It is a standard way of proving finite determination and decidability. The basic idea of filtration method is to generate a kind of quotient model from the original one so that its set of possible worlds is finite. Usually, the method is performed using a given finite set of sentences to which we pay attention with respect to purely logical problems. When, however, we deal with problems beyond pure logic, we often find ourselves paying attention to proper subsets of such given set of sentences. For the purpose, we introduce a concept of relative filtration, an extended definition of filtration with approximation in rough set theory. Finally we illustrate a formulation of human reasoning where a model is not kept fixed but is changed into a new one using the relative filtration whenever required.

51.2 Preliminaries A modal language LML (P) is formed from a given countable set of atomic sentences P in the usual way with a standard set of logical connectives incluT. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 8 5− 3 8 9, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

386

T. Murai, M. Nakata, and Y. Sato

ding the two modal operators. For a sentence p in LML (P), its subsentences are defined in the usual recursive way. Let Sub(p) be the set of subsentences of p. A set Γ of sentences is said to be closed under subsentences just in case def Sub(p) ⊆ Γ for every p in Γ . Let Pp = P ∩ Sub(p). Also, for the set Γ of def

sentences, let PΓ = ∪p∈Γ Pp . A Kripke model M is a structure , where W is a non-empty set of possible worlds and R is an accessibility relation on W , and V is a valuation, which assigns a subset in W to each atomic sentence p in P. Define M, w |= p iff w ∈ V (p). It means that a sentence p is true at a possible world w in M. |= is extended for every compound sentence in the usual way. The set of possible def worlds pM = {w ∈ W | M, w |= p} is called the proposition of p in M. Given a Kripke model M =, let Γ be a set of sentences closed under subsentences. Two worlds w, w in W is said to be Γ -equivalent, written w∼Γ w , when, for every sentence p in Γ , M, w |= p iff M, w |= p. Then, a filtration of M through Γ (or, Γ -filtration of M, for short) is defined as a def

structure MΓ =, where WΓ = W/∼Γ , RΓ is a relation on WΓ satisfying several conditions (for details, see Chellas[51.1], p.101), and, def

for each atomic sentence p in PΓ , VΓ (p) = {[w]∼Γ | w ∈ V (p)}. With respect to the filtration, the following remarkable result is obtained: for every p in Γ , M |= p iff MΓ |= p. Note that, if |Γ | = n, then |W/∼Γ | ≤ 2n .

51.3 Relative Filtration with Approximation Given a Kripke model M =, let Γ be a set of sentences closed under subsentences and Δ be a non-empty subset of Γ , which is also assumed to be closed under subsentences. In this paper, we call elements in Γ and Γ C (= LML (P) \ Γ ), respectively, explicit and implicit sentences. Also, we call elements in Δ and Γ \ Δ, respectively, focal and marginal sentences. The filtration that we want to formulate in what follows contains the set of possible worlds WΔ and accessibility relation RΔ . The difference is its valuation (let be V  , tentatively). Truth values for every focal atomic sentence in PΔ can be defined in the same way in the usual Δ-filtration of M. In general, however, we cannot determine truth values for marginal atomic sentences in PΓ \ PΔ . For example, consider the Kripke model M with W = {w1 , w2 , · · · , w6 } and valuation V given in Fig. 51.3. Let Γ = {p1 , p2 , p3 }(= PΓ ) and Δ = {p1 , p2 }(= PΔ ). Then, by equivalence relation ∼Δ , the following quotient set of four new possible worlds (equivalence classes) is generated: WΔ = {[w1 ]∼Δ , [w2 ]∼Δ , [w4 ]∼Δ , [w6 ]∼Δ }. The truth values for p1 and p2 in PΔ can be assigned in the same way in the usual Δ-filtration because every element in each equivalence class (newly generated possible world) has the same truth value of the original valuation. Note that

51. A Note on Filtration and Granular Reasoning

387

p1 M = [w1 ]∼Δ + [w2∼Δ ], p2 M = [w1 ]∼Δ + [w4 ]∼Δ , where + denotes the direct sum. Thus, in terms of rough set theory[51.3], for every sentence p in Δ, pM is a ∼Δ -definable set. For a marginal sentence p3 , since, for all w in [w1 ]∼Δ , w ∈ V (p3 ) and, for all w in [w6 ]∼Δ , w ∈ V (p3 ), so, we can assign, respectively, 1 and 0 to p3 at two newly generated worlds [w1 ]∼Δ and [w6 ]∼Δ . For a new world [w2 ]∼Δ , on the other hand, the two original worlds w2 , w3 in [w2 ]∼Δ have different states: w2 ∈ V (p3 ) and w3 ∈ V (p3 ). Thus we no longer uniquely determine V  (p3 ) with respect to [w2 ]∼Δ . We have the same result for [w4 ]∼Δ ={w4 , w5 }. Hence, in general, we can give only a partial definition of V  . V w1 w2 w3 w4 w5 w6

p1 1 1 1 0 0 0

p2 1 0 0 1 1 0

p3 1 1 0 1 0 0

··· ··· ··· ··· ··· ··· ···

−→ Relative filtration

V [w1 ]∼Δ [w2 ]∼Δ

p1 1 1

p2 1 0

p3 1 (which of 1,0 ?)

[w4 ]∼Δ

0

1

(which of 1,0 ?)

[w6 ]∼Δ

0

0

0

Fig. 51.1. Partiality of valuation when making relative filtration.

Here note that [w1 ]∼Δ ⊆ p3 M ⊆ [w1 ]∼Δ + [w2 ]∼Δ + [w4 ]∼Δ . by which we have ∼Δ (p3 M ) = [w1 ]∼Δ , ∼Δ (p3 M ) = [w1 ]∼Δ + [w2 ]∼Δ + [w4 ]∼Δ in terms of rough set theory. Thus, we can introduce the concept of approximation in rough set theory into relative filtration. This means that we have two kinds of definition of valuation in the following way. Definition 51.3.1. For every explicit atomic sentence p in Γ , lower and upper valuation through Δ relative to Γ (or, lower and upper Δ/Γ -valuation, for short) are defined by, respectively, def

1. VΔ/Γ (p) = {[w]∼Δ | [w]∼Δ ⊆pM } = {[w]∼Δ | [w]∼Δ ⊆∼Δ (pM )} def

2. VΔ/Γ (p) = {[w]∼Δ | [w]∼Δ ∩pM = ∅} = {[w]∼Δ | [w]∼Δ ⊆∼Δ (pM )}. Thereby, we have the following two kinds of filtration: Definition 51.3.2. 1. A lower filtration of M through Δ relative to Γ (or, lower Δ/Γ -filtration of M, for short) is defined by def

MΔ/Γ = < WΔ , RΔ , VΔ/Γ >,

388

T. Murai, M. Nakata, and Y. Sato

2. An upper filtration of M through Δ relative to Γ (or, upper Δ/Γ -filtration of M) is defined by def

MΔ/Γ = < WΔ , RΔ , VΔ/Γ > . Lemma 51.3.1. In lower and upper Δ/Γ -filtrations of M, for a marginal sentence p in set difference Γ \ Δ, we have, respectively 1. MΔ/Γ , [w]∼Δ |= p ⇒ M, w |= p, 2. M, w |= p ⇒ MΔ/Γ , [w]∼Δ |= p.

51.4 Example of Granular Reasoning Let us consider the following reasoning process: (p1) (p2) (c)

Socrates is Human. Human is Mortal. Socrates is Mortal.

First, we formulate a model. Given P = {Human, Mortal, · · · · · · }, consider a model M =, where W = {Socrates, Plato, Tweety, Zeus, · · · }, R is an arbitrary relation on W , and V is defined by V Human Mortal · · · · · · Socrates 1 1 ······ 1 1 ······ Plato 0 1 ······ Tweety 0 0 ······ Zeus ······ ······ ······ Let Γ = {Human, Mortal} and Δ = {Human}. At the first step, Premise (p1) can be translated into M, Socrates |= Human

(51.1)

in the usual way in rough set theory. At the second step, in order to translate Premise (p2), we need a lower Human/Γ -filtration. That is, if we define w∼Human w by M, w |= Human iff M, w |= Human, then we have W/∼Human = {HumanM , (HumanM )C }. Then, we can translate Premise (p2) into MHuman/Γ , HumanM |= Mortal. At the third step, by Formula (51.1), we have HumanM = [Socrates]∼Human so, by Formula (51.2), we have

(51.2)

51. A Note on Filtration and Granular Reasoning

MHuman/Γ , [Socrates]∼Human |= Mortal.

389

(51.3)

At the final step, by Lemma 51.3.1 and Formula (51.3), we can conclude M, Socrates |= Mortal,

(51.4)

which is just the translation of Conclusion (c). Hence, we can represent our example of reasoning by the following four steps: (51.1) M, Socrates |= Human, (51.2) MHuman/Γ , HumanM |= Mortal (51.3) MHuman/Γ , [Socrates]∼Human |= Mortal (51.4) M, Socrates |= Mortal.

51.5 Concluding Remarks The main characteristic of human reasoning is resource-boundedness. We cannot have unlimited ability of reasoning. Thus if we keep to fix our model for reasoning, then we must run with a great number of detailed facts. Thus we must ignore anything that is unnecessary to the current step of our reasoning. So what we should explore is a way of disregarding such many irrelevant things and our proposal is to adopt a filtration-like method. In fact, from the first to the second steps in the previous section, we use relative filtration, which plays a part like ’zooming in’ in reasoning process. Then we can disregard details of our world that have no connection with the step. From the second to the third steps, on the other hand, a kind of inverse operation of filtration is used as if its effect is ’zooming out.’ Then we can restore some of the details in order to have some conclusion about our world. Hence, if reasoning mechanism contains such kind of operations like zooming in and out, then we can focus our attention into what is essentially needed for each step of reasoning process. Acknowledgments. The first author was partially supported by Grantin-Aid No.12878054 for Exploratory Research of the Japan Society for the Promotion of Science of Japan.

References 51.1 Chellas, B.F. (1980): Modal Logic: An Introduction. Cambridge Univ. Press, Cambridge 51.2 Lin, T.Y. (1998): Granular Computing on Binary Relation, I Data Mining and Neighborhood Systems, II Rough Set Representations and Belief Functions. L. Polkowski and A. Skowron (eds.), Rough Sets in Knowledge Discovery 1: Methodology and Applications, Physica-Verlag, Heidelberg, 107-121, 122-140 51.3 Pawlak, Z. (1991): Rough Sets: Theoretical Aspects of Reasoning about Data, Kluwer, Dordrecht

52. A Note on Conditional Logic and Association Rules Tetsuya Murai1 , Michinori Nakata2 , and Yoshiharu Sato1 1

2

Division of Systems and Information Engineering, Graduate School of Engineering, Hokkaido University, Kita-ku, Sapporo 060-8628, JAPAN {murahiko, ysato}@main.eng.hokudai.ac.jp Department of Environment and Information Management for Social Welfare, Josai International University, Togane, Chiba 283-855, JAPAN [email protected]

Association rules in data mining are considered from a point of view of conditional logic and rough sets. In our previous work, given an association rule in some fixed database, its corresponding Kripke model was formulated. Then, two difficulties in the formulation were pointed out: limitation of the form of association rules and limited formulation of the models themselves. To resolve the defects, Chellas’s conditional logic was introduced and thereby, the class of conditionals in conditional logic can be naturally regarded as containing the original association rules. In this paper, further, an extension of conditional logic is introduced for dealing with association rules with intermediate values of confidence based on the idea of fuzzy-measure-based graded modal logic.

52.1 Introduction The recent rapid progress of computer technology enables us to analyze a massive number of transaction data in commercial database systems. Such direction has provoked various ways of knowledge discovery from large database and, among them, a mining of the so-called association rules proposed by Agrawal et al.[52.1] has obtained the widespread recognition that it is one of the most active themes of data mining. In our previous paper[52.9], we investigated logical meaning of association rules from a point of view of Chellas’s conditional logic[52.4] and Pawlak’s rough sets[52.10]. Thereby, we obtained a relationship between Chellas’s conditional logic and association rules with full confidence. The logic shows the difference between material implication and conditional, so our previous result enables us to deal with exact inference on association rules as conditionals as well as an extension of the form of association rules. In this paper, as a next step of our previous work, we present an extension of conditional logic based on the idea of fuzzy-measure-based graded modal logic (cf. Murai et al.[52.6, 52.7, 52.8]), in which we can represent association rules with intermediate degrees of confidence.

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 3 90− 3 94 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

52. A Note on Conditional Logic and Association Rules

391

52.2 Association Rules Let I be a finite set of items. Any subset in I is called an itemset in I. An itemset can be a (possible) transaction. A database D is defined as a set of actual transactions, so D ⊆ 2I . For an itemset X(⊆ I), its degree def

of support s(X) is defined by s(X) = |{T ∈ D | X ⊆ T }|/|D|, where | · | is a size of a set. Given a set of items I and a database D, an association rule[52.1, 52.2] is an implication of the form X =⇒ Y , where X and Y are itemsets in I with X ∩ Y = ∅. An association rule r = (X =⇒ Y ) holds in D with confidence c(r) (0 ≤ c ≤ 1) iff c(r) = s(X ∪ Y )/s(X). An association rule r = (X =⇒ Y ) has a degree of support s(r) (0 ≤ s ≤ 1) in D iff s(r) = s(X ∪ Y ). Mining of association rules is actually performed by generating all rules that have certain minimum support (denoted minsup) and minimum confidence (denoted minconf) that a user specifies. Consult, e.g., [52.1, 52.2, 52.3] for details of such algorithms for finding association rules.

52.3 Previous Works We describe association rules in a Kripke model. Given a set of items I and a database D, we construct a modal language LML (I) in the usual way[52.4], where we regard any item as an atomic sentence. Definition 52.3.1 ([52.9]). For a given association rule r = (x =⇒ y), its corresponding finite Kripke model Mr is defined as a structure , where (1) WD = D, (2) for any T, T  in WD , T RX T  iff T ∩X = T  ∩X, so RX is an equivalence relation on W , and (3) for any item x in I, Vr (x, T ) = true iff x ∈ T . Because RX is an equivalence relation, the modal operators defined in this model Mr satisfy axiom schemata of KT 5(= S5). Note that the model described here depends on the premise x of a given association rule. Definition 52.3.2 ([52.9]). For an association rule r = (X =⇒ Y ), let X = {x1 , · · · , xm } and Y = {y1 , · · · , yn }. Then, two sentences pX and pY are defined by ( def pX = x1 ∧ · · · ∧ xm = xi ∈X xi , ( def pY = y1 ∧ · · · ∧ yn = yi ∈Y yi . Then, we have the next theorem: Theorem 52.3.1 ([52.9]). For an association rule r = (X =⇒ Y ) and its corresponding model Mr , c(X =⇒ Y ) = 1 iff Mr |= pX → pY , where, in general, M |= p means that p is true at any world in M.

392

T. Murai, M. Nakata, and Y. Sato

Now we find the following two problems: 1. Limited form of association rules whose antecedent and consequent both can take only the form of conjunction. 2. Limited formulation of the models that depends on the fixed antecedent of a given association rule. To resolve these defects, in [52.9], we introduced Chellas’s conditional logic[52.4]. Given a set of items I and a database D, we construct a language LCL (I) for conditional logic[52.4], where the difference of formation rules is if p, p ∈ LCL (I) then (p  p ) ∈ LCL (I), where  expresses ’conditional.’ Definition 52.3.3 ([52.9]). For a given database D, its corresponding finite conditional model MD is defined as a structure , where (1) def

WD = D, (2) for any world (transaction) T in WD and any set of itemsets def

X , fD (T, X ) = X , and (3) for any item x in I, VD (x, T ) = true iff x ∈ T . Then we have the following theorem: Theorem 52.3.2 ([52.9]). Given a database D and its corresponding conditional model MD , for arbitrary association rule r = (X =⇒ Y ), c(r) = 1 iff MD |= pX  pY , 0 < c(r) < 1 iff MD |= ¬((pX  pY ) ∨ (pX  ¬pY )), c(r) = 0 iff MD |= pX  ¬pY . The theorem provides us the rigid correspondence between association rules with full and no confidence and a subclass of conditionals in conditional logic. Thus, in the framework of conditional logic, we can regard the set of conditionals as an extension of association rules.

52.4 Graded Conditional Logic In this section, we introduce a graded minimal conditional logic in order to make direct treatment of intermediate degrees of confidence of association rules. For minimal conditional models, see Chellas[52.4]. Given a set of items I, a language LgCL (I) for graded conditional logic is formed from I in the usual way, where the difference is if p, p ∈ LgCL (I) and 0 < k ≤ 1 then (p k p ) ∈ LgCL (I), where k is graded conditional for 0 < k ≤ 1. Let us formulate a finite graded conditional model MgD for a given database D.

52. A Note on Conditional Logic and Association Rules

393

Definition 52.4.1. A finite graded minimal conditional model MgD is a structure 

  0 .0 6

%$&7(5,$ 1 3 1 . 0 0

  0 .0 6

%$&7(5,$ 1 4 1 . 0 0

  0 .0 5

%$&7(5,$ 1 5 1 . 0 0

  0 .0 5

%$&7(5,$ 1 6 1 . 0 0

  0 .0 4

%$&7(5,$ 1 7 1 . 0 0

  0 .0 4

%$&7(5,$ 1 8 1 . 0 0

  0 .0 4

%$&7(5,$ 1 9 1 . 0 0

  0 .0 4

%$&7(5,$ 2 0 1 . 0 0

  0 .0 4

7 ( 6 1 2

7 %

7 ( / , 8 6 $$ ) * ' * ,) 1 & 7 5 ( & * &2 6 ( 6 2) $ . / /

3 5 &

5 6 (

>



>  > 

' ,1 ) B 7 &

 

> 

$&87( $&87(

>



$&87( $&87(

>





QRUPDO

>

>

>



2 3

   0 . 1 5

9,586

2 4

   0 . 1 5

9,586

2 5

   0 . 1 5



9,586

2 6

   0 . 1 4



9,586

2 7

   0 . 1 3

9,586

2 8

   0 . 1 1

9,586

2 9

   0 . 1 1

 

9,586

3 0

   0 . 1 0



9,586

3 1

   0 . 1 0

9,586

3 2

   0 . 2 4

> 

( 6 5 8 2 &

>  > 

>   )

>



 VWUHSWR

> 

> 

9,586

( 6 5 8 2 & B &

>



0

   0 . 1 6

 / / ( & B ) 6 &

> 



2 2

( 5 8 7 / 8 &

>



$&87(

9,586

>  7 

QHJD WLYH

QRUPDO > 



QRUPDO

$&87(  

> 

>  > 

> 

QRUPDO

> 

QHJD WLYH )

 >  > 



QHJD WLYH

>

  $&87(

9,586( 3 3    0 . 0 7

)

9,586( 3 4    0 . 0 7

)

9,586( 3 5    0 . 0 6

)

9,586( 3 6    0 . 0 6

)

9,586( 3 7    0 . 0 6

)

>  

  > 

 

>  



  S 

9,586( 3 8    0 . 0 5



9,586( 3 9    0 . 0 5



>  >

> 

> 

> 

9,586( 4 1    0 . 0 5

>  7 > 



9,586( 4 2    0 . 0 5

$&87(



9,586( 4 4    0 . 0 4

>  > 

 >  > 

9,586( 4 3    0 . 0 4

9,586( 4 5    0 . 0 4

' ,1 ) B 7 / 8 &

8 / * B ) 6 &

Q >

>



> >



>

> >



>

> 





> 

2 5 3 B ) 6 &





> 

>  > 

R Q R 0 OOB H &

>



DE QRUPDO

> 

O\ R 3 B OO H &

> 

> 

> 

/ / ( & B ) 6 &

> 

DE QRUPDO DE QRUPDO



>

6 8 & 2 ) B * ( (

( 9 $ : B * ( (

DE QRUPDO

>





%$&7(5,$ 2 1    0 . 0 3

9,586( 4 0    0 . 0 5

& % :

$&87(

%$&7(5,$ 1 1 1 . 0 0   0 . 0 9 %$&7(5,$ 1 2 1 . 0 0

( & 5 8 2 = / ,( 6



) >  >  

QRUPDO

S

72. Basket Analysis on Meningitis Data Takayuki Ikeda, Takashi Washio, and Hiroshi Motoda Institute of Scientific and Industrial Research, Osaka University, 8-1, Mihogaoka, Ibarakishi, Osaka, 567-0047, JAPAN

Basket Analysis is the most representative approach in recent study of data mining. However, it cannot be directly applied to the data including numeric attributes. In this paper, we propose an algorithm and performance measures for the selection and the discretization of numeric attributes in the data preprocessing stage for the wider application of Basket Analysis, and the performance is evaluated through the application to the meningitis data.

72.1 Introduction Basket Analysis is the most representative approach in the study of data mining[72.1], and has become to be widely used in the real world applications in recent years. Based on this background, we decided to apply Basket Analysis to the meningitis data given in this discovery challenge[72.2]. However, Basket Analysis has a drawback that it cannot handle data involving numeric information such as the meningitis data, because it is to mine the associations among discrete events in principle. Thus, the task to select numeric attributes having associations with other attributes and to discretize the values of the selected numeric attributes in the data must be introduced to the mining process. An approach is to embed the task into the mining algorithm. This approach is taken in the decision tree based mining such as C4.5[72.3]. The mining algorithm directly accepts the numeric data, selects attributes relevant to the class, and discretizes the values of the selected numeric attributes while developing the decision tree. However, this approach is not suitable for Basket Analysis since its algorithm does not include any process of intermediate estimation of value distributions for massive data. Accordingly, we followed another approach, which applies the selection and the discretization in the data preprocessing stage. One important issue in its development is that the selection of the numeric attributes must be performed while taking into account the dependency among the attributes. This is because the association is a representation of the strong dependency among the events characterized by the attributes. The second issue is that the points of the discretizations in the value ranges of the numeric attributes must be chosen to appropriately reflect the dependency of the data distribution among multiple attributes. The third issue is that the discretization must have appropriate granularity. If the granularity is too small, the excessive fragmentation of the dependent region reduces the number of data representing the association of the values in each fragmented region. T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 516− 524 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

72. Basket Analysis on Meningitis Data

517

The forth issue is to establish an efficient algorithm for the selection and the discretization under massive data, though this issue is not crucial for the meningitis data since the number of the data is very limited. The objectives of this paper are as follows. (1) Development of an approach for the selection and the discretization of numeric attributes addressing the aforementioned four issues. (2) Application of the approach and Basket Analysis to the meningitis data, the evaluation of their performance and the discussion on the discovered knowledge.

72.2 Method for Selection and Discretization 72.2.1 Algorithm First, the algorithm to select and discretize the numeric attributes we developed is described[72.4]. The entire flow chart of the algorithm is depicted in Fig.72.1. Given a performance measure, this algorithm takes the greedy strategy to conduct the selection and discretization for large database in an efficient manner, and thus does not ensure to achieve the optimum selection and discretization. The detail of the performance measure will be described in the later subsection. Initially the minimum value in the value range of data for a numeric attribute is set to be a candidate threshold value. Applying this threshold for the discretization, its performance is evaluated, and it is compared with the performance of the former candidate threshold if it exists. When the performance of the newest candidate threshold is better, the threshold and the performance are recorded. Increasing the threshold value in some small amount, this search process is repeated until all candidate thresholds for every attribute have been evaluated. Once this repetition is finished, the attribute and its threshold value having the optimum performance is selected and used to discretize the data at the threshold of the attribute. After determining the threshold value of a numeric attribute, the search of another attribute and its threshold is repeated until the number of the threshold becomes to a given upper limit. The process of the selection and the discretization is applied only to the numeric attributes in the data, and the discretized attribute is merged with the original categorical attribute. As easily seen by the loop structure of the algorithm, this algorithm needs only the computational time in the order of O(N D) where N and D are the number of data and the number of numeric attributes. Because of the linear order of the computational time in terms of the data size, this algorithm can process a large amount of data efficiently, and hence the issue of the efficiency described in the first section is addressed by this algorithm.

518

T. Ikeda, T. Washio, and H. Motoda

:KLOHWKHQXPEHURI GLVFUHWL]DWLRQ LVZLWKLQDOLPLW )RUHYHU\ QXPHULFDODWWULEXWH *HQHUDWHDFDQGLGDWHWKUHVKROG IRUGLVFUHWL]DWLRQ ,IWKHSHUIRUPDQFHHYDOXDWLRQRI WKHGLVFUHWL]DWLRQVKRZDQRSWLPDO

1


AICEq.(72.1)

(72.4)

524

T. Ikeda, T. Washio, and H. Motoda

This order matches with the order of the variety of itemset combinations in the association rules. The set of rules among various attributes and their thresholds suggests many potential mechanisms underlying the data. In conclusion, the performance measure for the discretization of the numeric attributes’ values strongly affects the results of Basket Analysis. The performance measure which selects variety of the attributes and many threshold values catches interesting relations among events for domain experts. This insight should be validated through the extensive analysis in the future.

References 72.1 Agrawal, R., Srikant, R.: Fast algorithms for mining association rules. In Proc. of the 20th VLDB Conference (1994) 487–499 72.2 http://wwwada.ar.sanken.osaka-u.ac.jp/pub/washio/jkdd/jkddcfp.html 72.3 Quinlan, R.: C4.5:Programs for Machine Learning. Morgan Kaufman (1993) 72.4 Tsukada, M., Inokuchi, A., Washio, T., Motoda, H.: Comparison of MDLP and AIC for the discretization of numeric attributes. Working notes of knowledge based systems workshop, JSAI, SIG-FAI-9802 (in Japanese) (1998) 45–52 72.5 Akaike, H.: A new look at the Bayes procedure. In Biometrika, 65 (1978) 53–59 72.6 Fayyad U.M., Irani, K.B.: Multi-Interval Discretization of Continuous-Valued Nbutes for Classification Learning. In Proc. of IJCAI-93: 13th Int. Joint Conf. on Artificial Intelligence 2 (1993) 1022–1027

73. Extended Genetic Programming Using Apriori Algorithm for Rule Discovery Ayahiko Niimi and Eiichiro Tazaki Department of Control and Systems Engineering, Toin University of Yokohama 1614 Kurogane-cho, Aoba-ku, Yokohama 225-8502, JAPAN [email protected]

Genetic programming (GP) usually has a wide search space and can use tree structure as its chromosome expression. So, GP may search for global optimum solution. But, in general, GP’s learning speed is not so fast. Apriori algorithm is one of algorithms for generation of association rules. It can be applied to large database. But, It is difficult to define its parameters without experience. We propose a rule discovery technique from a database using GP combined with association rule algorithm. It takes rules generated by the association rule algorithm as initial individual of GP. The learning speed of GP is improved by the combined algorithm. To verify the effectiveness of the proposed method, we apply it to the meningoencephalitis diagnosis activity data in a hospital. We got domain expert’s comments on our results. We discuss the result of proposed method with prior ones.

73.1 Introduction Various techniques have been proposed for rule discovery using classification learning. In general, the learning speed of a system using genetic programming (GP) [73.1] is slow. However, a learning system which can acquire structural knowledge by adjusting to the environment can be constructed, because GP’s chromosome expression is tree structure, and the structure is evaluated by fitness value for the environment. On the other hand, there is the Apriori algorithm [73.2], a rule generating technique for large databases. This is an algorithm for generation of association rules. The Apriori algorithm uses two indices for rule construction: a support value and a confidence value. Depending on the setting of each index threshold, the search space can be reduced. However, it is possible that an unexpected rule cannot be extracted by reducing the range of the search space. Moreover, the load of the expert who analyzes the rule increases when there are a lot of association rule candidates, and it is a possible that it becomes difficult to search for a useful rule. Some experience is necessary to set an effective threshold. Both techniques have advantages and disadvantages as above. In this paper, we propose an extended genetic programming using apriori algorithm for rule discovery. By using the combined rule generation learning method, T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 525− 53 2, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

526

A. Niimi and E. Tazaki

it is expected to construct a system which can search for high accurate rules in large databases. The purpose of this research is achieving high forecast accuracy by small number of rules.

73.2 Genetic Programming Genetic programming (GP) is a learning method based on the natural theory of evolution, and the flow of the algorithm is similar to genetic algorithm (GA). The difference between GP and GA is that GP has extended its chromosome to allow structural expression using function nodes and terminal nodes. [73.1] In this paper, the tree structure is used to express the decision tree. The decision tree construction by GP follows the following procedures. 1. An initial population is generated from a random grammar of the function nodes and the terminal nodes defined for each problem domain. 2. The fitness value, which relates to the problem solving ability, for each individual of the GP population is calculated. 3. The next generation is generated by genetic operations. a) The individual is copied according to the fitness value (reproduction). b) A new individual is generated by intersection (crossover). c) A new individual is generated by random change (mutation). 4. If the termination condition is met, then the process ends. Otherwise, the process repeats from the calculation of fitness value in step 2. Generally, there is no method of adequately controlling the growth of the tree, because GP does not evaluate the size of the tree. Therefore, during the search process the tree may become overly deep and complex, or may settle to a too simple tree. The technique by which GP defines an effective partial tree is proposed. The approache is automatic function definition (or Automatically Defined Function: ADF), and this is achieved by adding the gene expression for the function definition to normal GP [73.4]. By implementing ADF, a more compact program can be produced, and the number of generation cycles can be reduced. More than one gene expression of ADF can be defined in one individual. One example of our GP expression is shown following.(See Figure 73.1) In Figure 73.1, decision tree is expressed in the form similar to LISP-code. GP-TREE expresses one individual of GP, and GP-TREE is composed of the ADF definition part and the main tree part. “RPB” defines main GP tree. Both “ADF0” and “ADF1” defined as each ADF tree. “IFLTE”, “IFEQ” are function nodes. These functions requires four arguments(in following example, we use arg1, arg2, arg3, arg4). The definitions of them are following. (IFLTE arg1, arg2, arg3, arg4) if arg1 is less than or equal to (≤) arg2 then evaluate arg3, else then evaluate arg4

73. Extended Genetic Programming Using Apriori Algorithm (:G P -T R E E (:A D F 0 (IF E Q D (:A D F 1 (C )) (:R P B (IF L T E A (IF E Q (IF

527

A = "T " B = "T " "F " P N )) C = "F ": N C = "T ": P "T " B "T " E Q C "F " N P ) A D F 0 ) P )))

B = "F " D = "F ": P D = "T ": N A = "F ": P

Fig. 73.1. Expression of GP’s Chromosome (The left side is an individual expression of LISP-code and the right side is rewritten to the decision tree expression.)

(IFEQ arg1, arg2, arg3, arg4) if arg1 is equal to(=) arg2 then evaluate arg3, else then evaluate arg4. A, B, C and D express the attributes in database. “T” and “F” express attribute value, and “N” and “P” express class name.

73.3 Approach of Proposed Combined Learning To make up for the advantage and the disadvantages of the Apriori algorithm and GP, we propose a rule discovery technique which combines GP with the Apriori algorithm. By combining each technique, the search of high accurate rules from a large database is expected. An outline of our proposed technique is shown in Figure 73.2.

A p r io r i A lg o r ith m A s s o c ia tio n R u le s C o n v e r s io n ( r u le s - > tr e e s )

G e n e tic P r o g r a m m in g D e c is io n T r e e C o n v e r s io n ( tr e e s - > r u le s ) C la s s ific a tio n R u le s

Fig. 73.2. Flow Chart of Approach of Proposed Combined Learning

528

A. Niimi and E. Tazaki

The following steps are proposed for the rule discovery technique. 1. First, the Apriori algorithm generates the association rule. 2. Next, the generated association rules are converted into decision trees which are taken in as initial individuals of GP. The decision trees are trained by GP learning. 3. The final decision tree is converted into classification rules. This allows effective schema to be contained in the initial individuals of GP. As a result, it is expected to improve the GP’s learning speed and its classification accuracy. However, when GP is used for multi-value classification, the learning speed of GP may become slow due to increasing the number of definition nodes. Therefore, it is difficult to apply the proposed technique to multi-value classification. For conversion from the association rule into decision trees, we use the following procedures. 1. For the first process, the route of the decision tree is constructed, assuming the conditions of the association rule as the attribute-based tests of the decision tree. 2. In the next process, the conclusions of the association rule is appended on the terminal node of this route. 3. Finally, the class value of the terminal nodes which are not defined by the association rule are assigned by randamly choosing from the terminal nodes set. In conversion from the association rule to the decision tree, a rule which contains class attribute in the conclusion part is selected. One decision tree is converted based on one association rule. A too simple decision tree is generated by conversion, but the decision tree of high accuracy is not necessary to GP’s initial individuals, because of GP learning. The conversion does not make the amount of the calculation increase because it is simple conversion. For conversion from the GP’s decision tree to the classification rule, we use the process proposed by Quinlan [73.5].

73.4 Apply to Rule Discovery from Database We applied the proposed technique for the meningoencephalitis diagnosis data sets. This database was donated by S.Tsumoto[73.6]. We applied the proposed technique for “find factors important for diagnosis (DIAG2) to judge bacteria or virus”. We obtained following results of decision tree and rules generated by ADF-GP. In the proposed method, we took the association rule generated by Apriori algorithm as initial individuals of GP. We used 70 data for training, 140 data for test. 70 data was extracted at random. We studied these data by using the normal GP, and tuned of the GP parameter before experiment.

73. Extended Genetic Programming Using Apriori Algorithm

529

We defined some expressions. “A eq B” is express that attribute(A) is equal to attribute(B) if its attribute is discrete value. “A && B” represents to connect each part(A,B) of rules by ‘’and”. The left side of “→” express conditions of rule, and the right side of “→” express conclusion of rule (or class name). The section 73.4.1 shows the results using ADF-GP only. The section 73.4.2 shows the results using proposed technique. 73.4.1 ADF-GP Only The following rules are generated with ADF-GP. The generated rules are composed by the categorical attributes. === generated rules rule1: (EEG_FOCUS eq "-") && (RISK eq "n") -> rule2: (EEG_FOCUS eq "-") && (RISK eq "p") -> rule3: (EEG_FOCUS eq "-") && (RISK eq "n") -> rule4: (EEG_FOCUS eq "-") && (RISK eq "p") -> rule5: (EEG_FOCUS eq "+") && (RISK eq "n") -> rule6: (EEG_FOCUS eq "+") && (RISK eq "p") -> rule7: (EEG_FOCUS eq "+") -> BACTERIA rule8: (EEG_FOCUS eq "-") rule9: (EEG_FOCUS eq "+")

=== && (CT_FIND eq "normal") && (SEX eq "M") VIRUS && (CT_FIND eq "normal") && (SEX eq "M") BACTERIA && (CT_FIND eq "normal") && (SEX eq "F") VIRUS && (CT_FIND eq "normal") && (SEX eq "F") BACTERIA && (CT_FIND eq "abnormal") && (SEX eq "F") VIRUS && (CT_FIND eq "abnormal") && (SEX eq "F") BACTERIA && (CT_FIND eq "abnormal") && (SEX eq "M") && (CT_FIND eq "abnormal") -> BACTERIA && (CT_FIND eq "normal") -> VIRUS

To examine the availability and the accuracy of the generated rule, the size of the rule, the use frequency and the wrong classification frequency (wrong classification rate) to all data, the classification class by rules are shown in Table(73.1). In the table, the rule 6 is not used for all data. The rules (1 and 3) with high availability show low wrong classification rates. Other rules have high wrong classification rate independent of availability. To examine the classification accuracy of the generated rule set, each classification distribution to all data are shown in Table(73.2). The table shows that small number of data could not classify VIRUS and BACTERIA correctly. 73.4.2 Proposed Technique (Association Rules + ADF-GP) The following rules are generated with proposed technique. The generated rules are composed by the continuous value attributes.

530

A. Niimi and E. Tazaki

Table 73.1. Evaluation on test data by each rules (ADF-GP only) Rule 1 2 3 4 5 6 7 8 9

Size 4 4 4 4 4 4 3 2 2

Used 33 6 36 2 7 0 5 27 24

Wrong 4 2 3 0 0 0 1 8 6

( 12.12 %) ( 33.33 %) ( 8.33 %) ( 0.00 %) ( 0.00 %) ( 0.00 %) ( 20.00 %) ( 29.63 %) ( 25.00 %)

: : : : : : : : :

VIRUS. BACTERIA. VIRUS. BACTERIA. VIRUS. BACTERIA. BACTERIA. BACTERIA. VIRUS.

Table 73.2. Evaluation on test data by error distribution(ADF-GP only) (a) (b) ← classified as 87 11 (a):class VIRUS 13 29 (b):class BACTERIA total hits= 116

=== generated rules === rule1: (Cell_Poly VIRUS rule2: (Cell_Poly > 221) && (EEG_FOCUS rule3: (Cell_Poly > 221) && (EEG_FOCUS -> BECTERIA rule4: (Cell_Poly > 221) && (EEG_FOCUS && (SEIZURE == 0 ) -> VIRUS rule5: (Cell_Poly > 221) && (EEG_FOCUS && (SEIZURE != 0 ) -> BACTERIA

BECTERIA > 200) && (GCS 200) && (GCS > 121) > 200) && (GCS > 121)

The performance of the generated rule are shown in Table(73.3). In the table, the rule 3, 4 and 5 are not used for all data. The rule 1 and 2 have high availability and low wrong classification rates. Table 73.3. Evaluation on test data by each rules (proposed method) Rule 1 2 3 4 5

Size 1 2 3 4 4

Used 108 32 0 0 0

Wrong 10 0 0 0 0

( ( ( ( (

9.26 0.00 0.00 0.00 0.00

%) %) %) %) %)

: : : : :

VIRUS. BACTERIA. BACTERIA. VIRUS. BACTERIA.

To examine the classification accuracy of the generated rule set, each classification distribution to all data is shown in Table(73.4). The table shows that some rules classified BACTERIA as VIRUS by mistake, but almost rules have correct classification ability.

73. Extended Genetic Programming Using Apriori Algorithm

531

Table 73.4. Evaluation on test data by error distribution (proposed method) (a) (b) ← classified as 98 0 (a):class VIRUS 10 32 (b):class BACTERIA total hits= 130

73.4.3 Discussion for the Results In the results, the proposed method shows higher accuracy than ADF-GP, and dataset can be expressed using more small number of rules. The proposed method does not have pruning rules operation except for GP operations. GP operation is a kind of statistical operation. Thus, sometimes GP operation can obtain interesting rules, but otherwise, the result contains meaningless rules. For such problems, GP technique which contain the pruning operation are proposed [73.7], and it makes possible to build the pruning techniques in our proposed technique. Moreover, it is also possible in the experiment to remove meaningless rules by using the threshold in availability. When the experimental result is evaluated and cleaned by domain expert after experiment, the load for domain expert depends on the number of rules of results. In the proposed technique, the number of rules of results can be reduced compared with only ADF-GP. We got following comments on these results from domain expert(S. Tsumoto). Totally, the results obtained by ADF-GP are more interesting than the proposed methods. The results obtained by the proposed technique are very reasonable, but I do not see the meaning of “EEG FOCUS > 200” and “GCS > 121”. Please let me know what the authors mean by that. Please show me the results for other problems. The purpose of this research is to achieve high forecast accuracy by small number of rules. This purpose is not as same as expert’s interest on the experiment result. Because expert’s interesting rules were obtained by the normal ADF-GP, expert’s interesting rule can be obtained by the proposed technique by increasing the GP effect.

73.5 Conclusions In this paper, we proposed the rule discovery technique from the database using genetic programming combined with association rule algorithms. To verify the validity of the proposed method, we applied it to the meningoencephalitis diagnosis activity data in a hospital, and discussed the results of proposed method and normal ADF-GP with domain expert. As a result, an improvement of rules’ accuracy was seen, and proposed method can express

532

A. Niimi and E. Tazaki

dataset by the small number of rules. It can be concluded that the proposed method is an effective method to the improvement of the rules’ accuracy and can save the number of rules for the rule discovery problem. Though the comments of domain expert, using only ADF-GP method can be obtained more interesting rules than using proposed method. In the future, we will research the following 4 topics. The first topic is to apply the method to other verifications. We already applied proposed method for other problems [73.8] [73.9]. We need to discuss the problem suitable for proposed method through the applications to various problems. The second topic is to discuss the conversion algorithm from the association rule to a decision tree with high accuracy. The third topic is to extend the proposed method to multi-value classification problems. It is necessary for this problem to suppress increasing the number of definition nodes and to establish measures against the decrease at the learning speed by increasing nodes. The fourth topic is to obtain more interesting rules such as ADF-GP only.

References 73.1 Koza, J. R. (1992): Genetic Programming. MIT Press. 73.2 Terabe, M., Katai, O., Sawaragi, T., Washio, T., Motoda, H. (2000): Attribute Generation Based on Association Rules. Journal of Japanese Society for Artificial Intelligence, Vol.15 No.1,pp.187–197 (Japanese). 73.3 Kitsuregawa, M. (1997): Mining Algorithms for Association Rules. Journal of Japanese Society for Artificial Intelligence, Vol.12 No.4,pp.513–520 (Japanese). 73.4 Koza, J. R., Kinner, K. E.(ed.), et.al (1994): Scalable Learning in Genetic Programming Using Automatic Function Definition. Advances in Genetic Programming, pp.99–117. 73.5 Quinlan, J. R. (1993): C4.5: Programs for Machine Learning. Morgan Kaufman Publishers. 73.6 Tsumoto, S. (2000): Meningoencephalitis Diagnosis data description. [http://www.ar.sanken.osaka-u.ac.jp/jkdd01/menin.htm]. 73.7 Niimi, A., Tazaki, E. (1999): Extended Genetic Programming using Reinforcement Learning Operation. Proceedings of IEEE International Conference on Systems, Man, and Cybernetics, pp.596–600. 73.8 Niimi, A., Tazaki, E. (2000): Genetic Programming Combined with Association Rule Algorithm for Decision Tree Construction. Proceedings of the Fourth International Conference on Knowledge-Based Intelligent Engineering System & Allied Technologies, volume 2,pp.746–749. 73.9 Niimi, A., Tazaki, E. (2000): Rule Discovery Technique Using Genetic Programming Combined with Apriori Algorithm. Proceedings of the Third International Conference on Discovery Science, pp.273–277. 73.10 Niimi, A., Tazaki, E. (2000): Knowledge Discovery using Extended Genetic Programming from Biochemical Data. Proceedings of 49th KBS meeting, pp.45–49, Japan AI Society, SIG-KBS-A002 (Japanese).

7 4 . M e d ic a l K n o w le d g e D is c o v e r y o n th e M e n in g o e n c e p h a litis D ia g n o s is S tu d ie d b y th e C a sc a d e M o d e l Ta k a s h i Ok a d a Cen ter for In for m a ti on & Med i a Stu d i es , Kwa n s ei G a k u i n Un i ver s i ty 1-1-155 Ueg a h a r a , N i s h i n om i y a , Ja p a n

7 4 .1 I n t r o d u c t io n Th e i m p or ta n ce of ex tr a cti n g k n owl ed g e fr om d a ta b a s es i s wel l es ta b l i s h ed i n th e d om a i n of m ed i ca l s ci en ce. R ecen t a d va n ces i n h os p i ta l a u tom a ti on h a ve i n tr od u ced d a ta b a s es th a t s tor e va s t a m ou n ts of i n for m a ti on on p a ti en ts ’ ca s e h i s tor i es . Si n ce ex p er i m en ts i n vol vi n g h u m a n p a ti en ts a r e n ot a l wa y s p os s i b l e, effi ci en t a n d fl ex i b l e d a ta m i n i n g i s ex p ected to fa ci l i ta te n ew m ed i ca l d i s cover i es fr om a va i l a b l e d a ta . Th e ob j ecti ve of th i s p a p er wa s to a n a l y z e a m en i n g oen cep h a l i ti s d a ta s et, th e tes t d a ta a t th e JSAI KD D Ch a l l en g e 2001 wor k s h op [ 1] , a n d to h el p or g a n i z e th e s ea r ch for n ew i n for m a ti on on th i s d i s ea s e. Th e m eth od of a n a l y s i s i s th e ca s ca d e m od el d evel op ed b y th e a u th or . Secti on 2 b r i efl y i n tr od u ces th e m od el . Th e com p u ta ti on p r oced u r e for th e ch a l l en g e p r ob l em a n d th e r es u l ti n g r u l es a r e s h own i n Secti on 3 . We a l s o i n d i ca te th e u s efu l n es s of vi s u a l i n s p ecti on of d a ta g u i d ed b y th e ob ta i n ed r u l es . Th e l a s t s ecti on d i s cu s s es p os s i b l e i m p r ovem en ts i n d a ta m i n i n g u s i n g th e ca s ca d e m od el .

7 4 .2 T h e C a s c a d e M o d e l Th e m od el ex a m i n es a n i tem s et l a tti ce wh er e a n [ a ttr i b u te: va l u e] p a i r i s em p l oy ed a s a n i tem to con s ti tu te i tem s ets . L i n k s i n th e l a tti ce a r e s el ected a n d ex p r es s ed a s r u l es [ 2] . F i g u r e 1.1 s h ows a ty p i ca l ex a m p l e of a l i n k a n d i ts ex p r es s ed r u l e. Her e, th e p r ob l em con ta i n s fi ve a ttr i b u tes : A– E , ea ch of wh i ch ta k es (y , n ) va l u es . Th e i tem s et a t th e u p p er en d of th e l i n k h a s i tem [ A: y ] , a n d a n oth er i tem [ B : y ] i s a d d ed a l on g th e l i n k . Item s of th e oth er a ttr i b u tes a r e ca l l ed vei l ed i tem s . Th e ta b l es a tta ch ed to th e n od es s h ow th e fr eq u en ci es of vei l ed i tem s .

T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 53 3 -54 0, 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

53 4

T. Ok a d a

F i g . 1 . 1 . A s a m p l e l i n k , i ts r u l e ex p r es s i on a n d p r op er ti es of th e vei l ed i tem s .

In or d er to eva l u a te th e s tr en g th of a r u l e, th e wi th i n -g r ou p s u m of s q u a r es (W S S ) a n d b etween -g r ou p s u m of s q u a r es (B S S ) a r e d efi n ed b y th e fol l owi n g for m u la e[ 3 , 4 ] ,

W S S B S S i

i

=

n ⎛ ⎜ 1− 2⎝ n L = ∑ 2

∑

p i

a

(p (a ) − L

i

a

( a ) 2 ⎞⎟

(1)

, ⎠ p

U i

(a )

)

2

.

(2)

wh er e i d es i g n a tes a n a ttr i b u te; th e s u p er s cr i p ts U a n d L i n d i ca te th e u p p er a n d l ower n od es , r es p ecti vel y ; n s h ows th e n u m b er of s u p p or ti n g ca s es of a n od e; a n d p i (a ) i s th e p r ob a b i l i ty of ob ta i n i n g th e va l u e a for a ttr i b u te i . F i g u r e 1.1 s h ows th e W S S i a n d B S S i va l u es a l on g wi th th ei r s a m p l e va r i a n ces . A l a r g e B S S i va l u e i s evi d en ce of a s tr on g i n ter a cti on b etween th e a d d ed i tem a n d a ttr i b u te i . Th e tex tb ox a t th e r i g h t i n F i g u r e 1.1 s h ows th e d er i ved r u l e. Th e a d d ed i tem [ B : y ] a p p ea r s a s th e m a i n con d i ti on i n th e L HS, wh i l e th e i tem s i n th e u p p er n od e a r e p l a ced a t th e en d of th e L HS a s p r econ d i ti on s . Wh en a vei l ed a ttr i b u te h a s a l a r g e B S S i va l u e, on e of i ts i tem s i s p l a ced i n th e R HS of a r u l e. Th e m eth od of s el ecti n g i tem s fr om a vei l ed a ttr i b u te wa s d es cr i b ed i n [ 3 ] . We ca n con tr ol th e a p p ea r a n ce of a ttr i b u tes i n th e L HS b y r es tr i cti n g th e a ttr i b u tes i n th e i tem s et n od e. On th e oth er h a n d , th e a ttr i b u tes i n th e R HS ca n b e s e-

74 . Med i ca l Kn owl ed g e D i s cover y on th e Men i n g oen cep h a l i ti s D i a g n os i s

l ected b y s etti n g th e m i n i m u m B S S i va l u e of a r u i s n ot n eces s a r y for i tem s i n th e R HS of a r u l e s h a r p con tr a s t to a s s oci a ti on r u l e m i n er s , wh i ch r y ; E : n ] to d er i ve th e r u l e i n F i g u r e 1.1. Th es e ch m a k e i t p os s i b l e to d etect r u l es effi ci en tl y [ 5] .

l e (m i n -B S S i ) for ea ch a ttr i b u to r es i d e i n th e l a tti ce. Th i s eq u i r e th e i tem s et, [ A: y ; B : a r a cter i s ti cs of th e ca s ca d e m

53 5

te. It is in y ; D : od el

7 4 .3 R e s u lt s a n d D is c u s s io n 7 4 .3 .1 C o m p u t a t io n b y D I S C A S

Th e d a ta s et p r ovi d ed for th e JSAI KD D Ch a l l en g e 2001 con s i s ts of r ecor d s on 14 0 m en i n g oen cep h a l i ti s p a ti en ts [ 1] . E a ch r ecor d con ta i n s 4 0 a ttr i b u te va l u es . Al l n u m er i ca l d a ta wer e ca teg or i z ed a s s h own i n Ta b l e 1.1. F or ex a m p l e, th e a ttr i b u te “ COL D ” wa s con ver ted i n to on e of th e th r ee i tem s : “ col d ”0” , “ 0< col d ”5” or “ col d > 5” . We a n a l y z ed th e d a ta s et u s i n g D ISCAS s oftwa r e (ver s i on 2.1), wh i ch wa s d evel op ed b y th e a u th or . F a ctor s a ffecti n g d i a g n os i s , d etecti on of b a cter i a or vi r u s , a n d p r og n os i s wer e ex a m i n ed b y ch a n g i n g th e R HS a ttr i b u te. Th e r es u l ts a r e s h own s ep a r a tel y i n th e fol l owi n g s u b s ecti on s . Al l ca l cu l a ti on s wer e d on e u s i n g a 600-MHz P en ti u m III P C wi th 256 MB of m em or y . Th e p r u n i n g con d i ti on s wer e s et to m i n s u p = 0.01 a n d t h r e s = 0.05 (s ee r efer en ce [ 5] for th e m ea n i n g of th es e p a r a m eter s ). D ISCAS g en er a ted a l a tti ce wi th i n 2.5 m i n u tes th a t con ta i n ed a b ou t 50, 000 n od es for d i a g n os i s a n d cu l tu r e d etecti on p r ob l em s , wh i l e i t took 10 m i n u tes of ca l cu l a ti on to con s tr u ct a l a tti ce wi th T a b l e 1 . 1 . Ca teg or i z a ti on of n u m er i ca l 120, 000 n od es for th e p r og n os i s p r ob l em . a ttr i b u tes We a s s u m e th a t th e s i g n i fi ca n ce of a r u l e i s r ou g h l y p r op or ti on a l to i ts B S S Attr i b u te Sp l i tti n g va l u es va l u e, s o we s h ow th e r u l es wi th l a r g e B S S AG E 20, 3 0, 4 0, 50 COL D 0, 5 va l u es . However , we h a ve to b e ca r efu l i n HE AD ACHE 0, 3 , 6, 9 th e s el ecti on of r u l es fr om com p u ta ti on r eF E VE R 0, 3 , 6, 10 s u l ts , s i n ce s ets of r u l es th a t s h a r e m a n y of N AUSE A 0, 3 th e s a m e s u p p or ti n g i n s ta n ces s h ou l d b e L OC 0, 1 con s i d er ed d i ffer en t ex p r es s i on s of th e SE IZ UR E 0 s a m e p h en om en on . L et u s th i n k of a cl a s B T 3 6, 3 7, 3 8 , 3 9 STIF F 0, 1, 2, 3 s i fi ca ti on p r ob l em for [ Cl a s s : p os ] . If we G CS 14 ob ta i n th e fol l owi n g th r ee r u l es , th ey s h ow WB C 5000, 6000, 8 000, 10000 th e ex i s ten ce of a d a ta s eg m en t s h a r i n g CR P 0, 1, 3 i tem s [ A: y ] , [ B : n ] , [ C: y ] , a n d [ Cl a s s : E SR 0, 10, 25 p os ] . We b el i eve th a t th e s tr on g es t r u l e CSF _ CE L L 50, 125, 3 00, 750 s h ou l d b e s el ected fr om eq u i va l en t ex p r es Cel l _ P ol y 8 , 20, 50, 3 00 s i on s u s i n g th e B S S cr i ter i on , a l th ou g h Cel l _ Mon o 50, 125, 3 00, 750 CSF _ P R O 0, 60, 100, 200 oth er ex p r es s i on s a r e often u s efu l a s r eCSF _ G L U 4 0, 50, 60, 70 l a ted k n owl ed g e.

53 6

T. Ok a d a

IF [ A: a d d THE N THE N

y ] ed on [ B : n ] [ Cl a s s : p os ] [ C: y ]

IF [ C: a d d THE N THE N

IF [ B : a d d THE N THE N

y ] ed on [ B : n ] [ Cl a s s : p os ] [ A: y ]

n ] ed on [ A: y ] [ Cl a s s : p os ] [ C: y ]

7 4 .3 .2 D ia g n o s is Th e d a ta s et g u i d e i n d i ca tes th a t d i ffer en ti a l d i a g n os i s i s i m p or ta n t i n d eter m i n i n g wh eth er th e d i s ea s e i s b a cter i a l or vi r a l m en i n g i ti s . We a n a l y z ed th e d a ta s et s etti n g D IAG 2 (th e g r ou p ed a ttr i b u te of D IAG ) a s th e R HS a ttr i b u te. Al l a ttr i b u tes wer e em p l oy ed a s L HS a ttr i b u tes , ex cep t D IAG a n d th e 11 a ttr i b u tes wh os e va l u es wer e ob ta i n ed a fter th e i n i ti a l d i a g n os i s . Th ey wer e CUL TUR E , CUL T_ F IN D , THE R AP Y2, C_ COUR SE , C_ COUR SE (G r ou p ed ), CSF _ CE L L 3 , CSF _ CE L L 7, C_ COUR SE , COUR SE (G r ou p ed ), R ISK, a n d R ISK(G r ou p ed ). Th e top two r u l es i n th e fi r s t r u l e s et a r e s h own b el ow. Th es e a r e th e s tr on g es t r u l es l ea d i n g to th e d i a g n os i s of b a cter i a l a n d vi r a l m en i n g i ti s , r es p ecti vel y . I F T H E N I F

[ C e l l _ P o l y > D i a g 2 = B A C T E R I A

[ 2 0 < C e l l _ P o l y = T H E N D i a g 2 = V I R U S

5 0 ]

3 0 0 ] 3 0 . 0 %

a d d e d o n - > 1 0 0 . 0 % ;

a d d e d o n [ C S F _ C E L L 3 8 . 6 % - > 1 0 0 . 0 % ;

>

[ ] 7 5 0 ]

Si n ce th e ca s ca d e m od el h a s a s ea r ch r a n g e th a t i s l i m i ted i n th e p r op os i ti on a l ca l cu l u s d om a i n , i t ca n n ot d r a w u p on a n ex p er t’ s k n owl ed g e i n com p a r i n g Cel l _ P ol y a n d Cel l _ Mon o d i r ectl y . However , we ca n ea s i l y r ea ch th e s a m e con cl u s i on i f we i n s p ect th e s ca tter g r a m i n F i g . 1.2 r efer en ci n g th e con s tr a i n t: CSF _ CE L L = Cel l _ P ol y + Cel l _ Mon o. An oth er a n a l y s i s i s p u t i n to p r a cti ce om i tti n g th e a ttr i b u tes : Cel l _ P ol y a n d Cel l _ Mon o, a n d th e r es u l ti n g r u l es a r e ex p ected to l ea d to n ew k n owl ed g e, vi ewed fr om a n oth er p oi n t. Th e r u l es for wh i ch B S S > 3 .0 a r e i l l u s tr a ted i n Ta b l e 1.2.

F i g . 1 . 2 . Sca tter p l ot: Cel l _ p ol y vs CSF _ Cel l

74 . Med i ca l Kn owl ed g e D i s cover y on th e Men i n g oen cep h a l i ti s D i a g n os i s

53 7

T a b l e 1 . 2 . Str on g r u l es ob ta i n ed for D IAG 2

N o 1

2

3 †

Ma i n con d i ti on

P r econ d i ti on s

Ch a n g e i n D i a g 2 d i s tr i b u ti on (b a cter i a vi r u s )

[ CR P > 3 ]

†

B S S

Å (15 3 ) / 18 5.98 Å (19 5) / 24 , B S S = 5.8 0. Å (19 8 ) / 27 4 .23 o ch a n g es 62% Å93 % . ) / 14 0 Å (23 16) / 3 9. (7 17) / 24 Å (5 0) / 5. Å (7 0) / 7 3 .68

[ F OCAL : ] (27 78 ) / 105 If n o p r econ d i ti on i s a p p l i ed , (4 2 98 ) / 14 0 [ CT_ F IN D : a b n or m a l ] [ E E G _ F OCUS: ] (3 2 72) / 105 Th e p er cen ta g e of [ SE X : M] a l s If n o p r econ d i ti on i s a p p l i ed , (4 2 98 If [ F E VE R : 0, E E G _ F OCUS: -] i s th e p r econ d i ti on , [ N AUSE A= 0, (14 3 7) / 51 [ CT_ F IN D : a b n or m a l ] L OC_ D AT: ]

Va l u es i n p a r en th es es s h ow th e n u m b er of b a cter i a l a n d vi r a l ca s es , wh i l e th e va l u e a fter th e s l a s h g i ves th e n u m b er of a l l i n s ta n ces for a l l a ttr i b u te va l u es . T a b l e 1 . 3 . Str on g r u l es ob ta i n ed for D IAG Ch a n g es i n D i a g d i s tr i b u ti on (Ab s ces s B a cter i a B a cte(E ) TB (E ) B S S Vi r u s (E ) Vi r u s ) (9 24 8 1 3 0 68 ) / 14 0Å (2 21 7 0 0 0) / 3 0; [ Cel l _ P ol y > [ ] 8 .8 8 3 00] In cr ea s e i n B a cter i a , d ecr ea s e i n Vi r u s 1 If [ F OCAL : @ i s th e p r econ d i ti on , (4 19 4 0 14 64 ) / 105Å (1 16 4 0 0 0) / 21. If [ E E G _ F OCUS: @i s th e p r econ d i ti on , (8 19 4 1 11 61) / 104 Å (2 16 4 0 0 0) / 22. (0 6 0 0 13 3 0) / 4 9Å (0 1 0 0 10 0) / 11; [ SE X : F , [ E E G _ F OCUS 4 .3 5 CT_ F IN D : n or m a l ] : +] In cr ea s e i n Vi r (E ), d ecr ea s e i n Vi r u s 2 IF [ SE X : F ] i s th e p r econ d i ti on , (1 6 0 0 19 3 2) / 58 Å (0 1 0 0 15 2) / 18 . IF [ CT_ F IN D : n or m a l ] i s th e p r econ d i ti on , (0 15 4 0 19 63 ) / 101 Å (0 3 3 0 13 5) / 24 . IF n o p r econ d i ti on i s a p p l i ed , (9 24 8 1 3 0 68 ) / 14 0 Å (1 5 4 0 19 7) / 3 6. (8 19 4 1 11 61) / 104 Å (4 8 3 1 8 2) / 26; [ L OC_ D AT: 3 4 .27 [ E E G _ F OCUS: @ +] D ecr ea s e i n Vi r u s (8 19 4 1 11 61) / 104 Å (8 7 3 1 5 3 ) / 27 [ CT_ F IN D : 4 3 .95 [ E E G _ F OCUS: @ a b n or m a l ] D ecr ea s e i n Vi r u s , i n cr ea s e i n Ab s ces s [ CR P ” (0 3 0 0 5 3 4 ) / 4 2Å (0 1 0 0 5 0) / 6 [ F OCAL : +] CT_ F IN D : n or m a l , 3 .52 D ecr ea s e i n Vi r u s , i n cr ea s e i n Vi r u s (E ) E E G _ F OCUS: @ 5 IF [ CR P ”((*B)2&86 @ i s th e p r econ d i ti on , (2 5 1 1 8 3 5) / 52Å (2 2 0 1 7 1) /13 . (5 0 0 0 5 18 ) / 28 Å (5 0 0 0 1 0) / 6; [ E E G _ F OCUS: , [ CSF _ P R O”0] 2.53 Cel l _ P ol y ” 8 ] In cr ea s e i n Ab s ces s , d ecr ea s e i n Vi r u s 6 Ma j or ch a n g es i n oth er a ttr i b u tes : [ F OCAL = +] 29% Å 8 3 % , [ CT_ F IN D : a b n or m a l ] 21% Å 8 3 % , [ CSF _ CE L L ” 50] 3 9% Å 100% , [ Cel l _ Mon o ” 50] 3 9% Å 100% .

N o

Ma i n con d i ti on

P r econ d i ti on s

53 8

T. Ok a d a

F i g u r e 1.3 s h ows th e ch a n g es i n th e D i a g 2 d i s tr i b u ti on s b y Sp otfi r e [ 6] . Ax es wer e s el ected fol l owi n g R u l e 1 i n Ta b l e 1.2. We ca n s ee cl ea r i n cr ea s e i n b a cter i a r a ti o a t th e top l eft p i e ch a r t, a n d th e d i s tr i b u ti on ch a n g es i n oth er ch a r ts s eem r ea s on a b l e. Ta b l e 1.3 d ep i cts th e r u l es wh en we em p l oy ed D IAG a s th e R HS a ttr i b u te. Th e r u l es for wh i ch B S S > 3 .0 a r e s h own . Wh en n o r u l e a p p ea r ed to d i s cr i m i n a te a cl a s s , th e s tr on g es t r u l e r el a ted to th e cl a s s wa s a l s o i n cl u d ed , a l th ou g h n o r u l es wer e fou n d i n d i ca ti n g B ACTE (E ). Si x cl a s s es a p p ea r ed i n th e d i s tr i b u ti on l i s t, of wh i ch th e fi r s t fou r wer e b a cter i a l a n d th e l a s t two wer e vi r a l . Th e cl a s s s h owi n g a r em a r k a b l e ch a n g e i s u n d er l i n ed . R u l es 4 a n d 6 i n Ta b l e 1.3 i n d i ca te th e ch a r a cter i s ti c s eg m en t of a n a b s ces s , r el a ted to th e CT a b n or m a l i ty m en ti on ed i n [ 1] .

P ie C h a r t o f D ia g n o s is 







QHJ

SRV

)2&$/

F i g . 1 . 3 . Vi s u a l i z a ti on of R u l e 1 i n Ta b l e 1.2 B l a ck : b a cter i a ; wh i te: vi r u s .

7 4 .3 .3 D e t e c t io n o f B a c t e r ia o r V ir u s Ta b l e 1.4 s h ows th e r u l es wi th B S S > 2.0 wh en CUL T_ F IN D wa s u s ed a s th e R HS a ttr i b u te. Th e a ttr i b u tes for th e L HS wer e th e s a m e a s th os e i n th e d i a g n os i s p r ob . l em . Al l B S S va l u es a r e r el a ti vel y l ow. In fa ct, th e d i s tr i b u ti on ch a n g e s h own i n F i g u r e 1.4 s eem s to b e u n n a tu r a l for R u l e 1 i n Ta b l e 1.4 . Al l r u l es ex cep t r u l e 6 i n Ta b l e 1.4 d es er ve n o fu r th er d i s cu s s i on a fter s u ch vi s u a l i n s p ecti on s a r e a p p l i ed . F or th e p r ob l em of s p eci fi c cu l tu r e fi n d i n g s , we cou l d d etect n o s tr on g r u l es . Th e on l y ex cep ti on wa s R u l e 8 &FHOOBPRQR i n Ta b l e 1.4 , wi th wh i ch a l l s p eci es fou n d wer e h er p es . 



















F i g . 1 . 4 . Vi s u a l i z a ti on of R u l e 1 i n Ta b l e 1.4 . B l a ck : fou n d ; wh i te: n ot fou n d

74 . Med i ca l Kn owl ed g e D i s cover y on th e Men i n g oen cep h a l i ti s D i a g n os i s

53 9

T a b l e 1 . 4 . Str on g r u l es ob ta i n ed for CUL T_ F IN D N o

Ma i n con d i ti on

1

[ B T> 3 9.0]

P r econ d i ti on s

Ch a n g e i n CUL T_ F IN D d i s tr i b u ti on (F a l s e Tr u e) (18 5) / 23 Å (0 5) / 5

B S S

[ 125< Cel l _ Mon o”3 00] 3 .06 [ SE X : M, COL D ” [ 3 8 .0< B T”9.0] (18 6) / 24 Å (0 4 ) / 4 2.25 5000< WB C”6000] IF [ 5000< WB C”6000, CR P ”@ i s th e p r econ d i ti on , (3 1 10) / 4 1 Å (2 6) / 8 , B S S = 2.05. [ F OCAL :  (14 5) / 19Å (0 4 ) / 4 2.17 [ HE AD ACHE ”@ Cel l _ Mon o”50] [ 100< CSF _ P R O”200] [ KE R N IG : 1] (25 5) / 3 0Å (0 4 ) / 4 2.78 [ L OC> 1] [ SE X : F , COL D ”0] (27 7) / 3 4 Å (1 5) / 6 2.3 6 [ AG E > 50] [ SE X : F , CR P ”0] (24 8 ) / 3 2Å (0 4 ) / 4 2.25 [ F OCAL : +] [ 60< CSF _ P R O”100] (3 3 19) / 4 3 Å (1 5) / 6 2.17 Al wa y s a ccom p a n i ed b y [ CT_ F IN D : a b n or m a l ] . (14 4 ) / 18 Å (0 3 ) / 3 [ E E G _ F OCUS: +] [ 3 6.0< B T”7.0] 1.8 1 Al l a r e h er p es .

2 3 4 5 6 7 8

7 4 .3 .4

P r o g n o s is

COUR SE (G r ou p ed ) wa s em p l oy ed a s th e R HS a ttr i b u te a n d th e r es u l ti n g r u wi th B S S > 2.0 a r e s h own i n Ta b l e 1.5. On l y CSF _ CE L L 3 a n d C_ COUR SE om i tted fr om th e s et of L HS a ttr i b u tes . E ven i f we u s e C_ COUR SE a s th e R a ttr i b u te, we ca n n ot fi n d a s p eci fi c cou r s e ex cep t R u l es 3 a n d 9 i n Ta b l e 1.5. R 3 a p p ea r s i n ter es ti n g , a s i t i n d i ca tes th e ex i s ten ce of a d a ta cl u s ter .

l es a re HS u le

T a b l e 1 . 5 . Str on g r u l es ob ta i n ed for CUL T_ F IN D N o 1 2 [ 3 4

[ 5 6 7 8 9

10 11 12

Ch a n g e i n COUR SE (G r ou p ed ) d i s tr i - B S S b u ti on : (n eg p os ) [ F OCAL : +] [ D i a g 2: VIR US] (8 1 17) / 98 Å (9 11) / 20 2.8 4 [ N AUSE A> 3 ] [ STIF F = 2, CUL T_ F IN D : F ] (27 9) / 3 6Å (0 5) / 5 2.8 1 THE R AP Y2: AR A_ A] [ SE X : F ] (4 8 10) / 58 Å (0 4 ) / 4 2.74 Al wa y s a ccom p a n i ed b y [ D IAG : VIR US(E ), B T”3 6.0, L OC_ D AT = +, F OCAL = +, E E G _ F OCUS = +] . 3 a p h a s i a a n d 1 a m n es i a . [ CT_ F IN D : n or m a l , ON SE T: SUB ACUTE ] (69 8 ) / 77Å (0 3 ) / 3 2.4 1 E E G _ F OCUS: @ [ F OCAL : +] [ D i a g 2: VIR US, CR P = 0] (50 12) / 62Å (5 8 ) / 13 2.3 1 [ CSF _ G L U> 70] [ 0< CSF _ P R O”60] (3 2 5) / 3 7Å (0 3 ) / 3 2.24 [ COL D ”/2&B'$7  (16 3 ) / 19Å (0 3 ) / 3 2.13 [ N AUSE A> 3 ] 3 00< Cel l _ Mon o”750] [ L OC> 1] [ CUL T_ F IN D : F ] (8 9 18 ) / 107Å (2 5) / 7 2.09 [ SE X : M, (24 5) / 29Å (0 3 ) / 3 [ CSF _ G L U”4 0] 2.05 THE R AP Y2: n o_ th er a p y ] a l l d ea d [ F E VE R > 10] [ 4 0< CSF _ G L U”50] (26 5) / 3 1Å (1 4 ) / 5 2.04 [ D IAG : VIR US(E )] [ SE X : F , CUL T_ F IN D : F ] (3 9 8 ) / 4 7Å (5 7) / 12 2.05 [ 100< CSF _ P R O”200] [ STIF F = 2] (3 7 13 ) / 50Å (5 9) / 14 2.05 Ma i n con d i ti on

P r econ d i ti on s

54 0

T. Ok a d a

. is th th th (a

Th e n u m b er of s u p p or ti n g i n s ta n ces for R u l e 9 on l y 3 , b u t th ey a r e a l l d ea d . Th ey a r e s h own i n e b ottom r i g h t p i e ch a r t i n F i g u r e 1.5. If we om i t e p r econ d i ti on [ Th er a p y 2: n o_ th er a p y ] d u r i n g e vi s u a l i n s p ecti on , we ca n r ecog n i z e 5 p os i ti ves l l d ea d ) a m on g 12 p a ti en ts i n th e s a m e p i e ch a r t.

7 4 .4 C o n c lu d in g R e m a r k s











Th e ca s ca d e m od el ca n p r ovi d e m a n y s tr on g r u l es effecti vel y . Som eti m es s ets of r el a ted r u l es a r e ) 0 fou n d b y u s i n g d i ffer en t B S S va l u es a n d p r econ d i 6(; ti on s . We th en h a ve to r efer th em to ex p er t eva l u a ti on to d eter m i n e th ei r i m p or ta n ce a n d va F i g . 1 . 5 . Vi s u a l i z a ti on of l i d i ty . Of s p eci a l i n ter es t i s wh eth er th e B S S va l R u l e 9 i n Ta b l e 5. u es a r e con s i s ten t wi th th e i m p or ta n ce of th e r u l es . On l y n o_ th er a p y a r e s h own . Al th ou g h th e r es u l ti n g r u l es a r e p ower fu l , s evB l a ck : p os ; wh i te: n eg . er a l i m p r ovem en ts a r e ex p ected to b etter ex p r es s th em . Th e fi r s t i s th e op ti m i z a ti on of r u l es . If we m ove th e s p l i t va l u es of ca teg or i z a ti on s , a n d a d d /r em ove p r econ d i ti on cl a u s es , th en th e r es u l ti n g r u l es wi l l s u r el y b e i m p r oved . Th e s econ d i s th e p r es en ta ti on of r el a ted r u l es . If th ey a r e ex p r es s ed i n a g r ou p s or ted b y th ei r B S S va l u es , a n a l y s i s wi l l b e ea s i er . Th e vi s u a l i z a ti on wa s a l s o p r oved to b e u s efu l i f va r i a b l es i n r u l e con d i ti on s a r e u s ed a s a x es . An a l y s ts ca n often d etect r ea s on a b l e/n on s en s e r u l es b y th e vi s u a l i n s p ecti on .

R e fe r e n c e s 74 .1 74 .2 74 .3

74 .4

74 .5 74 .6

Wa s h i o, T.: h ttp : //wwwa d a .a r .s a n k en .os a k a -u .a c.j p /p u b /wa s h i o/j k d d /j k d d cfp .h tm l Ok a d a , T.: F i n d i n g D i s cr i m i n a ti on R u l es Us i n g th e Ca s ca d e Mod el , J . J p n . S o c . A r t i f i c i a l I n t e l l i g e n c e , 1 5 , p p .3 21-3 3 0 (2000). Ok a d a , T.: R u l e In d u cti on i n Ca s ca d e Mod el b a s ed on Su m of Sq u a r es D ecom p os i ti on , P r i n c i p l e s o f D a t a M i n i n g a n d K n o w l e d g e D i s c o v e r y (P r o c . P K D D ’ 9 9 ), p p .4 68 -4 75, L N A I 1704 , Sp r i n g er -Ver l a g (1999). G i n i , C.W.: Va r i a b i l i ty a n d Mu ta b i l i ty , con tr i b u ti on to th e s tu d y of s ta ti s ti ca l d i s tr i b u ti on s a n d r el a ti on s , S t u d i E c o n o m i c o - G i u r i d i c i d e l l a R . U n i v e r s i t a d e C a g l i a r i (1912). R evi ewed i n L i g h t, R .J., Ma r g ol i n , B .H.: An An a l y s i s of Va r i a n ce for Ca teg or i ca l D a ta , J . A m e r . S t a t . A s s o c . 6 6 , p p .53 4 -54 4 (1971). Ok a d a , T.: E ffi ci en t D etecti on of L oca l In ter a cti on s i n th e Ca s ca d e Mod el , P r o c . P A K D D 2 0 0 0 , L N A I 18 05, p p .193 -203 , Sp r i n g er -Ver l a g (2000). Sp otfi r e D eci s i on s i te: h ttp : //www.s p otfi r e.com

75. Meningitis Data Mining by Cooperatively Using GDT-RS and RSBR Ning Zhong1 , Ju-Zhen Dong1 , and Setsuo Ohsuga2 1

2

Dept. of Information Eng., Maebashi Institute of Technology, Japan [email protected] Dept. of Information and Computer Science, Waseda University, Japan

This paper describes an application of two rough sets based systems, namely GDT-RS and RSBR respectively, for mining if-then rules in a meningitis dataset. GDT-RS (Generalized Distribution Table and Rough Set) is a soft hybrid induction system, and RSBR (Rough Sets with Boolean Reasoning) is used for discretization of real valued attributes as a preprocessing step realized before the GDT-RS starts. We argue that discretization of continuous valued attributes is an important pre-processing step in the rule discovery process. We illustrate the quality of rules discovered by GDT-RS is strongly affected by the result of discretization.

75.1 Introduction Rough set theory constitutes a sound basis for Knowledge Discovery and Data Mining. It offers useful tools to discover patterns hidden in data in many aspects [75.8, 75.9]. It can be used in different phases of knowledge discovery process such as attribute selection, attribute extraction, data reduction, decision rule generation, and pattern extraction (templates, association rules). This paper describes an application of two rough sets based systems, namely GDT-RS and RSBR respectively, for mining if-then rules in a meningitis dataset. The core of the rule discovery process is GDT-RS that is a soft hybrid induction system for discovering classification rules from databases with uncertainty and incompleteness [75.10, 75.2, 75.3]. The system is based on the combination of Generalization Distribution Table (GDT) and the rough set methodology. A GDT is a table in which the probabilistic relationships between concepts and instances over discrete domains are represented. The GDT provides a probabilistic basis for evaluating the strength of a rule. Furthermore, the rough set methodology is used to find minimal relative reducts from the set of rules with larger strengths. Furthermore, in the pre-processing before using GDT-RS, a system called RSBR is used for discretization of real valued attributes. The system is based on the combination of the rough set method and Boolean reasoning proposed by Nguyen and Skowron [75.6, 75.11]. A variant of the rule selection criteria in GDT-RS is used in RSBR. Thus, the process of the discretization of real valued attributes does not only mean to find the minimal relative reduct, but T. Ter a n o et a l . (E d s .): JSAI 2001 Wor k s h op s , L N AI 2253 , p p . 54 1− 54 8 , 2001. © Sp r i n g er -Ver l a g B er l i n Hei d el b er g 2001

542

N. Zhong, J.-Z. Dong, and S. Ohsuga

also considers the effect of the discretized attribute values on the performance of our induction system GDT-RS. We argue that discretization of continuous valued attributes is an important pre-processing step in the rule discovery process. Rules induced without discretization are of low quality because they will not recognize many new objects. We illustrate the quality of rules discovered by GDT-RS is strongly affected by the result of discretization.

75.2 Rule Discovery by GDT-RS GDT-RS is a soft hybrid induction system for discovering classification rules from databases with uncertain and incomplete data [75.10, 75.2]. The system is based on a hybridization of Generalization Distribution Table (GDT) and the Rough Set methodology. The main features of GDT-RS are the following: – Biases for search control can be selected in a flexible way. Background knowledge can be used as a bias to control the initiation of GDT and in the rule discovery process. – The rule discovery process is oriented toward inducing rules with high quality of classification of unseen instances. The rule uncertainty, including the ability to predict unseen instances, can be explicitly represented by the rule strength. – A minimal set of rules with the minimal (semi-minimal) description length, having large strength, and covering of all instances can be generated. – Interesting rules can be induced by selecting a discovery target and class transformation. In [75.10, 75.3], we illustrated the first two features. This paper discusses the last two features of the GDT-RS. 75.2.1 GDT and Rule Strength Any GDT consists of three components: possible instances, possible generalizations of instances, and probabilistic relationships between possible instances and possible generalizations. Here the possible instances are all possible combinations of attribute values in a database; the possible generalizations for instances are all possible cases of generalization for all possible instances; the probabilistic relationships between possible instances and possible generalizations, represented by entries Gij of a given GDT, are defined by means of a probabilistic distribution describing the strength of the relationship between any possible instance and any possible generalization. The prior distribution is assumed to be uniform, if background knowledge is not available. Thus, it is defined by Eq. (75.1)

75. Meningitis Data Mining by Cooperatively Using GDT-RS and RSBR

Gij = p(P Ij |P Gi ) =

⎧ ⎪ ⎨ ⎪ ⎩

1 NP G i 0

543

if P Ij ∈ P Gi (75.1) otherwise

where P Ij is the j-th possible instance, P Gi is the i-th possible generalization, and NP Gi is the number of the possible instances satisfying the i-th possible generalization, that is,  nk (75.2) NP G i = k∈{l| P Gi [l]=∗}

where P Gi [l] is the value of the l-th attribute in the possible generalization P Gi . nk is the number of different attribute values in attribute k. “∗”, which 1 specifies  a wild card, denotes the generalization for instances . Certainly we have j Gij = 1 for any i. Let us recall some basic notions for rule discovery from databases represented by decision tables in rough set theory. A decision table is a tuple T = (U, A, C, D), where U is a nonempty finite set of objects called the universe, A is a nonempty finite set of primitive attributes, and C, D ⊆ A are two subsets of attributes that are called condition and decision attributes, respectively [75.8, 75.9]. By IN D(B) we denote the indiscernibility relation defined by B ⊆ A, [x]IN D(B) denotes the indiscernibility (equivalence) class defined by x, and U/B the set of all indiscernibility classes of IN D(B). In our approach, the rules are expressed in the following form: P → Q with S that is, “if P then Q with the strength S” where P denotes a conjunction of conditions (i.e. P ⊆ C), Q denotes a concept that the rule describes (i.e. Q ⊆ D), S is a “measure of strength” of the rule. Furthermore, S consists of three parts: s(P ), accuracy, and coverage, where s(P ) is the strength of the generalization P (i.e. the condition of the rule), the accuracy of the rule is measured by a noise rate function: r(P → Q), coverage denotes how many instances are covered by the rule. If some instances covered by the rule also belong to another class, the coverage is a set: {number of instances belonging to the class, number of instances belonging to another class}. The strength of a given rule reflects the incompleteness and uncertainty in the process of rule inducing influenced both by unseen instances and noise. The strength of the generalization P = P G is given by Eq. (75.3) under that assumption that the prior distribution is uniform s(P ) =

 l

p(P Il |P ) = card([x]IN D(P ) ) ×

1 NP

(75.3)

where card([x]IN D(P ) ) is the number of observed instances satisfying the generalization P . The strength of the generalization P represents explicitly 1

For simplicity, we would like to omit the wild card in some places in this paper.

544

N. Zhong, J.-Z. Dong, and S. Ohsuga

the prediction for unseen instances since possible instances are considered. On the other hand, the noise rate is given by Eq. (75.4) r(P → Q) =

card([x]IN D(P ) ∩ [x]IN D(Q) ) card([x]IN D(P ) )

(75.4)

where card([x]IN D(Q) ) is the number of all instances from the class Q within the instances satisfying the generalization P . It shows the quality of classification measured by the number of instances satisfying the generalization P which cannot be classified into class Q. The user can specify an allowed noise level as a threshold value. Thus, the rule candidates with the larger noise level than a given threshold value will be deleted. 75.2.2 A Searching Algorithm for Optimal Set of Rules We now describe an idea of a searching algorithm for a set of rules developed in [75.2]. We use a sample database shown in Table 75.1 to illustrate the idea. Let Tnoise be a threshold value. Table 75.1. A sample database

HHA U H a u1 u2 u3 u4 u5 u6 u7

a0 a0 a0 a1 a0 a0 a1

b b0 b1 b0 b1 b0 b2 b1

c c1 c1 c1 c0 c1 c1 c1

d y y y n n n y

Step 1. Create one or more GDTs. If prior background knowledge is not available the prior distribution of a generalization is calculated using Eq. (75.1) and Eq. (75.2). Step 2. Consider the indiscernibility classes with respect to the condition attribute set C (such as u1 , u3 , and u5 in the sample database of Table 75.1)  as one instance, called a compound instance (such as u1 = [u1 ]IN D(a,b,c) in the following table). Then the probabilities of generalizations can be calculated correctly. XXX XXX A U b c d XX a  u1 , (u1 , u3 , u5 ) u2 u4 u6 u7

a0 a0 a1 a0 a1

b0 b1 b1 b2 b1

c1 c1 c0 c1 c1

y,y,n y n n y

75. Meningitis Data Mining by Cooperatively Using GDT-RS and RSBR

545



Step 3. For any compound instance u (such as the instance u1 in the above table), let d(u ) be the set of the decision classes to which the instances in u belong. Furthermore, let Xv = {x ∈ U : d(x) = v} be the decision class corresponding to the decision value v. The rate rv can be calculated by Eq. (75.4). If there exist a v ∈ d(u ) such that rv (u ) = min{rv (u )|v  ∈ d(u )} < Tnoise then we let the compound instance u to point to the decision class corresponding to v. If does not exist any v ∈ d(u ) such that rv (u ) < Tnoise , we treat the compound instance u as a contradictory one, and set the decision class of u to ⊥(uncertain). For example, XXX X A b c d U XXXX a  u1 (u1 , u3 , u5 ) a0 b0 c1 ⊥ 

Let U be the set of all the instances except the contradictory ones.  Step 4. Select one instance u from U . Using the idea of discernibility matrix, create a discernibility vector (that is, the row or the column with respect to u in the discernibility matrix) for u. For example, the discernibility vector for instance u2 : a0 b1 c1 is as follows: aa  U U aa u1 (⊥) u2 (y) u4 (n) u6 (n) u7 (y) u2 (y) b ∅ a, c b ∅ Step 5. Compute all the so called local relative reducts for the instance u by using the discernibility function. For example, from instance u2 :a0 b1 c1 , we obtain two reducts {a, b} and {b, c}: fT (u2) = (b) ∧ ) ∧ (a ∨ c) ∧ (b) ∧ ) = (a ∧ b) ∨ (b ∧ c). Step 6. Construct rules from the local reducts for the instance u, and revise the strength of each rule using Eqs. (75.3) and (75.4). For example, the following rules are acquired 1 {a0 b1 } → y with S = 1 × = 0.5, and 2 1 {b1 c1 } → y with S = 2 × = 1 for the instance u2 :a0 b1 c1 . 2 Step 7. Select the best rules from the rules (for u) obtained in Step 6 according to its priority [75.10]. For example, the rule “{b1 c1 } → y” is selected for the instance u2 :a0 b1 c1 because it matches more instances than the rule “{a0 b1 } → y”.    Step 8. U = U − {u}. If U = ∅, then go back to Step 4. Otherwise, go to Step 9. Step 9. If any rule selected in Step 7 is covering exactly one instance then STOP, otherwise, select a minimal set of rules covering all instances in the decision table. The following table shows the result for the sample database shown in Table 75.1.

546

N. Zhong, J.-Z. Dong, and S. Ohsuga U u1 , u 3 , u 5 u2 , u 7 u4 u6

rules b0 → y b1 ∧ c1 → y c0 → n b2 → n

s(P) 0.25 1 0.17 0.25

accuracy 0.67 1 1 1

coverage {2, 1} 2 1 1

One can see that the discovered rule set is a minimal one having large strength and covering of all instances. Furthermore, the searching algorithm can be conveniently used to discover a rule set with respect to an interesting class (or a subset of classes) selected by the user as a discovery target. Thus, by using class selection/transformation, and combining with the some preprocessing steps such as discretization, we can obtain more interesting results.

75.3 Discretization Based on RSBR Discretization of continuous valued attributes is an important pre-processing step in the process for rule discovery in the databases with mixed type of data including continuous valued attributes. In order to solve the discretization issues, we have developed a discretization system called RSBR that is based on hybridization of rough sets and Boolean reasoning proposed in [75.6]. A great effort has been made (see e.g. [75.5, 75.1, 75.4, 75.7]) to find effective methods for discretization of continuous valued attributes. We may obtain different results by using different discretization methods. The results of discretization affect directly the quality of the discovered rules. Some of discretization methods totally ignore the effect of the discretized attribute values on the performance of the induction algorithm. RSBR combines discretization of continuous valued attributes and classification together. In the process of the discretization of continuous valued attributes we should also take into account the effect of the discretization on the performance of our induction system GDT-RS. Roughly speaking, the basic concepts of the discretization based on RSBR can be summarized as follows: – Discretization of a decision table T = (U, A ∪ {d}), where Va = [va , wa ) is an interval of real values taken by attribute a is a searching process for a partition Pa of Va for any a ∈ A satisfying some optimization criteria (like minimal partition) preserving some discernibility constraints [75.6]. – Any partition of Va is defined by a sequence of the so-called cuts v1 < v2 < . . . < vk from Va . – Any family of partitions {Pa }a∈A can be identified with a set of cuts.

75.4 Application in Meningitis Data Mining This section shows the results of mining in a meningitis dataset by using cooperatively GDT-RS and RSBR.

75. Meningitis Data Mining by Cooperatively Using GDT-RS and RSBR

547

In the meningitis dataset, 19 of 38 attributes are continuous valued attributes that must be discretized by RSBR before rule induction by GDT-RS. Since the quality of rules discovered by GDT-RS is strongly affected by the result of discretization of continuous valued attributes, we need to do the discretization of continuous valued attributes carefully. Furthermore, in this experiment, for each decision attribute with multiclass, we used two different modes of cooperatively using GDT-RS and RSBR: 1. All classes in a decision attribute are considered simultaneously when using RSBR for discretization. 2. Focus on an interesting class selected by a user as positive class (+) and other classes are considered as negative class (–). The GDT-RS and RSBR are cooperatively used to find the rules with respect to the focused positive class. After that, a class with respect to negative class is selected as a new interesting positive class, and then the RSBR and GDT-RS are cooperatively used again. Repeat this process until all interesting classes are selected as positive class. Here we show an interesting result. That is, finding factors important for predicting prognosis (COURSE and C COURSE). First we consider all classes when discretization. The following 2 of 11 reasonable rules are interesting ones. r1.1 : F EV ER(≥ 8) ∧ BT (< 37.1) → CU LT U RE(−) with coverage = {13, 2}, accuracy = 86%. r1.2 : F OCAL(+) ∧ CT F IN D(normal) → CU LT U RE(−) with coverage = {12, 2}, accuracy = 85%.

Then we focus on an interesting class, that is, CULTURE(–). The following 2 of 26 reasonable rules are interesting ones. r2.1 : COLD(< 9) ∧ BT (≥ 37.1) ∧ LOC DAT (−)∧ Cell M ono(< 429) → CU LT U RE(−) with coverage = 31, accuracy = 1. r2.2 : COLD(< 9) ∧ LOC DAT (−) ∧ Cell P oly(≥ 32)∧ CSF P RO(< 93) → CU LT U RE(−) with coverage = 30, accuracy = 1.

According to a medical doctor opinion, all these rules (r1.1 , r1.2 and r2.1 , r2.2 ) are reasonable, but the rules, r2.1 , r2.2 , which are learned by focusing on an interesting class, are are much better ones. This example shows that more interesting rules can be generated by selecting an interesting discovery target because the better result of discretization is obtained.

75.5 Conclusion We have presented an application of two rough sets based systems, GDT-RS and RSBR, for mining if-then rules from a meningitis dataset. The experimental results illustrate that the quality of rules discovered by GDT-RS is

548

N. Zhong, J.-Z. Dong, and S. Ohsuga

strongly affected by the results of discretization of continuous valued attributes. We need to do the discretization of continuous valued attributes carefully. Using cooperatively RSBR and GDT-RS is a good way for rule discovery in the datasets with mixed type of attributes and multi-class. Acknowledgements. The authors would like to thank Prof. S. Tsumoto for providing the meningitis dataset and evaluating the experimental results.

References 75.1 Chmielewski, M.R. and Crzymala-Busse, J.W. (1994): “Global Discretization of Attributes as Preprocessing for Machine Learning”, Proc. Thrid Inter. Workshop on Rough Sets and Soft Computing, 294–301. 75.2 Dong, J.Z., Zhong, N., and Ohsuga, S. (1999): “Probabilistic Rough Induction: The GDT-RS Methodology and Algorithms”, Z.W. Ras and A. Skowron (eds.) Foundations of Intelligent Systems, LNAI 1609, Springer, 621–629. 75.3 Dong, J.Z., Zhong, N., and Ohsuga, S. (2000): “Rule Discovery by Probabilistic Rough Induction”, Journal of Japanese Society for Artificial Intelligence, Vol.15, No.2, 276–286. 75.4 Dougherty, J, Kohavi, R., and Sahami, M. (1995): “Supervised and Unsupervised Discretization of Continuous Features”, Proc. 12th Inter. Conf. on Machine Learning, 194–202. 75.5 Fayyad, U.M. and Irani, K.B. (1996): “On the Handling of ContinuousValued Attributes in Decison Tree Generation”, Machine Learning, Vol.8, 87–102. 75.6 Nguyen, H. Son, Skowron, A. (1995): “Quantization of Real Value Attributes”, P.P. Wang (ed.) Proc Inter. Workshop on Rough Sets and Soft Computing, 34–37. 75.7 Nguyen H. Son and Nguyen S. Hoa (1998): “Discretization Methods in Data Mining”, L. Polkowski, A. Skowron (eds.) Rough Sets in Knowledge Discovery, Physica-Verlag, 451–482. 75.8 Pawlak, Z. (1991): Rough Sets, Theoretical Aspects of Reasoning about Data, Kluwer. 75.9 Skowron, A. and Rauszer, C. (1992): “The Discernibility Matrixes and Functions in Information Systems”, R. Slowinski (ed.) Intelligent Decision Support, Kluwer, 331–362. 75.10 Zhong, N., Dong, J.Z., and Ohsuga, S. (1998): “Data Mining: A Probabilistic Rough Set Approach”, L. Polkowski and A. Skowron (eds.) Rough Sets in Knowledge Discovery, Vol.2, Physica-Verlag, 127–146. 75.11 Zhong, N., Dong, J.Z., and Ohsuga, S. (2000): “A Rough Sets Based Knowledge Discovery Process”, Proc. Fourth Asian Fuzzy Systems Symposium, 415–420.

Subject Index

L1 -space

289

adaptation of software components 370 additional semantics 380 affordance 35 agent based simulation 162 agent-based approach 99 agent-based simulation 142, 174, 218, 227 analysis of image sequences 333 analysis of self-injurious behavior 395 Analytic Hierarchical Process(AHP) 344 approximate reasoning 251, 359 apriori algorithm 525 architecture of rough set processor 406 artificial market 110, 121, 132 association mining 508 association rule 380 Association rules 390 automatic community broadcasting system 51 automatic compositon of inductive application 500 Avatamsaka situation 153 Avatmsaka game 155 basket analysis 516 bayes’theorem 240 bayesian network 364 Boxed Economy Foudation Model CAI 468 calcui of information granules cascade model 533 case-based reasoning 270 classification 364 clustering method 400 co-evolution 185

251

227

collaboration technology 35 collaborative innovation 27 combination rule 322 communication costs 51 communication media 11 communication systems 51 complementarities 156 complex economic system 142 complex social system 142 complexity of agents 110 complexity of markets 110 computational simulation 99 computer programs for trading 121 Conditional logic 390 conditional probability relation 311 conditional rule 322 conflict analysis 349 conflict model 349 conflict profile 327 contribution 447 control systems 491 coordinatin game 156 coordination game 155 counterexample 468 creative activity 449, 462 cumulative progress 99 data dependency 301 Data Envelopment Analysis(DEA) 344 Data mining 390 data mining 269 decision making 59 decision table 240 democracy on-line 67 design information management 475 dialectical argumentation 414 dialogue-games 414 discovery context 481 discovery dialogue 414 discovery learning environment 468

552

Subject Index

discovery of communities 435 discovery of future directions 435

long-term idea-generation 455 lower possibility distribution 272

e-democracy 67 EM-clustering 333 emergence 43 equivalence relation 295 evaluation of chance discovery 425 evolutionary game theory 89 extended simulated annealing(ESA) 306

meta-agent 162 methodology of chance discovery 425 mining association rule 475 mining ordering rule 316 mixture distribution model 289 modeling process 99 morphology 43 multi-agent economics 208 multi-agent system 43 multi-objective generic algorithm 132 munagemnet system for ideas 455 mutual choice 174

failure prevention 475 fault detection 475 fractal 278 fuzzy c-means model 289 fuzzy knowledge based systems fuzzy multiset 283

491

generalization of rough set 311 generalized distribution table and rough set(GDT-RS) 541 genetic algorithm 185 genetic programming 525 GOLEM 339 Graded conditional logic 390 Granular reasoning 385 granule decomposition 359 graph–VPRSILP 340, 341 Human reasoning

385

imperfect data 354 inductive logic programming(ILP) 354 information system 278 interaction 35, 59 interval comparison matrix 344 interval probability function 322 interval-free active search 486 inverse simulation 91 IPD 185 keygraph 481 knowledge discovery in databases(KDD) 269 knowledge transaction 195 learning data set 306 learning from examples based on rough sets(LERS) 395 learning unbalanced positive class(LUPA) 508 level-two control systems 491 levy plan 162

network analysis 78 non-additivity 322 non-deterministic information nonlinear evoution 160 norms emergence 88 norms game 174

301

ontologies for inductive learning 500 open system 425 ordered information tables 316 ordering of objects 316 participation methology 67 path-dependent 157 personal information storage system 455 POC see Public Opinion Channel procedure for dependency 301 public discourse 59 Public Opinion Channel 75 public opinion channel 51 purchase consulting system 462 Relative filtration 385 retrieval 486 rough integral 375 rough measure 375 rough membership function 375 rough neural network 251 rough set 295 Rough set theory 385 rough set theory 240 rough sets with boolean reasoning(RSBR) 541 RS-ILP 354 rule discovery 525 rule induction 508

Subject Index shared knowledge 43 SI see Social intelligence small world 444 social intelligence 75 Social Intelligence Design 3 social intelligence design 75 social interaction 195 speech recognition 306 spillovers 156 stochastic process – Polya 158 susceptibility to consensus 327 time-series active search 486 trading agent 121, 132 tragedy of the common 162 TRUL 91

553

U-Mart 121, 132 upper possibility distribution 272 urn process – generalized version of Polya 158 – Polya 157 virtual community 19 virtual economy 218 virtual reality 19 visualization of communication VPRSILP model 339 Web usage graphs 340 World Trade League 208

11

Author Index

Abe, Hidenao 500 Aruka, Yuji 153 Borkowski, M.

Kamihashi, Kenichi 227 Kanasugi, Akinori 406 Kaneda, Toshiyuki 218 Kato, Yoshikiyo 475 Kawamura, Hidenori 174 Kawasaki, S. 508 Kita, Hajime 121, 132 Kitano, Satomi 227 Koga, Takatsugu 289 Komori, Mao 500 Kurahashi, Setsuya 88 Kurumatani, Koichi 208, 218 Kuwabara, Kazuhiro 11

375

Cardon, Alain 43 Cho, Sung-Bae 185 Chubachi, Yoshihide 227 Deguchi, Hiroshi 121, 218 Deja, Rafal 349 Dong, Ju-Zhen 541 Entani, Tomoe

344 Lee, Won Don 306 Lin, T.Y. 380 Liu, Chunnian 354 Louie, Eric 380 Luehrs, R. 67

Freeman, Rachel L. 395 Fruchter, R. 35 Fujihara, Nobuhiko 75 Fukuhara, Tomohiro 51 Fukumoto, Rikiya 132 Grzymala-Busse, Jerzy W. Guo, P. 272

395

Hashimoto, Shigeji 218 Hata, Yutaka 400 Hatazawa, Hiromitsu 500 Hirano, Shoji 239, 400 Hirashima, Tsukasa 468 Hirokane, Masaharu 227 Ho, T.B. 508 Hori, Koichi 455, 462, 475 Horiguchi, Tomoya 468 Iba, Takashi 227 Ichihashi, Hidetomo 344 Ikeda, Takayuki 516 Intan, Rolly 311 Inuiguchi, Masahiro 239, 295 Ishizuka, Mitsuru 425, 435, 444 Izumi, Kiyoshi 110

Maeda, Yutaka 322 Maheswari, V. Uma 339 Malsch, T. 67 Matsui, Hiroyuki 121, 218 Matsumoto, Masao J. 370 Matsumura, Naohiro 435 Matsuo, Yutaka 444 Matsuzawa, Yoshiaki 227 McBurney, Peter 414 Mehata, K.M. 339 Miyamoto, Sadaaki 283, 289 Mizuta, Hideyuki 142 Motoda, Hiroshi 516 Mukaidono, Masao 311 Murai, Tetsuya 385, 390 Nakata, Keiichi 59 Nakata, Michinori 385, 390 Namatame, Akira 85, 195 Nara, Yumiko 481 Nguyen, D.D. 508

550

Author Index

Nguyen, Hung Son 333 Nguyen, Ngoc Thanh 327 Niimi, Ayahiko 525 Nijholt, A. 19 Nishida, Toyoaki 3, 51 Nishimura, Takuichi 486 Ohguro, Takeshi 11 Ohsawa, Yukio 413, 435, 444, 481 Ohsuga, Setsuo 541 Ohuchi, Azuma 174, 208 Oka, Ryuichi 486 Okada, Takashi 533 Okuzaki, Tomohiro 400 Ono, Isao 121 Owada, Tatsuo 11 Pal, Sankar K. 261 Parsons, Simon 414 Pawlak, Zdzislaw 240, 375 Peters, James F. 359, 375 Polkowski, Lech 278 Prendinger, Helmut 425 Ramanna, S. 375 Riffel, Laura A. 395

Skowron, Andrzej 251, 333, 359, 375 ´ ˛zak, Dominik 349, 364 Sle Song, Chi-Hwa 306 Stepaniuk, Jaroslaw 359 Sugihara, Kazutomi 322 Sunayama, Wataru 449 Suraj, Z. 375 Suzuki, Keji 162 Szczuka, Marcin S. 333 Tachibana, Yoshiaki 500 Takabe, Yohei 227 Takadama, Keiki 99 Takata, Osamu 289 Tanaka, Hideo 272, 322, 344 Tanaka, Junichiro 227 Tanino, Tetsuzo 295 Tazaki, Eiichiro 525 Terano, Takao 88, 121, 218 Thomas, John C. 27 Torra, Vicen¸c 491 Tsumoto, Shusaku 239, 400 Tsuya, Ryunosuke 227 Uemura, Shunsuke Voss, K.

Sai, Ying 316 Sakai, Hiroshi 301 Sashima, Akio 218 Sato, Hiroshi 121 Sato, Kazuyo 195 Sato, Yoshiharu 385, 390 Schroeder, Stephen R. 395 Shibata, Hirohito 455 Shimohara, Katsunori 99 Shinkawa, Yoshiyuki 370 Shiozawa, Yoshinori 121 Shirai, Yoshinari 11 Shoji, Hiroko 462 Siromoney, Arul 339

51

67

Washio, Takashi

499, 516

Yachida, Masahiko 449 Yairi, Takehisa 475 Yamagata, Yoshiki 142 Yamaguchi, Takahira 500 Yamamoto, Masahito 174 Yamashita, Tomohisa 174 Yao, Y.Y. 311, 316 Yuzawa, Taro 218 Zhong, Ning

354, 541