You are on page 1of 398

ATLANTIS C OMPUTATIONAL I NTELLIGENCE S YSTEMS VOLUME 2 S ERIES E DITOR : DA RUAN

Atlantis Computational Intelligence Systems Series Editor: Da Ruan, Belgian Nuclear Research Centre (SCK • CEN) Mol & Ghent University, Gent, Belgium
(ISSN: 1875-7650)

Aims and scope of the series The series ‘Atlantis Computational Intelligence Systems’ aims at covering state-of-the-art research and development in all fields where computational intelligence (CI) is investigated and applied. The series seeks to publish monographs and edited volumes on foundations and new developments in the field of computational intelligence, including fundamental and applied research as well as work describing new, emerging technologies originating from computational intelligence research. Applied CI research may range from CI applications in the industry to research projects in the life sciences, including research in biology, physics, chemistry and the neurosciences. All books in this series are co-published with World Scientific. For more information on this series and our other book series, please visit our website at: www.atlantis-press.com/publications/books

A MSTERDAM – PARIS

c ATLANTIS PRESS / WORLD SCIENTIFIC

Computational Intelligence in Complex Decision Systems

Da Ruan
Belgian Nuclear Research Centre (SCK • CEN) Mol & Ghent University, Gent, Belgium

A MSTERDAM – PARIS

Atlantis Press 29, avenue Laumi` re e 75019 Paris, France For information on all Atlantis Press publications, visit our website at: www.atlantis-press.com Copyright This book is published under the Creative Commons Attribution-Non-commercial license, meaning that copying, distribution, transmitting and adapting the book is permitted, provided that this is done for non-commercial purposes and that the book is attributed. This book, or any parts thereof, may not be reproduced for commercial purposes in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system known or to be invented, without prior permission from the Publisher.

Atlantis Computational Intelligence Systems Volume 1: Linguistic Values Based Intelligent Information Processing: Theory, Methods, and Applications – Zheng Pei, Da Ruan, Jun Liu, Yang Xu

ISBN: 978-90-78677-27-7 ISSN: 1875-7650

e-ISBN: 978-94-91216-29-9

c 2010 ATLANTIS PRESS / WORLD SCIENTIFIC

Preface

In recent years there has been a growing interest in the need for designing computational intelligence to address complex decision systems. One of the most challenging issues for computational intelligence is to effectively handle real-world uncertainties that cannot be eliminated. These uncertainties include various type of information that is incomplete, imprecise, fragmentary, not fully reliable, vague, contradictory, deficient, and overloading. These uncertainties result in a lack of the full and precise knowledge of the decision system including the determining and selection of evaluation criteria, alternatives, weights, assignment scores, and the final integrated decision result. Computational intelligent techniques including fuzzy logic, neural networks, and genetic algorithms etc., as complimentary to the existing traditional techniques, have shown great potential to solve these demanding, real-world decision problems that exist in uncertain and unpredictable environments. These technologies have formed the foundation for computational intelligence.

To overview the role of computational intelligence in information–driven complex decision systems and illustrate the potential use and practical applications of computational intelligence related techniques in complex decision systems, this edited volume presents recent research results and provides a state-of-the-art on the future research directions. The book gathers a peer-reviewed collection of a number of representative applications of Computational Intelligence in Complex Decision Systems. It contains 13 chapters written by 25 authors from Australia, Belgium, Brazil, China, India, Japan, Portugal, Spain, Turkey, the UK, and the USA. These contributions and applications cover Complex Decision Systems ranging from Computational Intelligence (Chapter 1 by Kahraman et al.), Dynamically Changing Input Data (Chapter 2 by Pais et al.), Possibilistic Linear Programming (Chapter 3 by Guo), Virtual Reality (Chapter 4 by Machado and de Moraes), Many-Valued Temporal Logic and Reasoning (Chapter 5 by Lu et al.), Statistical Approach (Chapter 6
v

vi

Computational Intelligence in Complex Decision Systems

by Kunsch), Web based Assessment Tool (Chapter 7 by Xu), Intelligent Policy Simulator (Chapter 8 by Rao), Computing with Words (Chapter 9 by Mendel and Wu), Policy Realization (Chapter 10 by Kahraman and Kaya), Load Dispatch (Chapter 11 by Zhang et al.), Smart Home (Chapter 12 by Muˇ oz et al.), to Agile Supply Chain Strategies (Chapter 13 n by B¨ y¨ k¨ zkan). u u o Each chapter of the book is self-contained. Academic and applied researchers and research students working on complex decision systems can also benefit from this book.

March 2010 Da Ruan Belgian Nuclear Research Centre (SCK • CEN) Mol & Ghent University, Gent, Belgium

. . . . . . . . . . Learning processes . . . . . . . . . . . . . . . . . . . . . . .4. . .2. . . Kahraman. . .3 1. .4. . . . . . . 24 Literature Review . . . Computational Intelligence: Past. . . . .4. . . . . . . .5 1. . . . . . . . . .8 Hybrid Systems . . . . . .1 Computational intelligence in complex decision systems . . v 1 1 5 6 9 Fuzzy Set Theory . . . . . . . . . 23 1. . . . . . . . . . . .6. . . . . . . . . . . . . . . . . .2 1. .2. . . . . . . . . . . . . . . . . . . . . . . . . . . .1 1. .4 Network Architectures . . . . . . . I. . . .1 1. . . . . . . .2 Introduction . . . . . . . . . . . . . . 29 Computational Intelligence Journals . . . . . . . . .2 1. 12 Evolutionary Computation . . . . . . . .4. . . . . 24 1. . . . . . . . . . . Today. . . . . . 21 Particle swarm optimization . . . . . . 20 Ant colony optimization . . . . . . .Contents Preface 1. . .3 1. . . . . . . . .6 1. . . . . . . . . . . . . . . . . . . . 36 Bibliography . . . .7 Genetic algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7 1. . . 14 Genetic programming . . . and Future ˙ C. . . . . . . . . . . 1. . Kaya. . . . and D. . . . .5 1. . . 38 vii . . . . . . . . . . .4. . . . . . . . . .4. 14 1. . . . 18 Evolution strategies . . . Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . 19 Classifier systems . . . . . . . . . . . 19 Evolutionary programming . . . . . . . . . . . . . . . . . . . . . .6 1. . . . . . .1 1. . Cınar ¸ 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. . . . . . . . . .4 1. . 34 Computational Intelligence’s Future and Conclusions . . . . . . .

. . . Marcos de Moraes 4. . . . . . . . 55 Phase B. . . . . . . . . 87 4. . . . . . . . . . . . . . . . . . . . .F. . 67 Fuzzy Decisions in Possibility Linear Programming Problems . . . . . . . .viii Computational Intelligence in Complex Decision Systems 2. . . . . . . . . . . . . . .1 2. . . . . . . . . Decision Evaluation Process with Feedback . . . . . . . . . 52 Background about Landing Site Selection . . . . . 89 85 Training in Virtual Reality Systems . Uncertainty in Dynamically Changing Input Data T. . . .2. . Sim˜ es o 2. . . . . . . . . . . . 54 Phase A. . . . . . . . . . . .1 4. . . . . . . . . . . 65 Bibliography . . . . 91 Collecting and Using Data from VR Simulators . . . 47 Description of Methodology . . . .C. . . . . . 65 3. . . .3 4. . . . . . . . . .4 Future Research Directions . . . . . . . . . 88 Haptic Devices . . . . . . . . . 94 .4. . . . Ribeiro. . . . . . .3. . . .2. . Intelligent Decision Making in Training Based on Virtual Reality L. . . . . . . . . . . . 49 2. . . . . . . . . . .3. . .3 Example in Practice . . . . . . . . . . . . . . . . .2 Introduction . . . . . . . .1 Online and Offline Evaluation . . .4 Interaction Devices . . . . . . . Guo 3. . 82 Bibliography . . 68 3. . . . . . . . . . . . . .2. .2 2. . . . . . . . . . .2. . . .1 2. . .2 Training Aspects . . . Decision Evaluation Process with Feedback . . . . . . . . . . . . . . . . . . and L. . . .2 Phase A. . . . . Data Preparation in Dynamic Environments . . . . . . . . . . . . . . . 64 2. . . . . . . dos Santos Machado and R. . . .2 47 Problem Statement . . . . . . . . . . . . 79 Conclusions . . . . . . . . . . .A. . . . . . . . . . . . . . . . . . . .1 2. . 50 Phase B. . . . . . . . . . . . . . . . . . . Pais.1 3. . . . .2 4.3. . . . . . . . .4 Triangular Possibility Distributions of Fuzzy Decision Variables 69 Exponential Possibility Distributions of Fuzzy Decision Variables 73 67 Numerical Examples . . . . . . . . . Decision Making under Uncertainty by Possibilistic Linear Programming Problems P. . .3 3. . . . .1 3. . 93 4. .2 3. . . . . Data Preparation in Dynamic Environments .2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2. . . . . . .3 2.1 4. . . R. . . . . . . . . . 85 Virtual Reality . 54 2. . . . . . . . . 83 4. . . . .

130 Inference Rules . . . . . . . . . . . . . . . . . . . . .1 4. . 109 Comparison of Hidden Markov Models . . . . . . . .7 Evaluation Methods Based on Hidden Markov Models . . .6. . . . Wang 5. . . . . . . . . . . . . . . . . . . 108 Continuous Hidden Markov Models . . . . . . . . . . 103 Sequential Methods . 95 Using Fuzzy Logic . . . . . . . . 100 Fuzzy Gaussian Mixture Models . . . . . J. 110 Maximum Likelihood . . . . . . . . . . . . . . . . . . 113 Quantitative and Qualitative Evaluation Based on Naive Bayes . . 136 125 Many-valued Temporal Propositional Logic Systems . .2. . . . . . . . . . . . . . . . . . . . . . 108 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . .1. . . . . . . . .6. . . 98 Gaussian Mixture Models . .1 5. . . . .4 4. . .8. . . . . . . . . . . . 100 4. .3 4. . . . . . . . . . . .4 4. .5 Evaluation Based on Expert Systems .2 4.2 5. . . . . .5. . 110 Fuzzy Hidden Markov Models . . 125 5. . . . . . . . 132 Logical Calculus .9 4. . . . . . . . . . . . . . . . . . . . . .C. . . . .4 Problem Description: a Realistic Scenario . . .2 4.7. . . . . . . . 95 4.3 4. . 118 Future Trends and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . .10 Considerations About Evaluation Methods . . . . . . . . . . . . . . . . . 115 Bayesian Networks . . . . . .1 5. . . . . . . . . Liu. . 113 Naive Bayes . . .2.5. . . . . 105 Discrete Hidden Markov Models (DHMM) . . .2. . . . . . . . . J. . . . . . . . .2 5. . . . . . . . . . . . . .1 4. . 134 Soundness and Completeness Theorems . . . .8. . . . . . . A Many-Valued Temporal Logic and Reasoning Framework for Decision Making Z. . . . . . . . . 116 4.3 Using Classical Logic . .6 Evaluation Based on Mixture Models . . . Lu. . . . .2 4. . . .7. .8. .1. . . . . . . . . . . . . .1 4. . . . 96 Combining Expert Systems . . . . . . .Contents ix 4. 128 Syntax and Semantics . .6. . . . . . . . . . . . . . . . 111 4. . . 129 . .1 4. . .3 5. . . . . .2 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 Relevant Works . . . . .5 4. .5. . . . . . . . . . . . . . . . . . . . . . . . . .8. .8. . . . .2. . . .7. . . . . . . . . . . . . . . . . . . . . . Augusto. . . . . . .3 4. . . . . . . . . . . . . . . . . . 119 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . .2 5. .7. . . . . . . . . . and H. 120 5. . 111 Fuzzy Bayes Rule . . . .8 Evaluation Methods Based on Bayesian Models . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Properties of the space of n-vectors . . . . .4 5. . . . . . 145 6. . . . . . . . . . . . . .3 6. . . . . . . .4 Introduction . . . . . . . .3 Practical Many-valued Temporal Reasoning . . . . . . . . . . . . . . . .1 6. . . . . . 158 The 2-D case . . . . . .1 6. . . . . . . . . . . . . . . . . . . . 185 Illustration of the ER Approach . . . . . . . . . . . . . . . . . . . . . . . . . . .2 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. . . 176 Projection in the XY -plane of the 74 ranking vectors (n = 4) on the unit sphere in R3 . . . . . . . . . .6 Scenarios . . . . . .2 Background . . . . . . . . . . . . . . . . . . . . . . . . . 138 5. . . . . . .1 A. . . . . . . . . . . . . . . . . . . . . . 155 The inverse procedure for finding preference weights . . . . . .2 7. . . . . . . . .3 7. . . . . . . . . . . . . . . . . . . . . .3. . . 162 A MCDM procedure . . 181 7. . . . . . . . . . . . . 173 A. 138 Backward Reasoning Algorithm .2 6. .3. . . . . 183 The ER Approach . . . . 165 An application of the cardinal procedure . 144 Bibliography . .1 5. . . . . . . . . . . . . . . . .-L. . . . . . . . . . . . 192 183 . . . . . . . . 165 Conclusions and Outlook . 160 The 3-D Case . . . . . . .4. . Xu 7. . . . . . 147 Basic Elements of the MCDM Methodology . . . A Web Based Assessment Tool via the Evidential Reasoning Approach D. . . . . . . . . . . . . . . . .2 Forward Reasoning Algorithm . . . . . . 144 Conclusions . . . . . . . . . . . . . . . . Kunsch 6. A Statistical Approach to Complex Multi-Criteria Decisions P. . . . . .3 6. . . . . . . . . . . . . . . . . . . . . . . . . .2. . . . . . . . . . . . . . . . .x Computational Intelligence in Complex Decision Systems 5.3. . . 139 5. .5 An ordinal procedure for finding the Rank Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 147 2-D and 3-D Visualisation in the Ordinal Case . . . . . . . .1 6. . . . . . . . . . 149 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186 Interfaces for Data Input . . . . . .1 7.5 5. . . 160 An Extension of the Approach to Cardinal Scales . . . . . . . . . . . . . 149 The maximum number of criteria . . . . . . . . . . . . . . . .3. . . . . . . . . . . . . . . . . .2. . 176 Bibliography . . . . .1 6. . . 180 Appendix A . . .2 6. 140 Discussions . . . .4 6. . . . . . .L. . . . . .2. . . . . . .

. . . . . . . . b. . . .6 8. . . . . . . . 216 Construction of the DiNI Atoms for Peace Policy Simulator . . .3. . . . . . . . . . .7. . .6 8. . . . . . .7. . . . h and d for the DiNI Atoms for peace policy simulator . . . . . . . . . . . . . . . . . . .2 8. . . .7 Inspiration from Nature: The Genetic Evolution of Altruism . . . . . . . . . . . 197 Applications . . b. . . . . .7. .5 7. . . . . . . . . . . . . 207 A Suggested Policy Simulator Based on the Genetic Evolution of Altruism 208 8. . . 199 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . .4 8. . .8 8. . . .7. . . . . . . . . . . . . . . . . 221 DiNI Atoms for Peace Simulation and Results . .4 8. . 211 DiNI Model Construction . . .A Brief Background . . . 205 Existing Intelligence Methods in Discerning Proliferation Intentions of Nation-states . 209 Peace as a State of Dynamic Equilibrium: Altruism in the Nuclear World . . . . . . . . . . . . . . . . . . . . . 210 Modeling Basis for DINI . . . . . . . . . . . . . 219 a Evaluating the AfP Experience of USA vis-` -vis India a The 1960s . . . . 216 8.1 Calculating Parameter Values c. 212 DiNI Model Simulation and Results . . . . . 199 Bibliography . . . h and d for DiNI . . .7 Interfaces for Outcome Display . . . . . . . . . . 212 8. . .7. . . . .5 8. . . . . . 225 Atoms for Peace: Some Policy Options for the USA . . . .7. . . . .Base Case . . . . . . . . . . . . . . . . . . 218 The AfP Experience of USA vis-` -vis India . . . . . . . . . . . . . . . . . .1 8. . . . .3 Introduction . . . . . . .1 8. . . . . . . . . 228 Ethical and Security Aspects of DiNI Policy Simulator . . . . . . . . . . . . . . . . . . . . . .5 8. . . .5. 209 DiNI Policy Simulator: Validity and Verification . .7. .6 7. .3 8. . . . . . . . . . . . . . . . .9 Nuclear Cooperation Between USA and India . .Nature’s Altruism Model in NetLogo . . . . . . . . . . . . . . . .7 8. . . . . . . . . . . . .3. . . . .2 8. . . .3. .3 8. . . . . . 227 205 Advantages of the DiNI Policy Simulator . . .1 8. . .2 8. . . . . . . . An Intelligent Policy Simulator for Supporting Strategic Nuclear Policy Decision Making S. . . . . . . . . . . . .Contents xi 7. . . . . . . 217 Calculating Parameter Values for c. . . . . Rao 8. . . . . . 229 . . . . . . . 215 Policy Simulation on DiNI: The Experience of USA vis-` -vis India a under the Atoms for Peace program . . . 203 8. . . .

. . . . . . . . . . 273 Multiple Criteria Decision Making . . . . . . . . . . . . . . . . . . . . . . . . 287 Conclusions . . . .5 Introduction . 253 Complete Reviews . . .4. . .2 Interval Type-2 Fuzzy Sets (IT2 FSs) . . . . .4. . . . . . . . . . . . . . . . 258 233 Examples . . . . . . . . . . . 230 9. . . . . . . 229 Conclusions . . . . . .2 10. . . .3 Introduction . . . . . . . . . . . . . . . . . . . . . . .5 9. . . . . 240 CWW Engine . . . . . . . . . . . . . . . . . 270 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . Mendel and D. . . . . . . .4. . . . . . . . . . .4 10. . . . . . . . . . . . 239 Encoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233 The Journal Publication Judgment Advisor (JPJA) . . . . . . . . .4.M. . . . . . .1 9. . . . . . . . . 266 Appendix A . . . . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . .5 Modified Paper Review Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kaya 10. . .4. . . . .1 9.3 9. .1 10. . . . .11 Further Research Directions . . . . . . . . . . . . . . . 252 Aggregation at the AE Level . . . . . . . . .10 8.4 9. . . . . . . . . . . . Wu 9. . . . . . . . Computing with Words for Hierarchical and Distributed Decision-Making J. . .1 9. .4 9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .xii Computational Intelligence in Complex Decision Systems 8. . . . . . . . . . . . . . . . . . . . . . .4 9. . . . . . . . . . . . 245 Aggregation of Presentation Sub-criteria . . . . 299 . . . . . . . . . . . . . . . . . . . . . . 242 Decoder . . . . . .2 9. 275 The Proposed Methodology . . . Realizing Policies by Projects Using FMCDM ˙ C.3 10. . . . .3. . . . . . . . . . 265 Bibliography . . . . 249 Aggregation at the Reviewer Level . . . . . . . . . . . . 237 Perceptual Computing for the JPJA . . 279 An Application . . . Kahraman and I. . . . . . . . 265 Linguistic Weighted Average (LWA) . . .2 9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230 Bibliography . . .2 9. . . . . . 239 9.3. . . . . .1 A. . 243 Aggregation of Technical Merit Sub-criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262 A. . 297 273 Bibliography . . . . . . . . . . . . . . . . . . . . . . . .3 9. . . .

. . Y. . . . . .C. . . . . . .1 Models for Day-ahead Markets . . . . . . . . . . . . . . 329 12. . . . . 302 11. . . . . . . . . . . . . .4. . . . 347 12.1 Information Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Evolutionary Computation Methods for Fuzzy Decision Making on Load Dispatch Problems G. . . . . . . . . .2 A Load Dispatch Model for a Day-ahead Market . . . . . . . .5 Decision Mechanism: Two Approaches . . . . . . Lu 11. . . .2. . . . . . . . . 322 12. 352 12.3.2 An Example of Uncertain Environment/Economic Load Dispatch 317 11. . . . . .2. . . . . . . . . . . . . . . . . .1 Same Conclusion Supported by Different Type of Information . . . .2 A Fuzzy Multi-object Non-linear Optimization Method . 350 12. Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants A. . . . . . . . . 334 Domain Criteria in the SmartTV System .2. . . . . . . . . . . . . . . . 305 11. . . and J. .2 Inconsistent Conclusions Supported by Different Type of Information . . . . . . . . . . . . . . . Zhang.3 Illustrations on Load Dispatch for Day-ahead Market .3 12. . . . . . Gao. . Mu˜ oz. . . . . .1 An example of load dispatch in a day-ahead market . . . . . . . . . . . . . . 316 11. . . . . . . . . .4 Arguments in the SmartTV System . . . . . . . . 332 12. . . . . . . . . . . . .2 Introduction . . . . . . . . . . Augusto n ı 12. . . . . 304 11. . . . . . . Bot´a. . . . . . .1. . 302 11.4. .2. . . . . . .2 A Distributed Approach . . . . . . . . . . G. .1. . . . . . . 343 12. . . . . . . . . . . . . . . . . . 344 12. . . J. . .1 A Centralized Approach . 325 SmartTV System Architecture . . . . . . . 316 11. . . . . .4 Conclusions and Further Research . . . .2 Decision Module . . .4.1 12. . . . . . . .3 Inconsistent Conclusions Supported by the Same Type of Information . 301 11.Contents xiii 11. .5. . . . . . 321 Bibliography . . .5. . .1. . . . . 313 11. . . . . . . 306 11. . . . . . . . . . . . . . . . .1 Introduction . . . . . . .3 An Uncertain Environment/Economic Load Dispatch Model . . . . . . . . 352 12. . . . . . . 330 12.1 Evolutionary Computation . .A.3. and J. . . . . . . . Zhang. . . . . . . . .2 Evolutionary Computation Methods and Fuzzy Decision Making . . . . . . . . . . . . . 353 325 301 . . . .

. . . . . . . . 373 Multi Criteria Decision Making with Choquet Integral . .6 12. . . . . . . . . . . . . 379 Concluding Remarks . .2 Elucidation of the Aggregation Result . . .2 Introduction . . . . . . . . . . . . . . . . . . . . . 384 373 Bibliography . . . . . 376 13. . . . . . . .1 13. . . . . . . . . . . . . . .2. . . . .3 13. . . 365 Conclusions and Further Research Directions . . . . . . . .3 SmartTV Scenarios . . . . . .4 Evaluation of Agile Supply Chain Strategies . . . . . . . . . . . . . . . . . . . . . 367 Bibliography . . . . . . 375 13. . . . . . . . . . 384 Subject Index 387 . . . . .5. . B¨ y¨ k¨ zkan u u o 13. . . . . 369 13. . . . . . . . . . . . . . . . . . . . . . Applying a Choquet Integral Based Decision to Evaluate Agile Supply Chain Strategies G. . . . . . . . . . . . . .1 Background and Notation . . . . .2. . .xiv Computational Intelligence in Complex Decision Systems 12. . . . 377 13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356 12. . . . . . . . . . .7 Comparing Decision Mechanisms in the SmartTV System . . . . . . . . . . .

it must consistently make a decision to select of one from a number of alternative ways of allocating the available resources for a given purpose. DOI 10. The primary goal of all living systems is survival and selection eliminates the weakest solutions variants while evolution as a process is purposeless. and Didem Cınar ¸ ˙ Istanbul Technical University. The intelligence was defined as the ability of solving the hard problems. 34367. Then. Department of Industrial Engineering. Intelligence comprises decision making. kayai. a detailed classification of the existing methodologies will be made. Later.tr Computational intelligence is defined as the study of the design of intelligent agents.1 Introduction A definition of intelligence which is needed to be defined to research in a field termed artificial intelligence and computational intelligence has only rarely been provided. In this chapter the history of computational intelligence with a wide literature review will be first given. As an example. © 2010 Atlantis Press/World Scientific 1 . Today. However there are basic questions in this definition. a special literature review on computational intelligence in complex decision systems will be also given. such as how hard the problem is or who decides which problem is hard [11]. Biological organisms satisfy the condition of intelligence as the derivation from their competition for the available resources.edu. The direction of computational intelligence in the future will be evaluated. ¸ ˙ Istanbul. and the definitions in the literature have often been of little operational value. 1.Chapter 1 Computational Intelligence: Past. computational intelligence means different things to different people.Macka. The weakest solution is the one which does D.1007/978-94-91216-29-9_1. For any intelligent system. Turkey E-mails: {kahramanc. Computational Intelligence in Complex Decision Systems. Since its relation to other branches of computer science is not well-defined. the international computational intelligence journals will be handled and the characteristics of these journals will be examined. and Future Cengiz Kahraman. Atlantis Computational Intelligence Systems 2. ˙ Ihsan Kaya. cinard}@itu. Ruan.

CI is generally accepted to include evolutionary computation. which consists of neural networks. CI can be broadly defined as the ability of a machine to react to an environment in new ways. and additionally when it (begins to) exhibit: (i) computational adaptivity. The birth of Computational Intelligence (CI) is attributed to the IEEE World Congress on Computational Intelligence in 1994 Orlando. This term combines elements of learning. A system is computationally intelligent when it: deals with only numerical (low-level) data. a significant characteristic of CI is that the precise model needs not to be established when dealing with imprecise. Chellapilla and Fogel [11] defined the intelligence as “the capability of a system to adapt its behavior to meet its goals in a range of environments and the life process itself provides the most common form of intelligence”. in some sense. IEEE Computational Intelligence Society defines its subjects of interest as neural networks. Compared with the traditional artificial intelligence. Florida. Since that time not only a great number of papers and scientific events have been dedicated to CI. uncertain. Cınar I. Despite the relatively widespread use of the term CI. but numerous explanations of the term have been published. and D. This basic idea has been recurred through the transitions of generations. . So intelligence is the basic property of decision maker. fuzzy logic and evolutionary computing. neural networks. In order to have a brief outline of history of the term the founding and most interesting definitions will be summarized now: The term CI was first introduced by Bezdek in 1994. (iii) speed approaching human-like turnaround. making useful decisions in light of current and previous information. there is no commonly accepted definition of the term. fuzzy systems. intelligent has been investigated by researchers until now.2 C. Kahraman. (ii) computational fault tolerance. adaptation. and incomplete information. It is a necessity to answer the question what the “computational intelligence” is. ¸ not demonstrate adequately suitable behavior and it is eliminated stochastically. evolution and fuzzy logic (rough sets) to create systems that is. A good definition of the field is therefore impossible. including swarm intelligence. The approach taken by the journals and by the book authors is to treat CI as an umbrella under which more and more methods will be added. and combinations thereof. But intelligence involves more than being restricted to biological organisms. and (iv) error rates that approximate human performance”. is a novel technology to bring intelligence into computation. . CI. fuzzy systems and evolutionary computation. . has a pattern recognition component. Bezdek says “. because different people include or exclude different methods under the same CI heading. does not use knowledge in the AI sense. and so on. ˙ Kaya.

computational intelligence combines artificial neural networks.. . While some techniques within CI are often counted as artificial intelligence techniques (e. where the structure of models. and biological intelligence according to the level of complexity as shown in Figure 1. solutions. “Computational intelligence is the field of studying how to build intelligent agents” [14]. [. neural networks (NN) and evolutionary computation (EC) were integrated under the name of CI as a hybrid system [122]. “Computational Intelligence is a new paradigm of knowledge-based processing that dwells on three main pillars of granular computing. and Future 3 Eberhart et al. association. neural and fuzzy-neural networks. According to Bezdek [6]. Computational processes deal with numeric data from sensors. Artificial intelligence is based on symbol processing techniques and augments computational processes with rules and other nonnumerical knowledge tidbits. Today. wetware. In the face of these difficulties fuzzy logic (FL). artificial intelligence. genetic algorithms. in natural or artificial systems” [102].Computational Intelligence: Past.g. and abstraction. etc.” “Computational intelligence is the study of the design of intelligent agents. An artificial neural network is a branch of CI that is closely related to machine learning. evolutionary computing. discovery. such as generalisation. . As such. is imposed.] The central scientific goal of computational intelligence is to understand the principles that make intelligent behavior possible. In general. Rule-based reasoning and case-based reasoning are in the style of artificial intelligence. connectionist systems and cybernetics. self-organization map. “Computational intelligence is the study of adaptive mechanisms to enable or facilitate intelligent behavior in complex and changing environments. intelligence can be subdivided into CI. CI techniques are generally bottom-up.1. [29] defined CI as follows: “Computational intelligence is defined as a methodology involving computing (whether with a computer. Neural networks. or neural networks) there is a clear difference between these techniques and traditional logic based artificial intelligence techniques. and evolutionary computing” [107]. . etc.) that exhibits an ability to learn and/or deal with new situations such that the system is perceived to possess one or more attributes of reason. where order and structure emerges from an unstructured beginning. typical artificial intelligence techniques are top-to-bottom. CI is further closely associated with soft computing. Fuzzy logic is the most representative example of integrating numeric and semantic information. genetic algorithms are examples of CI. swarm intelligence and fuzzy systems” [31].

” According to Marks [70]. Another definition is making computer programs for doing intelligent things. its components should have the following characteristics: considerable potential in solving real world problems. ability to learn from experience. and ability of adapting in response to dynamically changing conditions and constraints. fuzzy and evolutionary systems were brought together under the rubric of Computational Intelligence.4 C. a relatively new term offered to generally describe methods of computation that can be used to adapt solutions to new problems and do not rely on explicit human knowledge. . Because of the cloudy definition of intelligence. For both statements the questions difficult to response are occurred: what . genetic algorithms. Fogel [35] summarized CI as “. Cınar I. Artificial intelligence is defined generally as the study of intelligent behavior. and artificial life. different statements of both artificial intelligence and CI have been developed [25]. Kahraman. ˙ Kaya. fuzzy systems. CI consists of neural nets. and D. capability of self-organizing. These technologies of neural. .1 Structure of Intelligent Systems The IEEE Computational Intelligence Society categorizes CI into three broad subjects: • Neural Networks • Fuzzy Systems • Evolutionary computation. evolutionary programming. Irrespective of the way CI is defined above. ¸ Figure 1. which includes ◦ Evolutionary Algorithms ◦ Swarm Intelligence On the other hand.

combination of these techniques. 1.7. A signal transport from axon to dendrites and passing through other neurons by synapses [37]. Neural networks. and artificial life are the basic terms of CI. issues of architecture and integration emerge as central. ANN is used effectively for pattern recognition and regression. CI involves many methods and there is no consensus about methods under the same CI heading.2 Neural Networks Artificial neural networks (ANN) are optimization algorithms which are recently being used in variety of applications with great success. CI is used on problems that are difficult to solve using artificial systems. and 1.2 contains a schematic diagram of real neuron. respectively. axon and dendrites. CI’s future and conclusions are discussed in Section 1. genetic algorithms.4. intelligent agents and probabilistic reasoning models. including swarm intelligence are the subjects which are subjects of interest of IEEE Computational Intelligence Society. For this aim. fuzzy logic and evolutionary computing are explained in Sections 1. CI includes tools such as fuzzy logic. In recent years. Also hybrid systems. Brain is formed by neurons and Figure 1. Humans are capable of solving these problems so they are requiring intelligence [25]. experts prefer ANN over classical statistical methods as a forecasting model. fuzzy systems. A perceptron consists of node as cell body and connections as synaptic links [15]. are defined in Section 1. ANN consist of many nodes and connecting synapses. established already in the mid 1950s. evolutionary programming. The literature review and trends of CI in recent years are analyzed in Section 1.3. and Future 5 are the intelligent behavior and intelligent things? [11] Artificial Intelligence (AI) which was the first large scientific community.Computational Intelligence: Past. 1. The main parts of a neuron are cell body. ANN models take inspiration from the basic framework of the brain.2. fuzzy systems and evolutionary computation. Figure 1. The international CI journals and the characteristics of these journals are examined in Section 1. Extrapolating from historical data to generate forecasts. evolutionary computing. Nodes operate in parallel and communicate with each other through connecting synapses. According to the books and journals in literature. neural networks. Today.3 shows a representation of artificial neuron called perceptron. According to Marks [70] neural nets.8.6. They are worked on comprehensive toolbox. solving the .5. This increasing interest can be explained by some basic properties of ANN. This chapter aims to analyze the past and present of CI and make some evaluations for its future. artificial neural networks.

3 A simple perceptron complex nonlinear problems successively and high computation rate are features that are the reasons for many experts to prefer ANN.2 A schematic diagram of real neuron [15] Figure 1. Each layer has a certain number of neurons which are the basic processing elements of ANN.2. The neurons connect from one layer to the next layer. and D. In feedforward neural networks signals flow from the input layer to the output layer. ˙ Kaya. Kahraman. Neurons are connected with the other neurons in further layers and each connection has an associated weight. Cınar I. 1.1 Network Architectures Neural networks can be classified according to their architectures as single layer feedforward networks. there is no requirement for any assumptions in ANN [88]. Moreover. multilayer feed-forward networks and recurrent neural network. not to the previous layer .6 C. ¸ Figure 1.

so if the output exceed the threshold value the value of the function will be 1. Activation function defines the output of a neuron. x 0 f (x) = ⎩0. In Figure 1.4 A sample single layer feed-forward neural network or within the same layer. some hidden layers and an output layer.4. ⎧ ⎨1. is an operation unit in ANN.6(a). and Future 7 Figure 1.6-A the threshold value is 0 and this is the most widely used one. Then the weighted sum is passed through a function which is called activation function and this activated value is the output of the neuron. so input layer is not counted when the architecture is defined [140]. In Figure 1. if the output is 0 the value of the function will be 0 and if the output is lower than the threshold value the value of the function will be −1 [140]. Today. There are three basic activation function types [52]: (i) Threshold Function: This type of function is shown in Figure 1. The process starts with summation of weighted activation of other neurons through its incoming connections. (ii) Piecewise-Linear Function: Piecewise-linear function is shown in Figure 1. Each neuron.6-B and . there is a sample illustration of single layer feedforward neural network architecture with four input and four output nodes. Generally if the signal coming through the previous layer is bigger than the threshold value. A sample multilayer feed-forward neural network is shown in Figure 1.Computational Intelligence: Past. the output value of threshold function is 1. otherwise 0. except the neurons in input layer. The network is composed of an input layer. Only the output layer consists of computation nodes.5. x < 0 (1) Some threshold functions can take the value of −1 or 1.

¸ defined as f (x) = ⎧ ⎪1. ⎪ ⎪ ⎪ ⎪ ⎨ x − 1 2 + 1 2 (2) 1 1 x x+ . Sigmoid functions are widely used activation functions since it is easy to derivative (Figure 1. Cınar I. ˙ Kaya. Kahraman. and D.5 A sample multilayer feed-forward neural network [52] Figure 1.8 C.6-C). (iii) Sigmoid Function: Because of the learning process in neural networks.6 Basic activation function types . nonlinear and differentiable activation functions are used as activation functions. As the slope parameter approaches infinity the function becomes threshold function. x − 2 This function can be explained as an approximation of a nonlinear function [52]. An example of sigmoid functions called logistic function is defined as f (x) = 1 1 + e− α x (3) where α is the slope parameter of sigmoid function. − ⎪ 2 2 ⎪ ⎪ ⎪ ⎪ 1 ⎩0. Figure 1.

wi j is the synaptic weight of neuron i to neuron j.7.Computational Intelligence: Past. Training is another name of seeking the correct weight values. 1. While classical statistical techniques only estimate the coefficient of independent variables.2. 1. Recurrent neural networks can be commented as subsume of single layer and multilayer networks [103].2. ANN selects proper weights during training and keeps them for further use to predict the output [15].2 Learning processes The most important section of ANN is training. Among the several supervised learning algorithms.1 Supervised learning Supervised learning is used for regression and classification problems which concerns with the relationships between inputs and output values. In recurrent networks. This structure provides a short time memory and is used for remembering the past [1].2. and Future 9 Figure 1. unsupervised and reinforcement learning. neurons have self connections and/or connections to nodes in the previous layers besides the forward connections between layers. backpropagation algorithm is the most suitable method for training multilayer feed-forward . and θ j is the bias which is the constant value of the activation function. Today. Applications in literature have shown that neural networks with one hidden layer adequately learn the continuous functions. Second hidden layer can be used for continuous functions which have some break points that inhibit the continuity [140].7 Recurrent network The net input of neuron j is net j = ∑ wi j xi + θ j i (4) where xi is the output of the neuron i in the previous layer. Three types of learning processes are supervised. The last type of network architecture is recurrent networks shown in Figure 1.

10

C. Kahraman, ˙ Kaya, and D. Cınar I. ¸

networks [145]. While training the ANN, the weights are differentiated among neurons. In the learning phase, an input is presented to the network along with the desired output (y p ) and the weights are adjusted to produce the desired output. The most popular learning algorithm is the back-propagation (BP) algorithm. BP is a gradient descent algorithm which improves the resulting performance of the ANN by reducing the total error by adjusting the weights along its gradient. Root Mean Square Error (RMSE) is the most widely used error value, which is calculated as E= 1 (yk − ok )2 2∑ k (5)

where ok is the output vector and k is the index for output units [52]. During backpropagation learning, weights are modified according to their contribution to the error function. For two layer feed-forward neural networks (one hidden layer and one output layer), the change of weights which are between neurons in hidden layer and ones in output layer can be found by least squares method and formulated as following: ∂E Δvh = −η = η ∑ rt − yt zth ∂ vh t

(6)

where t is the index of sample, vh is the value of weight between output and the hth hidden neuron, yt is the output value found by neural network, rt is desired output value, zth is the input value of hth hidden neuron and η is the learning rate which determines the magnitude of changes to be made in the parameter [52]. (rt − yt ) is the error between desired and output value and it is multiplied by zth since the change of weight depends on the magnitude of input. So, for an input with small value an unnecessary training process will not be applied [1]. While the weights between input and hidden neurons are updated, least squares method is not applied. Because the error term resulted by hidden neuron cannot be found. The change of these weights can be found by chain rule as follows: ∂E Δwh j = −η ∂ wh j ∂ E t ∂ yt ∂ zth = −η ∑ t t t ∂ y ∂ zh ∂ wh j = −η ∑ − rt − yt vh zth 1 − zth xtj
t

= η ∑ rt − yt vh zth 1 − zth xtj
t

(7)

where (rt − yt ) vh can be thought as error term of hth hidden neurons and it is backpropagated to the hidden neurons. (rt − yt ) is the error term of output and is weighted with

Computational Intelligence: Past, Today, and Future

11

the weight of hidden neuron vh . zth 1 − zth is derivative of sigmoid function and xtj is the derivative of weighted sum [1]. The performance can be analyzed with RMSE, however RMSE can be quite high or low depending on the unit of variables. In order to calculate the performance of different forecasting models, relative error (RE) can be used: RE = ∑(yk − ok )2
k

∑ y2 k
k

(8)

Although BP is a simple and popular algorithm, there are several disadvantages. It is slow and needs much iteration to solve a simple problem [109]. An important disadvantage of BP is local convergence [126]. Solutions found by BP are dependent on the initially random weights [41]. So in order to find global optimum, BP is run with different random initial weights. Another drawback of BP is determining the parameters (number of hidden layers, number of hidden neurons, learning rate etc.) for finding the global optimum [145]. Many different combinations of the parameters are tried to generate a confidence that a global solution has been found. In the literature, most of the ANN applications use trial-and-error method to find appropriate parameters [126, 145]. 1.2.2.2 Unsupervised learning Another learning type is unsupervised learning. In supervised learning the values of inputs and outputs are known and the aim is seeking the structure of between inputs and outputs. In unsupervised learning only the input values are known and the aim is finding the pattern of inputs. Input space has a pattern and regular structures can be investigated [1]. After the pattern of input is defined, training procedure is completed and the classes of new input values can be determined. Clustering problems are the examples of unsupervised learning. The aim of cluster analysis is determining the groups in data set or grouping the data. Such as, a firm sustaining demographic features and arrival frequencies of its customers can cluster the customers within different groups and define customer profiles [1]. 1.2.2.3 Reinforcement learning In the last type of learning process, reinforcement learning, signals, which are produced through a sequence of decisions and provide some measure of performance, are the sources of learning. Although the reinforcement learning provides developing profitable decision policies with minimal domain information, it generally requires a large number of training

12

C. Kahraman, ˙ Kaya, and D. Cınar I. ¸

episodes and extensive CPU time [73]. Games are the examples of reinforcement learning. A single movement has not a meaning for a player. The sequence of movement is desired to give the best result. Chess is a popular game for reinforcement learning [1]. 1.3 Fuzzy Set Theory Zadeh published the first paper, called “fuzzy sets”, on the theory of fuzzy logic (1965). He characterized non-probabilistic uncertainties and provided a methodology, fuzzy set theory, for representing and computing data and information that are uncertain and imprecise. Zadeh [149] defined the main contribution of fuzzy logic as “a methodology for computing with words” and pointed out two major necessities for computing with words: “First, computing with words is a necessity when the available information is too imprecise to justify the use of numbers, and second, when there is a tolerance for imprecision which can be exploited to achieve tractability, robustness, low solution cost, and better rapport with reality.” While the complexity arise and the precision is receded, linguistic variables has been used to modeling. Linguistic variables are variables which are defined by words or sentences instead of numbers [148]. Many problems in real world deal with uncertain and imprecise data so conventional approaches cannot be effective to find the best solution. To cope with this uncertainty, fuzzy set theory has been developed as an effective mathematical algebra under vague environment. Although humans have comparatively efficient in qualitative forecasting, they are unsuccessful in making quantitative predictions. Since fuzzy linguistic models permit the translation of verbal expressions into numerical ones, thereby dealing quantitatively with imprecision in the expression of the importance of each criterion, some methods based on fuzzy relations are used. When the system involves human subjectivity, fuzzy algebra provides a mathematical framework for integrating imprecision and vagueness into the models [57]. Uncertainties can be reflected in mathematical background by fuzzy sets. Fuzzy sets have reasonable differences with crisp (classical) sets. Crisp set A in a universe U can be defined by listing all of its elements denoted x. Alternatively, zero-one membership function, μA (x), which is given below can be used to define x. ⎧ ⎨1, x ∈ A μA (x) = ⎩0, x ∈ A /

(9)

Unlike crisp sets, a fuzzy set A in the universe of U is defined by a membership function

Computational Intelligence: Past, Today, and Future

13

μA (x) which takes on values in the interval [0, 1]. The definition of a fuzzy set is the extended version of a crisp set. While the membership function can take the value of 0 or 1 in crisp sets, it takes a value in interval [0, 1] in fuzzy sets. A fuzzy set, A, is completely characterized by the set of ordered pairs [152]: A = (x, μA (x)) | x ∈ X membership function of A satisfies the inequality (10)

A fuzzy set A in X is convex if and only if for every pair of point x1 and x2 in X, the

μA δ x1 + (1 − δ )x2
where δ ∈ [0, 1] [49].

min μA x1 , μA x2

(11)

In fuzzy logic, basic sets operations, union, intersection and complement, are defined in terms of their membership functions. Let μA (x) and μB (x) be the membership functions of fuzzy sets A and B. Fuzzy union has the membership function of

μA∪B (x) = max μA (x), μB (x)
and fuzzy intersection has the membership function of

(12)

μA∩B (x) = min μA (x), μB (x)
and fuzzy complement has the membership function of

(13)

μ (x) = 1 − μA(x)
A

(14)

Most real world applications, especially in engineering, consist of real and nonnegative elements such as the Richter magnitudes of an earthquake [110]. However there are some applications that some strict numbers cannot be defined for given situation on account of uncertainty and ambiguity. Fuzzy set theory provides a representing of uncertainties. Zadeh [147] mentioned about fuzzy sets as: “The Notion of a fuzzy set provides a convenient point of departure for the construction of a conceptual framework which parallels in many respects the framework used in the case of ordinary sets, but is more general than the latter and, potentially, may prove to have a much wider scope of applicability, particularly in the fields of pattern classification and information processing.” There is a huge amount of literature on fuzzy logic and fuzzy set theory. In recent studies, fuzzy set theory has been concerned with engineering applications. Certain types of uncertainties are encountered in a variety of areas and fuzzy set theory has pointed out to be very efficient to consider these [57]. Automatic control, consumer electronics, signal processing, time-series prediction, information retrieval, database management, computer vision, data classification, and decision-making are some areas in which fuzzy logic and fuzzy set theory are applied [50].

14

C. Kahraman, ˙ Kaya, and D. Cınar I. ¸

1.4 Evolutionary Computation Evolutionary computation concept based on Darwin’s evolution theory by applying the biological principle of natural evolution to artificial systems for the solution of optimization problems has received significant attention during the last two decade, although the origins were introduced the late 1950’s with works of Bremermann in 1962, Friedberg in 1958 and 1959, and Box in 1957, and others [3]. The domain of evolutionary computation involves the study of the foundations and the applications of computational techniques based on the principles of natural evolution. Evolutionary algorithms employ this powerful design philosophy to find solutions to hard problems from different domains, including optimization, automatic programming, circuit design, machine learning, economics, ecology, and population genetics, to mention but a few. Evolutionary computation for solving optimization and other problems has considerable advantages. The first and important advantage is adaptability for fluxional situations. Unlike many traditional optimization procedures where the calculation must be restarted from the beginning if any variable in the problem changes, evolutionary computation does not need to restart from beginning since the current population serves as a memoir of stored knowledge that can be applied on the fly to a dynamic environment. Therefore, traditional optimization methods are more computationally expensive than the evolutionary computation. Another advantage of an evolutionary computation to problem solving comes in being able to generate good enough solutions quickly enough for them to be of use. There are several techniques of evolutionary computations, among which the best known ones are genetic algorithms, genetic programming, evolution strategies, classifier systems and evolutionary programming; though different in the specifics they are all based on the same general principles [3, 20, 100]. An evolutionary computation (algorithm) begins by initializing a population of candidate solutions to a problem. New solutions are then created by randomly varying those of the initial population. This population is then selected in a guided random search using parallel processing to achieve the desired end.

1.4.1 Genetic algorithms Genetic algorithms (GAs) which was first developed by Holland in 1975 are based on mechanics of natural selection and genetics to search through decision space for optimal solutions. The metaphor underlying GAs is natural selection. Genetic algorithm is a search

Computational Intelligence: Past, Today, and Future

15

technique used in computing to find exact or approximate solutions to optimization and search problems. GAs can be categorized as global search heuristics. GAs are a particular class of evolutionary algorithms (also known as evolutionary computation) that use techniques inspired by evolutionary biology such as inheritance, mutation, selection, and crossover (also called recombination). GAs are stochastic search methods based on the genetic process of biological organisms. Unlike conventional optimization methods, GAs maintain a set of potential solutions (populations) in each generation. A GA is encoding the factors of a problem by chromosomes, where each gene represents a feature of problem. In evolution, the problem that each species faces is to search for beneficial adaptations to the complicated and changing environment. In other words, each species has to change its chromosome combination to survive in the living world. In GA, a string represents a set of decisions (chromosome combination), that is a potential solution to a problem. Each string is evaluated on its performance with respect to the fitness function (objective function). The ones with better performance (fitness value) are more likely to survive than the ones with worse performance. Then the genetic information is exchanged between strings by crossover and perturbed by mutation. The result is a new generation with (usually) better survival abilities. This process is repeated until the strings in the new generation are identical, or certain termination conditions are met. A generic flow of GA is given in Figure 1.8. This algorithm is continued until the stopping criterion is reached. GAs are different from other search procedures in the following ways: (1) GAs consider many points in the search space simultaneously, rather than a single point; (2) GAs work directly with strings of characters representing the parameter set, not the parameters themselves; (3) GAs use probabilistic rules to guide their search, not deterministic rules. Because GAs consider many points in the search space simultaneously there is a reduced chance of converging to local optima. In a conventional search, based on a decision rule, a single point is considered and that is unreliable in multimodal space. GAs consist of four main sections that are explained in the following subsections: Encoding, Selection, Reproduction, and Termination [58, 32, 55, 56].

1.4.1.1 Encoding While using GAs, encoding a solution of a problem into a chromosome is very important. Various encoding methods have been created for particular problems to provide effective implementation of GAs for the last 10 years. According to what kind of symbol is used as the alleles of a gene, the encoding methods can be classified as follows:

16

C. Kahraman, ˙ Kaya, and D. Cınar I. ¸

Figure 1.8

The fundamental cycle and operations of basic GAs [38]

• Binary encoding, • Real number encoding, • Integer or literal permutation encoding, • General data structure encoding. 1.4.1.2 Selection During each successive generation, a proportion of the existing population is selected to breed a new generation. Individual solutions are selected through a fitness-based process, where fitter solutions (as measured by a fitness function) are typically more likely to be selected. Certain selection methods rate the fitness of each solution and preferentially se-

Computational Intelligence: Past. . Generally. Popular and wellstudied selection methods include roulette wheel selection and tournament selection. as this process may be very time-consuming. for reasons already mentioned above. Other methods rate only a random sample of the population. thus slowing or even stopping evolution.1. the average fitness will have increased by this procedure for the population.1. It is always problem dependent.3 Reproduction The next step of GAs is to generate a second generation population of solutions from those selected through genetic operators: crossover (also called recombination). 1. New parents are selected for each child. 1. This helps keep the diversity of the population large. Most functions are stochastic and designed so that a small proportion of less fit solutions are selected. and the process continues until a new population of solutions of appropriate size is generated. crossover is a genetic operator used to vary the programming of a chromosome or chromosomes from one generation to the next. along with a small proportion of less fit solutions. since only the best organisms from the first generation are selected for breeding. By producing a “child” solution using the crossover and mutation. The purpose of the mutation in GAs is to allow the algorithm to avoid local minima by preventing the population of chromosomes from becoming too similar to each other. It is defined over the genetic representation and measures the quality of the represented solution. These processes ultimately result in the next generation population of chromosomes that is different from the initial generation. a new solution is created which typically shares many of the characteristics of its “parents”. upon which genetic algorithms are based.4 Crossover In GAs. It is an analogy to reproduction and biological crossover. preventing premature convergence on poor solutions. Today.4.1.4. 1.4. and/or mutation. and Future 17 lect the best solutions. a pair of “parent” solutions is selected for breeding from the pool selected previously.5 Mutation Operator The premature convergence of a new generation can be prevented by using the mutation operator. For each new solution to be produced. The fitness function is the important section of selection.

instead of encoding possible solutions to a problem as a fixed-length character string. Create a new population by applying the following operations: i) Copy existing individuals to the new population. and D. The basic steps in constructing a genetic program are as follows [80]: 1. . 1.1. • Allocated budget (computation time/money) reached.4. Common terminating conditions are as follows: • A solution that satisfies minimum criteria is found. It is defined by Koza [61] as “. It is a specialization of genetic algorithms where each individual is a computer program. 2. ˙ Kaya. Therefore. ii) Randomly select a pair of existing trees and recombine subtrees from them to produce a new tree. The operations of reproduction and recombination are carried out with the probability . Kahraman. Cınar I. they are encoded as computer programs. • Combinations of the above. ¸ 1. Calculate the fitness of each tree in the initial population according to a suitable criterion. 3. Create an initial randomly generated population of trees.2 Genetic programming Genetic programming (GP) developed by Koza [61. The individuals in the population are programs that — when executed — are the candidate solutions to the problem. populations of computer programs are genetically bred using the Darwinian principle of survival of the fittest and using a genetic crossover (recombination) operator appropriate for genetically mating computer programs.6 Termination This generational process is repeated until a termination condition has been reached. 62] is an evolutionary algorithm-based methodology inspired by biological evolution to find computer programs that perform a user-defined task. • The highest ranking solution’s fitness such that successive iterations no longer produce better results. .18 C.” In GP.4. • Fixed number of generations reached. it is a machine learning technique used to optimize a population of computer programs according to a fitness landscape determined by a program’s ability to perform a given computational task. genetic programming paradigm.

4. However.9. Repeat these operations.Computational Intelligence: Past. evolutionary programming first generates offspring and then selects the next generation. and Future 19 of selection for the operations skewed toward selecting individuals with higher levels of fitness. and (ii) select the next generation from the mutated and the current solutions. It is becoming harder to distinguish it from evolutionary strategies. except that the program structure is fixed and its numerical parameters are allowed to evolve.3 Evolution strategies Evolution strategies use natural problem-dependent representations. 5. Some of its original variants are quite similar to the later genetic programming. Mutation is the major genetic operator in evolution strategies. where mutation is used to generate new solutions (offspring) . and primarily mutation and selection as search operators were introduced by Ingo Rechenberg and Hans Paul Schweffel in the 1960s [100] as a method for solving parameter-optimization problems. it is desirable to develop an algorithm that can handle a large population efficiently [43]. It also plays the role of a reproduction operator given that the mutated individual is viewed as an offspring for the selection operator to work on. It has been applied with success to many numerical and combinatorial optimization problems in recent years. 4. These two steps can be regarded as a population-based version of the classical generate-and-test method. Like with evolution strategies.4 Evolutionary programming Evolutionary programming (EP) was first used by Fogel in 1960s in order to use simulated evolution as a learning process aiming to generate artificial intelligence. Optimization by EP can be summarized into two major steps as (i) mutate the solutions in the current population. Today. 1. The schematic diagram of the EP algorithm is depicted in Figure 1.4. when large numbers of processors are available. the advantage of using evolution strategies in parallel cannot be fully exploited. Fogel used finite state machines as predictors and evolved them. when the number of offspring is smaller than the number of available processors. One of the commonly proposed advantages of evolution strategies (ESs) is that they can be easily parallelized. Consequently. ESs with λ offspring per generation (population size λ) are usually parallelized by distributing the function evaluation for each of the offspring on a different processor. 1. keeping a record of the overall fittest individual. Calculate the fitness of each individual in the new population.

Kahraman. Cınar I. rather than a ‘pure’ evolutionary algorithm. called classifiers.20 C. They can be thought of as restricted versions of classical rule-based systems. CSs exploit evolutionary computation and reinforcement learning to develop a set of condition-action rules (i. GAs. such as evolution strategies. that represents the current knowledge of the system. and (3) the GAs. and D. presented by Holland in 1970’s. which distributes the reward received from the environment to the classifiers accountable for the rewards ob- . all CS models. Formulating EP as a special case of the generate-and-test method establishes a bridge between EP and other search algorithms.4. ˙ Kaya. comprise four main components: (i) a finite population of condition action rules. (iii) the reinforcement component. which adapt the system’s knowledge by rule discovery [100]. (ii) the performance component. 1. with the addition of input and output interfaces. which adapts the behavior by credit assignment. are evolution-based learning systems.e. (2) the apportionment of credit system. tabu search. There are many models of CSs and therefore also many ways of defining what a learning classifier system is.5 Classifier systems Classifier systems (CSs).9 Schematic diagram of the evolutionary programming algorithm [144] and selection is used to test which of the newly generated solutions should survive to the next generation. more or less. which performs the inference and defines the behavior of the whole system. which governs the interaction with the environment. simulated annealing.. and others. and thus facilitates cross-fertilization among different research areas [146]. ¸ Figure 1. the classifiers) which represent a target task that the system has learned from on-line experience. A classifier system consists of three main components: (1) the rule and message system. Nevertheless.

The working principles of ACO based on a real ant colony can be explained as follows [22]: Real ants are capable of finding the shortest path from a food source to the nest without using visual cues.g. Consider Figure 1. ant colony optimization (ACO) which is a class of optimization algorithms modeled on the actions of a real ant colony was introduced by Dorigo and colleagues as a novel nature-inspired metaheuristic for the solution of hard combinatorial optimization problems. they are capable of adapting to changes in the environment.10-A: ants are moving on a straight line that connects a food source to their nest.Computational Intelligence: Past. Also. It is well known that the primary means for ants to form and maintain the line is a pheromone trail. the shorter path around the obstacle will more rapidly reconstitute the interrupted pheromone trail compared to those who choose the longer path. by chance. It is interesting to note that those ants which choose. which is responsible for discovering better rules and improving existing ones through a GA [45]. (iv) the discovery component. This elementary behaviour of real ants can be used to explain how they can find the shortest path that reconnects a broken line after the sudden appearance of an unexpected obstacle has interrupted the initial path (Figure 1. once the obstacle has appeared. 1. The most interesting aspect of this autocatalytic process is that finding the shortest path around the obstacle seems to be an emergent property of the interaction between the obstacle shape and ants distributed behaviour: although all ants move at approximately the same speed and deposit a pheromone trail at approximately the same rate. A very similar situation can be found on the other side of the obstacle (Figure 1.10-B). In fact. finding a new shortest path once the old one is no longer feasible due to a new obstacle. and Future 21 tained.10-D). the shorter path will receive a greater amount of pheromone per time unit and in turn a larger number of ants will choose the shorter path. Today. Thus. The inspiring source of ACO is the foraging behavior of real ants. those ants which are just in front of the obstacle cannot continue to follow the pheromone trail and therefore they have to choose between turning right or left.10-C). it is a fact that it takes longer to contour obstacles on their longer side than on their shorter side which makes the pheromone . Due to this positive feedback (autocatalytic) process. all the ants will rapidly choose the shorter path (Figure 1.6 Ant colony optimization In the early 1990s. and each ant probabilistically prefers to follow a direction rich in pheromone. e. In this situation we can expect half the ants to choose to turn right and the other half to turn left. Ants deposit a certain amount of pheromone while walking.4.

22

C. Kahraman, ˙ Kaya, and D. Cınar I. ¸

trail accumulate quicker on the shorter side. It is the ants’ preference for higher pheromone trail levels which makes this accumulation still quicker on the shortest path.

Figure 1.10 Real Ant Colony Movements from Nest to Food [22].

In general, the ACO approach attempts to solve an optimization problem by repeating the following two steps [21]: • candidate solutions are constructed using a pheromone model, that is, a parameterized probability distribution over the solution space; • the candidate solutions are used to modify the pheromone values in a way that is deemed to bias future sampling toward high quality solutions. Dorigo and Gambardella [22] applied this philosophy to solve travelling salesman problem. After the initial proof-of-concept application to the traveling salesman problem (TSP),

Computational Intelligence: Past, Today, and Future

23

ACO has been applied to many other optimization problems such as assignment problems, scheduling problems, and vehicle routing problems. Recently, researchers have been dealing with the relation of ACO algorithms to other methods for learning and optimization. 1.4.7 Particle swarm optimization Particle swarm optimization (PSO) which is population based stochastic optimization technique developed by Kennedy and Eberhart [59], inspired by social behavior of bird flocking or fish schooling is a global optimization algorithm for dealing with problems in which a best solution can be represented as a point or surface in an n-dimensional space. PSO shares many similarities with evolutionary computation techniques such as Genetic Algorithms (GA). The system is initialized with a population of random solutions and searches for optima by updating generations. However, unlike GA, PSO has no evolution operators such as crossover and mutation. In PSO, the potential solutions, called particles, fly through the problem space by following the current optimum particles. Each individual of the population, called a ‘particle’, flies around in a multidimensional search space looking for the optimal solution. Particles, then, may adjust their position according to their own and their neighboring-particles experience, moving toward their best position or their neighbor’s best position. In order to achieve this, a particle keeps previously reached ‘best’ positions in a cognitive memory. PSO performance is measured according to a predefined fitness function. Balancing between global and local exploration abilities of the flying particles could be achieved through user-defined parameters. PSO has many advantages over other heuristic techniques such that it can be implemented in a few lines of computer code, it requires only primitive mathematical operators, and it has great capability of escaping local optima. In a PSO system, multiple candidate solutions coexist and collaborate simultaneously. Each solution candidate, called a ‘particle’, flies in the problem search space (similar to the search process for food of a bird swarm) looking for the optimal position to land. A particle, as time passes through his quest, adjusts its position according to its own ‘experience’, as well as according to the experience of neighboring particles. Tracking and memorizing the best position encountered build particle’s experience. For that reason, the PSO algorithm possesses a memory (i.e. every particle remembers the best position it reached during the past). PSO system combines local search methods (through self experience) with global search methods (through neighboring experience), attempting to balance exploration and exploitation. Two factors characterize a particle status on the search space: its position and velocity [123]. Particle swarm optimization has been successfully applied in many research and application

24

C. Kahraman, ˙ Kaya, and D. Cınar I. ¸

areas in recent years. It is demonstrated that PSO gets better results in a faster, cheaper way compared with other methods. Another reason that PSO is attractive is that there are few parameters to adjust.

1.5 Hybrid Systems Solving complex problems, such as financial investment planning, foreign exchange trading, and knowledge discovery from large/multiple databases, involves many different components or sub-tasks, each of which requires different types of processing. To solve such complex problems, a great diversity of intelligent techniques, including traditional hard computing techniques (e.g., expert systems) and CI techniques (e.g., fuzzy logic, NN, and GAs), are required. For example, in financial investment planning, NN can be used as a pattern watcher for the stock market; GAs can be used to predict interest rates; and approximate reasoning based on fuzzy logic can be used to evaluate financial risk tolerance ability of clients. These techniques are complementary rather than competitive, and thus must be used in combination and not exclusively. These systems are called hybrid intelligent systems [150]. Although the sufficient results found by CI techniques, more effective solutions can be obtained when used combination of these techniques. Each combination has an aim to decrease the limitation of one method. For example, genetic algorithm has been used to improve the performance of ANN in literature. Usage of genetic algorithms in ANN for training or finding the appropriate architecture can keep from getting trapped at local minima. Another example is using fuzzy inference with other CI techniques. A fuzzy inference system can take linguistic information (linguistic rules) from human experts and also adapt itself using numerical data to achieve better performance. Generally, using hybrid systems provides synergy to the resulting system in the advantages of the constituent techniques and avoids their shortcomings [50]. The reader can find more details about intelligent hybrid systems, fuzzy logic and neural networks in Martin and Ralescu [71], Ruan [118], Song et al. [127], Zhang and Zhang [150], Kahraman [51], and Munakata [75].

1.6 Literature Review In this section a literature review on CI has been realized. In Table 1.1, the results of this review are given. Here, the papers whose title including “computational intelligence” is listed for the years 1992 to 2008. In the first column, these papers are classified as roughly based on their CI technique(s) which is (are) used in the paper.

Computational Intelligence: Past, Today, and Future

25

Table 1.1 CI Papers CI Techniques Authors Xue and Dong [143] Wang and Zou [137] Fuzzy logic Zimmermann [153] Nucci et al. [82], Papageorgioua et al. [89] Gu et al. [39] Karray et al. [54] Fuzzy logic, Genetic algorithms Sebastian and Schleiffer [125] Naso and Turchiano [78] Naso and Turchiano [78] Pedrycz and Reformat [99] Corne et al. [17], Park et al. [90], Park and Oh [92] Liang and Chunguang [67] Tzafestas and Tzafestas [132] Pal and Pal [87] Douligeris et al. [24] Fuzzy logic, Genetic algorithms, Neural networks Alvarado et al. [2] Oh et al. [84] Huang et al. [46] Oh et al. [83], Pedrycz and Reformat [99], Pedrycz [97], Park et al. [91], Wilson et al. [141] Karr [53], Castellano et al. [8] Riziotis and Vasilakos [108] Ruan [114] Stathakis and Athanassios [129] Sun and Kalenchuk [130] Dasey and Tzanakou Duch and Hayashi [27] Fuzzy logic, Neural networks Vasilakos et al. [134] Innocent et al. [47] Duch et al. [26] Lau et al. [65] Chen et al. [12] Ng and Leng [81] Genetic algorithms Lai et al. [64] Li [66] Wang et al. [136] Genetic algorithm, Neural networks Genetic programming, Neural networks Ling et al. [69] Nakayama et al. [76] Kubota [63] Narendra [77] Main Topics Conceptual Design Environmental Planning Health Conceptual Design Facility layout Planning Engineering Design Manufacturing (AGVs) Automated guided vehicles dispatching control Decision Trees Modeling Product packaging Forecasting Pattern Recognition Telecommunication Networks Petroleum Industry Design Conceptual Design Optimization Manufacturing Computer systems Photonics technology Sensory Evaluation Image Classification Conceptual Design Medicine Classification and prediction Network Research projects Data analysis Supply Chain Construction performance Conceptual Design Power systems Optimization Modeling Forecasting Optimization Control of robots Cement stabilization

26

C. Kahraman, ˙ Kaya, and D. Cınar I. ¸

Table 1.1 CI Techniques Authors

continued. Main Topics Conceptual Design Process discovery (data analysis) Optimization Inferential model development Nuclear Engineering Foreign investment Forecasting Earth sciences and Environment Online banking Export behavior Forecasting Energy management Sound transmission

Venugopal and Naredran [135], Chang and Tsai [10], Park et al. [93] Cass and DePietro [7] Guimar˜ es et al. [40] a Neural networks Warne et al. [138] Uhrig and Hines [133] Plikynas et al. [101] Pavlidis et al. [95] Cherkassky et al. [16] Quah and Sriganesh [104] Neural networks, Fuzzy logic Evolutionary algorithms, networks; Edwards et al. [30] Santos et al. [124] Matics and Krost [72] Neural Xu et al. [142]

Evolutionary computation, Genetic algorithms, Neural networks Evolutionary computation, Particle swarm optimization, Neural networks, granular Computing, Rule-based computing Neurocomputing, Evolutionary computing, Granular computing Neuro-fuzzy modelling General overview CI techniques are reviewed

Reformat et al. [107]

Software quality analysis

Pavlidis et al. [94]

Game Theory

Pedrycz [96]

Software engineering

Pedrycz and Peters [98]

Software engineering

Kothamasu et al. [60] Feigenbaum [33] Chen [13], Isasi et al. [48] Pal and Mudi [86] Ovaska et al. [85] Dote and Ovaska [23]

Hard turning CI Finance and Economic Decision making The fusion of hard computing and soft computing techniques Industrial applications

According to Table 1.1, the papers which include CI techniques have gained more attention in recent years. It should be noticed that two or more techniques (hybrid systems) are

Computational Intelligence: Past, Today, and Future

27

Figure 1.11

Distribution of articles whose title including CI.

integrated rather than using only one CI technique. In Table 1.2, the literature review results for the number of papers whose titles including “computational intelligence” are given for the years from 1994 to 2008. When a search was made with respect to paper’s title including “computational intelligence”, totally 146 articles were viewed. At the same time, these papers were classified based on their keywords. As it is seen that 44 of these papers include the keyword fuzzy set theory (FST), 69 of them include the keyword neural networks (NNs), 39 of them include the keyword genetic algorithms (GAs). In the other half of Table 1.2, the literature review has been extended to the papers whose keywords including “computational intelligence” for the years from 1994 to 2008. When a search was made for the papers whose keywords including “computational intelligence”, totally 185 articles were viewed. At the same time, these papers were classified based on their keywords. In this case, total number of papers whose keywords including FST, GAs, and NNs were determined as 69, 57, and 90, respectively. In Figure 1.11, a pie chart which summarizes the first half of Table 1.2 is given. According to Figure 1.11, CI has gained more attention year by year. 65 % of these papers have been published in the last 4 years. The papers related to CI are also classified based on their subject areas and the obtained results are summarized in Table 1.4. The frequency distribution of these papers is illustrated

28

C. Kahraman, ˙ Kaya, and D. Cınar I. ¸

Table 1.2 Search Result Based on the Publication Years Year 2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 Total T: CI 29 27 19 19 11 8 5 3 4 5 9 3 1 1 2 146 T: CI + K: FST 3 7 6 9 2 2 0 3 3 4 1 1 1 1 1 44 K: CI + FST 1 13 3 8 9 7 5 4 7 3 3 2 1 1 1 1 69 T: CI + K: GAs 5 4 9 8 3 1 0 1 2 3 2 0 1 0 0 39 K: CI + GAs 0 7 4 8 10 6 5 3 2 0 2 3 2 2 0 0 57 T: CI + K: NNs 8 12 14 11 3 3 1 3 2 4 3 2 1 1 1 69 K: CI + NNs 1 12 12 8 14 7 8 5 6 1 3 5 2 2 2 2 90 T: CI + K: EC 1 2 1 2 1 0 1 1 1 1 1 0 0 1 1 14 K: CI + EC 0 5 4 2 2 1 1 1 3 1 1 1 1 1 2 1 27 T: CI + K: PSO 0 0 1 1 0 0 0 0 0 0 0 0 0 0 0 2 K: CI + PSO 0 3 2 1 3 1 0 0 0 0 0 0 0 0 0 0 10 T: CI + K: ACO 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 2 K: CI + ACO 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 2

Year 2009 2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 Total

K: CI 3 37 28 19 26 13 10 8 11 5 4 7 7 3 2 2 185

in Figure 1.12 as a bar chart. It is clearly understood that the most of CI papers are related to “computer science” while “engineering” and “mathematics” have consecutive importance. The pie chart for the papers, whose title including “computational intelligence” is given in Figure 1.13. According to this figure, NNs are the most used CI technique with 41 % of the papers. The second and third preferences are FST and GAs with percentages 26 and 23, respectively.

These papers are also analyzed based on their publication years. Today.Computational Intelligence: Past.15. the largest increase in percentage is observed from 2004 (8 %) to 2005 (16 %) while the largest decrease in percentage is viewed from 2005 (16 %) to 2006 (9 %).15.14 and 1.6.1 Computational intelligence in complex decision systems CI techniques are required to solve complex problems. Management and Accounting Energy Environmental Science Social Sciences Multidisciplinary Medicine Chemistry Neuroscience Chemical Engineering Agricultural and Biological Sciences Health Professions Economics. the papers using FST have the highest percentage in 2005 with 20 %. 1. and Future 29 Table 1. the papers using NNs have the highest percentage in 2005 with 16 %. According to Figure 1.3 Classification Results Based on Subject Areas Computer Science Engineering Mathematics Decision Sciences Physics and Astronomy Biochemistry. According to Figure 1. and knowledge discovery from large/multiple databases. 1. They involve many different components or sub-tasks. In addition. such as financial investment planning. each of which requires different .14. foreign exchange trading. Genetics and Molecular Biology Materials Science Earth and Planetary Sciences Business. Also the largest increase in percentage is observed from 2004 (5 %) to 2005 (20 %) while the largest decrease in percentage is viewed from 2005 (20 %) to 2006 (14 %). Two pie charts illustrate the distribution of the papers with respect to the used technique in Figs. Econometrics and Finance 76 71 26 9 8 7 7 5 4 4 4 3 3 3 2 2 2 2 1 1 NNs and FST are the most used two CI techniques in the papers.

Cınar I. . For example. ANN can be used as a pattern watcher for the stock market.13 The distribution of CI techniques based on titles. expert systems. including operations research.30 C. ¸ Figure 1. system science/engineering. These techniques can be divided into two categories [150]: • traditional hard computing techniques. Kahraman. in financial investment planning. Many techniques have been developed for complex problem solving and decision making. ˙ Kaya. and D. types of processing. and approximate reasoning based on fuzzy logic can be used to evaluate financial risk tolerance ability of clients. GAs can be used to predict interest rates.12 Frequency distribution of subject area classification. and • CI techniques. Figure 1.

In this subsection a literature review on computational intelligence in complex decision systems is given briefly. and Future 31 Figure 1. Many international conferences are organized to gather the latest development on CI and they publish those works in their proceedings. Today.15 Distribution of the papers using FST based on their publication years.14 Distribution of the papers using NNs based on their publication years. In the following. construction and real time performance of a mobile monitoring system based on a Khepera robot. They implemented . some papers published in these proceedings are summarized: Benitez-Read and Rojas-Ramirez [4] presented the design. Figure 1.Computational Intelligence: Past.

autoencoder radial basis functions. . and studied their computational time complexity. Rao [106] suggested a new computationally intelligent approach based on the Discerning Nuclear Intentions Model to support policy decision-making in a complex. The methods analyzed were autoencoder multi-layer perception. [5] compared computational intelligence methods to analyze HIV in order to investigate which network was best suited for HIV classification. support vector machines and neuro-fuzzy models. Rubin et al. They also used one-dimensional image processing for the recognition of landmarks located at each test point. Reduction of manual human intervention in the decision process was achieved through the implementation of a neuro-fuzzy computational intelligence processor to automate the major decision making process of tasks distribution. non-linear manner. It compared the performance of three soft-computing models. Castillo and Melin [9] presented the development and design of two software tools which include a graphical user interface for construction. They also investigated the distributed network layout and system security measures. and D. [42] presented a new tissue classification system which combined both the classical fuzzy c-means classification algorithm and the human medical knowledge on geometric properties of different tissues and organs. Kahraman. Muhsin et al. The identification of test points was carried out by means of a Kohonen neural network. namely Multi-layer Perceptrons. Adaptive Neuro-Fuzzy Inference Systems and General Growing and Pruning Radial Basis Function. They designed a user friendly interface so that medical knowledge could be easily transformed into these data structure in an interactive way. [121] explored the interdependent roles of CI. Cınar I.32 C. ˙ Kaya. Betechuoh et al. [74] presented an effective development process for a distributed workflow management system which was utilized to automate the management processes related to computer and peripheral repair requests in the Faculty of Engineering and Technology (FET) at the University of Jordan. ¸ a cascade fuzzy control algorithm for the robot navigation. Fuzzy sensor fusion related to the perception of the environment was used to reduce the complexity of the navigation function. edition and observation of the intelligent systems for computational Intelligence. Han et al. Ssali and Marwala [128] introduced a novel paradigm to impute missing data that combined a decision tree with an auto associative neural network based model and a principal component analysis-neural network based model. Quah and Ng [105] presented methodologies for equities selection based on soft-computing models which focused on applying fundamental analysis for equities screening. Ferreira and Ruano [34] presented RBF neural networks in order to implement a model-based predictive control methodology for a research greenhouse. This system had been successfully applied to MRI images for classification of tissues of thigh.

Today. including the problem of finding good symbolic descriptors for continuous data. The advantages and disadvantages of different intelligent techniques were summarized. neuro fuzzy methods. and explained why agent perspectives were suitable for modeling. Instead of eigenvalue computation using the complete power system model. Weber and Wu [139] described architecture for designing CI that included a knowledge management framework. and Future 33 data mining. The reader can find full conference papers on CI applications in complex decision systems in Ruan et al. problem. The HIMS++ model took advantage of object-oriented modeling. designing and constructing hybrid intelligent systems. [111]–[120]. heuristic method. tradeoffs between accuracy and simplicity at the rule-extraction stage.Computational Intelligence: Past. Individual) 3 layers. [26] focused on the extraction and use of logical rules for data understanding. Liau and Schmitt-Landsiedel [68] described a diagnosis method that worked with industrial semiconductor ATE (Automatic Test Equipment) for analyzing the robustness of the circuit (e. Zhang and Lin [151] proposed a new research area called Computational Web Intelligence based on both CI and Web Technology to increase the Quality of Intelligence of e-Business. Duch et al. Teeuwsen et al. memory test chip) and used CI techniques. They also presented application system–agent-based hybrid intelligent system for data mining. Zhang and Zhang [150] discussed the importance of hybrid intelligent systems for complex problem solving and decision-making. and decision trees. and tradeoffs between rejection and error level at the rule optimization stage. [44] analyzed Vehicle Dispatching Problem for Cooperative Deliveries from Multiple Depots (VDP/CD/MD) problem in the transport industries. and fuzzy inference in (Atomic. All aspects of rule generation. They presented some basic concepts and existing knowledge on hybrid intelligent systems. Molecular.g. Ndousse and Okuda [79] proposed a computationally based expert system for managing fault propagation in internetworks using the concept of fuzzy cognitive maps. and those learned in external contexts. allowing the system to learn from its own experiences. so it could find a utility decision (vehicles plan) close to expert dispatcher. and machine learning. They proposed a hierarchical multiplex structure (HIMS++ ) calculation model to solve the VDP/CD/MD. Hirota et al. the new proposed OSA methods were based on computational intelligence such as neural networks. . [131] introduced newly developed methods for on-line oscillatory stability assessment in large interconnected power systems. and application were described. optimization.

which is including in particular symbolic multistrategy. fuzzy and evolutionary computation. statistical models. fuzzy systems. ¸ 1. The International Journal of Computational Intelligence and Applications. The journal is focusing on theories. ˙ Kaya. etc. neuro-pattern recognition. Covering a wide range of issues – from the tools and languages of AI to its philosophical implications – the journal provides a vigorous forum for the publication of both experimental and theoretical research.7 Computational Intelligence Journals This section includes the brief descriptions of the CI journals. pattern recognition.” The International Journal of Computational Intelligence that is a scholarly open access. fuzzy-evolutionary. The main goal of the journal is to provide the scientific community and industry with a vehicle whereby ideas using two or more conventional and CI based techniques could be discussed.34 C. and hybrid systems). The IJCIA welcomes works in areas such as neural networks. web intelligence and semantic web. modern knowledge-based systems. Areas include neural. interdisciplinary. statistical models. It focuses on topics about machine learning. image/audio/video compression and retrieval. encouraging “new ideas. image/audio/video compression and retrieval. evolutionary-symbolic. discovery science and knowledge mining. hybrid intelligent systems. symbolic machine learning. The journal is designed to meet the needs of a wide range of AI workers in academic and industrial research. Engineering and Technology) since 2005. neuro-evolutionary. combining two or more areas. has been published as quarterly since 1985 by Wiley-Blackwell Publishing. The journal dedicated to the theory and applications of CI (ANN. fuzzysymbolic. pattern recognition. as well as surveys and impact studies. has been published since 2001 by World Scientific. and cognitive learning. quarterly and fully refereed journal has been published by WASET (World Academy Science. peer-reviewed. The journal. hybrid intelligent systems.” The International Journal of Computational Intelligence Research that has been published by Research India Publications since 2005 is a free online journal. evolutionary-symbolic. methods and applications in CI. evolutionary computation. and D. fuzzyevolutionary. Kahraman. Key application areas of AI. neuro-symbolic. evolutionary computation. IJCIA. agents and multiagent systems. In the description of the journal it is clearly stated that: “This leading international journal promotes and stimulates research in the field of artificial intelligence (AI). neuro-symbolic. symbolic machine learning. Computational Intelligence: An International Journal. The aim and scope . such as neurofuzzy. Cınar I. fuzzy logic.

The journal is an open access journal interested in neural computing. fuzzy systems. the area is heterogeneous as being dwelled on such technologies as neurocomputing. multi-agent systems etc. Applied Computational Intelligence and Soft Computing. artificial intelligence and engineering. and . physicists. engineers. which becomes quite apparent in light of the current research pursuits. neural engineering and artificial intelligence. In a nutshell. psychologists. and artificial intelligence investigators among others can publish their work in one periodical that bridges the gap between neuroscience. The journal publishes original papers on foundations and new developments of CI with an emphasis on applications. where new theories with a sound biological understanding have been evolving.” The Journal of Computational Intelligence in Bioinformatics which publishes articles describing recent fundamental contributions in the field of artificial intelligence and CI theory and their applications in Bioinformatics has been published by Research India Publications since 2008 with three issues in a year. published by Hindawi Publishing in 2009. probabilistic reasoning. computational proteomics and systems biology. and mathematics. engineering. The current experimental systems have many of the characteristics of biological computers and are beginning to be built to perform a variety of tasks that are difficult or impossible to do with conventional computers. focuses on the disciplines of computer science. evolutionary algorithms. cognitive scientists. where neuroscientists. and metabolic pathway engineering. The scope of the journal includes developing applications related to all aspects of natural and social sciences by employing the technologies of CI and soft computing. fuzzy logic and evolutionary computation.Computational Intelligence: Past. The Journal of Advanced Computational Intelligence and Intelligent Informatics has been published bimonthly by Fuji Technology Press since 1997. An other journal. artificial life. The topics covered include many CI methods. The Computational Intelligence and Neuroscience has been published by Hindawi Publishing since 2007. and Future 35 of the journal is defined by “Computational intelligence is a well-established paradigm. Today. The journal focuses on the synergetic integration of neural networks. encompassing areas such as computing in biomedicine and genomics. computer scientists. including current and potential applications of CI methods and techniques. The International Journal of Computational Intelligence Systems that aims at covering state-of-the-art research and development in all fields where CI is applied has been published by Atlantis Press since 2008. This journal tries to cover all “advances in computational molecular/structural biology.

and applied research. Although it may seem like a dangerous utopia perhaps deeper understanding of developmental processes. methods and applications in ANN. In the future. hybrid systems and other areas of artificial intelligence.4 shows web page addresses of CI Journals introduced above. implementation and management of neural networks. The International Journal of Computational Intelligence: Theory and Practice aims at publishing papers addressing theories. Table 1. .8 Computational Intelligence’s Future and Conclusions In the future. and D. The journal is related to all areas of CI design and applications: applications oriented developments. but also observing and guiding personal development. development. It should be a technology with access to extensive knowledge. fuzzy systems. successful industrial implementations. methods and applications of CI in organizations. genetic algorithms. fuzzy logic. coupled with a technology that can recognize behavioral patterns. The Journal of Computational Intelligence in Finance has been published by Finance & Technology Publishing since 1993. Duch [25] says “. gently challenging children at every step to reach their full physical as well as mental potential. . . uncertain reasoning techniques. and related machine learning methods has been published by Lawrence Erlbaum Associates since 1996. ¸ building intelligent systems for industrial applications. make sense of observations. evolutionary computation. Cınar I. technology reviews. Kahraman. intelligent agents. one of the most important topics is the answer of the question how we will use CI.” CI techniques will probably be more used in robotics. but it also should help humans to make wise decisions presenting choices and their possible consequences. The journal was focused on the financial applications of CI predictive methods. CI education. not only monitoring the health hazards. will lead to a better world in which no human life is wasted. The journal that publishes original. 1. partner . The journal has been published by Serials Publications since 2006. cognitive and emotional brain functions. . understand natural language. real human needs. The IEEE Computational Intelligence Magazine has been published by the IEEE Computational Intelligence Society since 2006. pet robots. The International Journal of Computational Intelligence and Organizations is a quarterly journal focusing on theories. design tools. high-quality articles dealing with the design. ˙ Kaya. plan and reason with extensive background knowledge.36 C. Computational intelligence should protect us starting from birth.

the robot needs to accumulate the mapping structure between its perceptual system and action system through interaction with the human step by step.atlantis-press.php/ijcitp/ \newlineindex http://elib.cs. because the human tries to find out the causal relationship between human contact and its corresponding robotic action. Furthermore. As a result.ieee.org/ijci/ http://www. Therefore. and Future 37 Table 1.cs. CI will become a solid branch of .com/bw/journal. Today.siu. will continue to be very important.sfu. and entertainment robots will be more developed than now as the next generation of human-friendly robots.fujipress.Computational Intelligence: Past.net/ijcir/ http://www.jsp? \newlinepunumber=10207 http://ourworld. CI techniques will be used to realize this aim successfully. although inspirations drawn from cognitive and brain sciences.compuserve.ripublication.jp/JACIII/ http://sofia.asp?ref=0824-7935 http://www. CI techniques can be used to construct the interrelation between the human and robot through their actual interaction.com/jcib.softcomputing. The communication of a robot with a human requires the continuous interaction with the human.htm http://www.wiley.com/journals/acisc/ http://www.ca/Collections/CMPT/ cs-journals/P-Erlbaum/J-Erlbaum-IJCIO.com/homepages/ftpub/ \newlinejcif.com/publications/ijcis/ http://www.com/ijcia/ http://www.com/journals/cin/ http://www. the human can construct complicated relationship by expanding or integrating the found or constructed relationships.edu/index.4 Web pages of CI Journals Journal Name Computational Intelligence: An International Journal International Journal of Computational Intelligence International Journal of Computational Intelligence and Applications International Journal of Computational Intelligence Research Journal of Computational Intelligence in Bioinformatics International Journal of Computational Intelligence Systems Computational Intelligence and Neuroscience Applied Computational Intelligence and Soft Computing Journal of Advanced Computational Intelligence and Intelligent Informatics International Journal of Computational Intelligence: Theory and Practice International Journal of Computational Intelligence and Organizations IEEE Computational Intelligence Magazine Journal of Computational Intelligence in Finance Web Page http://www.hindawi.waset.ocean.worldscinet.htm robots. or biology in general. From this perspective.org/xpl/RecentIssue.hindawi.html http://ieeexplore.

Expert Systems with Applications.. DePietro. Montero. P. Chap.J. 11. IEEE Transactions on Evolutionary Computation. World Scientific Proceedings Series on Computer Engineering and Information Science 1. [5] Betechuoh. Neural Networks. 26. In: Zurada.S. INES 2008. Cambridge.L. [2] Alvarado. genetic algorithms. Clusterization. F. Rojas-Ram´rez. [6] Bezdek.. Engineering Applications of Artificial Intelligence. What is computational intelligence?. B. The future development of hybrid systems should incorporate various disciplinary knowledge of reservoir geo-science and maximize the amount of useful information extracted between data types so that reliable extrapolation away from the wellbores could be obtained. and approximation capabilities of CI systems are improving day by day and so many methods have been already developed that it is always possible to find alternative solutions with these CI techniques. Computational intelligence methods for process discovery. pp. The MIT Press. J.. Lu. J.. 12th International Conference on Intelligent Engineering Systems – Proceedings. 1997. Computational Intelligence in Decision and Control . Introduction to machine learning. R. T. Kahraman.). In spite of the conflict on scope of the CI. Marks. 1994. R. Evolutionary computation: comments on the history and current state.J. While the problems getting more complex.. J.M. (Eds..P. CI is coming into view that includes a highly interdisciplinary framework useful for supporting the design and development of intelligent systems. [4] Benitez-Read. ˙ Kaya.. pp. 1998. 2008..Proceedings of the 8th International FLINS Conference. 3–8. 3–17. Computational Intelligence. The trend of development of CI techniques will continue until systems as intelligent as human are created. Cınar I.. H.. and fuzzy systems are common methods encountered in CI literature. J. Imitating Life. 1–12. Marwala. London. J. 127–132. Martinez.V.). J. Cantu. E.. . D. Schwefel. 11. 2004. and D. L.. the importance of CI is increasing. 2004. Manana. 2008. and Kerre.38 C. E. This is because the accurate results and different aspects of CI techniques for several problems... there are many numbers of applications in industrial and academic researches. but the researchers do not have any consensus on these models. 1 (1). Bibliography [1] Alpaydın. CI’s future is related to using of the hybrid systems.E. A mobile monitoring system controlled by compuı tational intelligence techniques.. M. C. CI involves innovative models that often come with machine learning. 675–681. D’hondt.. C.. ¸ computer science more than now in the future. 981–986.. J.. IEEE Computer Society Press. E. England.. T. Sheremetov. [7] Cass. Robinson. Ruan. (Eds.. Computational intelligence for HIV modeling.. U. [3] Back. classification. Hammel. Autonomous agents and computational intelligence: the future of AI application for petroleum industry. L.

Neural Networks: A Review from a Statistical Perspective. Springer Studies in Fuzziness and Soft Computing. Studies in Computational Intelligence (J. 2002. Computational Intelligence Techniques for Building Transparent Construction Performance Models. O.. 92 (5). 1996. S. 771–805. IEEE Transactions On Information Technology In Biomedicine.. L. Krasnopolsky. 2–54. 1999. Fanelli. [17] Corne. and comparisons. I. T. Computational intelligence in economics and finance: Carrying on the legacy of Herbert Simon. 216–224..J.. Ltd.. and Titterington. Ant colony optimization theory: A survey. A. Hayashi. [12] Chen.Y. 771–805 [29] Eberhart. What is Computational Intelligence and what could it become?. Tzanakou. Dobbins. P. Encyclopedia of Life Support Sciences.A. Zalzala. Y. Eiben.. See. 92 (5). Blum. Using computational intelligence techniques to model subglacial water systems. Pedrycz. 87 (9).E. D. [27] Duch. C. Pitsillides. 1471–1496. [18] Craenen. Web Intelligence and Agent Systems. 54.pdf [19] Dasey. [16] Cherkassky.. [25] Duch.. R. M. L. pp. M. D. [20] Dimopoulos. Computer Communications.cs. L. C.. 43. 6 (3). 253–272. C. A. 1997.. 2006. Neural Networks...W. 93–113. [24] Douligeris.. [9] Castillo. S. Murray.M. Proceedings of the International Joint Conference on Neural Networks. J. 1166–1171. 37– 60.) Challenges for Computational Intelligence.vu. [28] Duch. W... Using ART1 neural networks with destructive solid geometry for design retrieving systems. and Cybernetics. Statistical Science. 2000. [10] Chang.J. 2. T. W. Panno. 63. 2000. B. pp.. 9 (1).. 2000. Tsai. Springer.. Kacprzyk Series Editor). IEEE International Conference on Systems. J. Zurada. 2008. D. Information Sciences. V. 243 – 278. Mandziuk. 2005. 1997. 121–131... D. Z..A.M.. 4 (2).. H. Detection of Multiple Sclerosis with Visual Evoked Potentials—An Unsupervised Computational Intelligence System. Man. Solomatine. Turton. [11] Chellapilla. http://www.. K. Games. Proceedings Of The IEEE. 1999. Setiono.. C. Proceedıngs of the IEEE. 1243–1264.Computational Intelligence in Complex Decision Systems 39 [8] Castellano. and Chen P. EOLSS. W.. Computational intelligence methods and data understanding. 1. Gambardella.G. A. and Intelligence. 1820–1827. J. [14] Chen.. 25. 113–121... Simpson. 2004.P. G. Heidelberg [26] Duch.M.. [13] Chen. B.S. 2008. 4 (3). Computational Intelligence. Ovaska. [21] Dorigo.. 344. Setiono. 2007. R. Proceedings of the IEEE. P. Computers in Industry. E. Computational Intelligence in Telecommunications Networks. solutions. 34..M. IEEE Transactions on Evolutionary Computation. Computational intelligence in earth sciences and environmental applications: Issues and challenges. EOLSS Co... Evolution.M.M. Recent developments in evolutionary computation for manufacturing optimization: problems. 2004. W.. Computational Intelligence PC Tools.. W. Y. M. Computational Intelligence for Decision Support. 2001. Computational Intelligence Methods for Rule-Based Data Understanding. Zurada. 19. Industrial applications of soft computing: A review. 1413–1414. Theoretical Computer Science. and Valdes. 2005. Openshaw. Duch.. S. 27–41.nl/ci/papers/Comp-IntellCraenen-Eiben. S.. 2000. Computational intelligence methods for rule-based data understanding. Academic . Computational intelligence software: Type-2 fuzzy logic and modular neural networks. Melin.. Journal of Geographical Systems.... B. Ant colonies for the travelling salesman problem. Boca Raton. vol. 73–81.. [23] Dote. and Fogel.. W. V. Torsello. C.... 2006. 256–269. Proceedings of the IEEE 89 (9). [15] Cheng. 170..M. R. 1994. Neural Networks. CRC Press. BioSystems. R.. A. (eds. 2008. [22] Dorigo.. Computational Intelligence techniques for Web personalization. J.

Lu. Neural Networks In Computer Intelligence. 2005. 2003. (Eds. 8 (4). Kahraman.. Neural Networks. 39 (4). [43] Hansen. [38] Gen. Reducing the Time Complexity of the Deu randomized Evolution Strategy with Covariance Matrix Adaptation (CMA-ES).. Kaya. P. 2000. Review of Computational intelligence: imitating life. Learning classifier systems: New models. [36] Fogel... Lanzi..D. R. successful applications. Koumoutsakos. New York: John Wiley and Sons. A. 2002. An integrated computational intelligence approach to product concept generation and evaluation. [34] Ferreira. 2008. P.P.N. N. Computational intelligence in bioinformatics. Taleb-Ahmed. Cheng. P. 1654–1663... 46 (1). Ruan. D.. M¨ ller. [49] Jahanshahloo. Pinti. A.S.. K. Ruano. Extension of the TOPSIS method for .Proceedings 1. Applied Soft Computing. Comparing backpropagation with a genetic algorithm for neural network training. 2006. 23–30. 11 (1). A. Computational Intelligence to Model the Export Behavior of Multinational Corporation Subsidiaries in Malaysia.H. 2008. ¸ Press. 7–12. J. 163–170.. Lotfi. 3582–3589. and Petrovic-Lazarevic. W. Journal of the American Society for Information Science and Technology. F. [39] Gu. Montero. Quintana. 2004 2nd International IEEE Conference ‘Intelligent Systems’ . Y. BioSystems. 1562–1565. [45] Holmes. Computational Intelligence.). [30] Edwards.... pp.. 23 (2). 72. 111–116.. World Scientific Proceedings Series on Computer Engineering and Information Science 1. 2007. M.. 56 (11). 2121–2124.R. Ann CIRP.. R.G. An integrated design methodology for life cycle engineering. R. W. 2003. IEEE Trans. Genetic Algorithms and Engineering Optimization. 1177– 1186. R. J. 50 (1). Corne. S. IEEE Transactions On Magnetics. Abraham. Computational intelligence approach to real-world cooperative vehicle dispatching problem..W. R. 1997. J.M..E. 82. O. Cınar I. Applied Computational Intelligence for Finance and Economics. W. and D. [37] Fu.L.... 567– 583. A. S. Chen. P. Wilson. 32–40. 2003. Z. J.. G. B. Izadikhah. E. A. 1999. Xianyi. 27. Martinez... Evolutionary Computation. [48] Isasi.M. [32] Engin.. [47] Innocent. Application of computational intelligence methods to greenhouse environmental modeling. Artificial Intelligence in Medicine. McGraw-Hill. ˙ Kaya. Chen. Sosale.. 2006. 2003. Garibaldi. Dong. D. Celik. A.A. D’hondt. Computational Intelligence in Decision and Control . [41] Gupta.L.. Boston. H. Journal of the ACM. P. Some Challenges and Grand Challenges for Computational Intelligence.. D... P. Ram´rez.. 21. The fuzzy medical group in the centre for computational Intelligence. Barros.H. and Kerre. 2004.. M. F. 1–18. K. [42] Han. 1–4. E. L. [40] Guimar˜ es... P... S...D. Saez.... 71–74. [46] Huang.. 679–684. United States of America. Hashemian.. pp. A fuzzy approach to define sample size for attributes control ¸ I.. Information Processing Letters. Mechanism and Machine Theory 41. [35] Fogel. 1994. K. 2001. Computational Intelligence: An Introduction. Proceedings of the International Joint Conference on Neural Networks.H. Omega... [33] Feigenbaum.I. S.40 C. Wiley. F. [31] Engelbrecht..Proceedings of the 8th International FLINS Conference.. 2003. ˙ 2008.. Mochon. P. Integration of human knowledge for automatic tissue classification on medical images. Stolzmann.E. pp.A.. M. D.Z. J. G.. Bo. An Object-Oriented Library Based a ı on Computational Intelligence Techniques for Optimization in Electromagnetics. L.R. chart in multistage processes: An application in engine valve manufacturing process.. A.. 447–452. J. 6. and Sexton. John. [44] Hirota. 1995.

L.. Kaya.R. Control of a phosphate processing plant via a synergistic architecture for adaptive. multistage processes: An application in engine piston manufacturing process. 373–401. 2003.G.. Cambridge. R. Zaneldin.L. with variables and attributes. J. Chan. Comparison of computational intelligence and statistical methods in condition monitoring for hard turning. Koza. 2008.. C.. Li. Schmitt-Landsiedel. Proceedings of the 2005 IEEE International Conference on Computational Intelligence for Measurement Systems and Applications.). J.. Lam.C.. F. 597–610. 14.. An automatic parameter extraction technique for advanced CMOS device modeling using genetic algorithm.R. O. Expert Systems.. Advances in Identification of Nonlinear Characteristics of Packaging Based on Computational Intelligence. L. 181. 2005.. Facility Location Selection Using A Fuzzy Outranking Method. charts. Lai. 2001. Stanford University. Liau.T.W. Neuro-Fuzzy modeling and control. Hegazy. Genetic Programming. Monitoring the supply of products in a supply chain environment: A fuzzy neural approach.J. and Sun. B. Information Sciences. Z. Y... E. A new approach to define sample size at attributes control chart in I. Kubota.. 2002. 2005. R. 179 (10). Engineering Applications of Artificial Intelligence.. Y. 1544–1551 Jang. (Ed. 273– 278. J. Shabeeb. H.. Fuzzy multi-criteria decision making theory and applications with recent developments...S. 84 (2).. H. Artificial neural networks in renewable energy systems applications: a review. Wong. 1552–1566. E. Ling. JourI. Transmission and Distribution. Expert Systems with Applications. S. Leung. 616–620. ˙ 2009a. Karr. 21–30... Renewable and Sustainable Energy Reviews. Microelectronic Engineering. 5. J. Kothamasu. 27 (1). intelligent control. 2000. W. 1998. Koza. 83 (3). Applied Mathematics and Computation. ˙ Cınar. 2007. F. 2005. K. ˙ Engin. 1942–1948. 2007. Proceedings of the American Control Conference. International Journal of Production Research. 1995. 251–263.Computational Intelligence in Complex Decision Systems 41 [50] [51] [52] [53] [54] [55] [56] [57] [58] [59] [60] [61] [62] [63] [64] [65] [66] [67] [68] [69] decision-making problems with fuzzy data. MA: MIT Press.R. Comparison between evolutionary programming and a genetic algorithm for fault-section estimation. 145 (5).L. 15–20. 378–406. Sichanie... D.S. C. A. International Journal . Huang. A genetic algorithm approach to determine the sample size for control charts I. Technical Report STANCS-90–1314. 1995.Y. Elbeltagi. E. Lau. Information Sciences. Mechanics Research Communications.. Kalogirou. Computational intelligence semiconductor ATE . ˙ 2009b.. Springer. 16 (1). Eberhart.. 1992.. 8719–8734. and Verduin.. 2005. Computational Intelligence Tools for Solving the Facilities Layout Planning Problem. 1990.. C... Genetic programming: a paradigm for genetically breeding populations of computer programs to solve problems. F.. Wong. Chunguang.C. 235–243.. Journal of Materials Processing Technology.Ultrashort worst case test. 260–272.W. 36 (5). 19 (4). 403–429.. ¸ nal of Multiple-Valued Logic and Soft Computing.T. A. S. F. 171. K. Department of Computer Science. A genetic algorithm approach to determine the sample size for attribute control I. Computational intelligence for structured learning of a partner robot based on imitation. 2008. H. 3954–3958. I. Kennedy. CIMSA 2005.. H. 43 (3).K. T.. Proceedings of the IEEE International Conference on Neural Networks... Hui. H. December.. Kaya. 38–48. 2000. Kahraman. Proceedings of the IEEE. Kaya. Kaya. Liang. D. Gwyn. pp. S.A.. Karray. N. 183 (1). Computational Intelligence Techniques for Home Electric Load Forecasting And Balancing. C. IEE Proceedings: Generation. H. Particle swarm optimization.

T. 1997. 15. 521–527...K. SSD’08. 2006. The Journal of Financial and Quantitative Analysis. Man. IJCAI ’95 Workshop. Computers and Geotechnics. R.. 201–214. S. Canada. T. 3.. G. J.F. B. 28. [73] Moriarty. 439–445. A. and Krost. S. IEEE Transactions on Neural Networks. W.S. 2002. [72] Matics. R. B. Neurocomputing. 1995. M. Efficient Reinforcement Learning through Symbiotic Evolution. R. (Eds. Park.N. Cınar I.. A. [71] Martin. 2002. Giordano R. 2005. International Journal of Intelligent Systems.). E. and Cybernetics—Part b: Cybernetics.S.. 33. 2002. 5th International MultiConference on Systems. Th. S. Machine Learning. Pedrycz. [82] Nucci. Z.. Weller. A. Ravazoulad. 683–697. 397–431. H. 72–79. Ralescu. El-Mousa. A Fuzzy Logic Algorithm For Identification of The Harvesting Threshold During PGA production by Bacillus megaterium. Cruz. 483–486. [78] Naso. A distributed workflow management system utilizing computational intelligence for decision support. G.. P. Okuda. Glotsosc.-S.P. IEEE Transactions on Systems. A. [87] Pal. Fusion of soft computing and hard computing in industrial applications: An overview.. 1996. 3. Implicit rule-based fuzzy-neural networks using the identification algorithm of GA hybrid scheme based on information granulation.T. 16 (7). C.. 2005. 22. [80] Neely. M. Is technical analysis in the foreign exchange market profitable? A genetic programming approach. Sasaki.. [76] Nakayama. Fundamentals of the New Artificial Intelligence Neural. H. S. pp. 2008... S. Evolutionary. Arakawa. T.G.D. Application of genetic algorithms to conceptual design of a micro-air vehicle. 773–779. 11–32. [74] Muhsin.. [86] Pal. Computational Intelligence for Decision-Making Systems.. [75] Munakata.. S. 371–391.. [79] Ndousse.A.I. Fuzzy and More. Signals and Devices. Fuzzy Logic in Artificial Intelligence.. Spyridonosc. [83] Oh. 18. S.B. R. 1996. Computational versus artificial. 405–426. C. and Mudi. Omkar.. Sivapullaiah.. P. Brazilian Journal of Chemical Engineering. Suresh. Advanced Engineering Informatics. Selected Papers. R. 3 (5). Kahraman. [81] Ng. Brain tumor characterization using the soft com- .S. Multi-layer hybrid fuzzy polynomial neural networks: a design in the framework of computational intelligence. Dittmar..R...R.H.. Park.. Leng.. 196–208. 1992. 2002. L. Micro combined heat and power home supply: Prospective and adaptive management achieved by computational intelligence techniques. T. [89] Papageorgioua. [84] Oh. P. Kamiya. W.-K. Nikiforidisc. Applied Thermal Engineering... 4. 1993.P. Optimization and Engineering. Pal. Groumposa.. Simulation-Based Optimization Using Computational Intelligence. 208–226.. G. Engineering Applications of Artificial Intelligence. 16 (4). Man and Cybernetics Part C: Applications and Reviews.. 2008.K. 2055–2061. Multicriteria Meta-Heuristics for AGV Dispatching Control Based on Computational Intelligence.. August 19–21.. [88] Pal.. 2008.J... N.J.H. Silva. D.. and Miikkulainen. 22 (4).E... IEEE Transactıons on Neural Networks..42 C. C. Springer. ¸ of Computational Intelligence and Applications.. 247–263. 32 (4). 32 (2). Al-Taee. R. and Classification. Prediction of uncon.J.. pp. and Mitra. G.D. ˙ Kaya. 1995. Computational Intelligence for Pattern Recognition. H.ned compressive strength of soft grounds using computational intelligence techniques: A comparative study.. Fuzzy Sets. P. 64. [70] Marks. E. Styliosb. 35 (2). C. 1558–1562.V. D. 2003.N. Montreal. International Journal of Pattern Recognition and Artificial Intelligence. Springer.. P. 3.K. R. N. VanLandingham. IEEE Transactions on Systems. D. T. Multilayer Perceptron.. Gomes. [77] Narendra. 1–6. S... H. Pedrycz. 2008. and Turchiano. T. Computational intelligence for distributed fault management in networks using fuzzy cognitive maps. and D. 737–739. 2002. [85] Ovaska. 2005.. IEEE International Conference on Communications.

Artificial Intelligence in Medicine. 2001. D. 7.. [94] Pavlidis.. 150 (2). [90] Park. (Eds. 2053–2062. 2008. Genetically optimized logic models...A. Seo. [103] Puskorius.. Mackworth.. 304–332. W. 10 (5). Vrahatis. New York. Information and Software Technology. Rule-based fuzzy-neural networks using the identification algorithm of the GA hybrid scheme.. Pizzi.. 956– 961. D. A. Applied Soft Computing.. G. and Kerre. J. D.. Automation and Systems. in: Proceedings of Second International Symposium on Environmentally Conscious Design and Inverse Manufacturing. 2008. Vrahatis.E. Software quality analysis with the use of computational intelligence. 321–331. C..F. Research in International Business and Finance. ... S. 2003.. 1994. Computational Intelligence in Decision and Control . Soft Computing.S.S. N. Neurocomputing. International Journal of Control. E... (Eds. 321–326. pp. K. 175. Genetic tolerance fuzzy neural networks: From data to fuzzy hyperboxes. Computational Intelligence in Software Engineering. Analysis of foreign investment impact on the dynamics of national capitalization structure: A computational intelligence approach. 35.-K. 823–828.. [91] Park.. L. Oh. 19. 2000. A nature-inspired computational approach to discern unhealthy nuclear intentions of nation states. 2003. Oh. 351–371.G.V. [105] Quah. Real-time credit card fraud detection using computational intelligence. Journal of Computational and Applied Mathematics. Reformat. 2003. 101–110. M.. S. Singapore. [96] Pedrycz...E. [97] Pedrycz.K. M. Montero. W... 113–136. Automation and Systems. P. 1998..D.J. D. 820–828. Pedrycz. B.-H.. [106] Rao. S. and Poliakova. Siper. 2005. 2002.. and Feldkamp.. D’hondt. 1721–1732. 2002.). W. Approximate life cycle assessment of classified products using artificial neural network and statistical analysis in conceptual product design. World Scientific. Computational Intelligence – A Logical Approach. Computing Nash equilibria through computational intelligence methods. 405–417. Martinez. 45. Wallace. 19.. 1 (3). Lu.P. 8 (1). 2007. S. Oxford University Press. D. J... 1–23. 2006.. [102] Poole. W. N.. K. Peters.S.G. L.. V.N. pp. M. Utilizing computational intelligence for DJIA stock selection. N. 1403–1413. Rule-based fuzzy polynomial neural networks in modeling software process data. R. W. [101] Plikynas. Tasoulis. Fuzzy polynomial neural networks: Hybrid architectures of fuzzy modeling. Plagianakos. Computational Intelligence Methods for Financial Time Series Modeling. 2007. Evolutionary computation in medicine: an overview. [104] Quah.K. W. Computational intelligence and visual computing: an emerging technology for software engineering. M. [100] Pena-Reyes..Proceedings of the 8th International FLINS Conference. [92] Park. J. [107] Reformat. Pedrycz.Y.Computational Intelligence in Complex Decision Systems 43 puting technique of fuzzy cognitive maps. IEEE Transactions on Fuzzy Systems. B. International Journal of Bifurcation and Chaos.. 1 (1). A.T. vol. Sakalauskas. J. Neurocontrol of Nonlinear Dynamical Systems with Kalman Filter Trained Recurrent Networks. 16 (7). Parsopoulos. D..-K. H. 5 (2).J. Lee. 2005. 70 (7–9). [93] Park. Ng.. Ruan. W. [99] Pedrycz. Advances In Fuzzy Systems—Applications and Theory.. Sriganesh. World Scientific Proceedings Series on Computer Engineering and Information Science 1. International Journal of Control.A.J. Goebel. J. Expert Systems with Applications.. 607–621. M.. 16... 2005.N..). J.. 33–44. Oh.K. IEEE International Conference on Neural Networks – Conference Proceedings. Fuzzy Sets and Systems. 279–297. [98] Pedrycz. [95] Pavlidis. T. 1998. IEEE Transactions on Neural Networks. M. L.

Kahraman. (Eds. D’hondt. D. USA. E.. 2005. Singapore. and genetic algorithms... and Kerre.. Intelligent technoloı gies and soft computing in nuclear science and engineering. M. (Eds..E. D. Intelligent Systems. 2000. Singapore.. R. Computational intelligent systems for applied research. D. and Kerre.. H.. Expert Systems with Applications. Ceruti. and Vasilakos. A.. Proceedings of the FLINS’98. D. 2008.). R.. 2007.).P. Communications & Control. L.. 2006. [119] Ruan. 547–589. J. and Kerre. P. De Cock.E. Montero. Govaerts. Nachtegael... World Scientific Publishing Co. E. Computational Intelligence Applications to Power Systems. McGraw-Hill. and Johnson. [109] Rooij. Applied artificial intelligence. C. G.). Fantoni. Applied computational intelligence. D’hondt.S. [120] Ruan. Y.E. 1996. N.H. World Scientific.V.44 C. Comparative evaluation of genetic algorithm and backpropagation for training neural networks... J.. 21 (3). (Eds. S. P. Dai.. [126] Sexton. and Kerre. T. Jain. I. Ltd. [127] Song. Information Sciences. T.J.... World Scientific.. World Scientific. Computational intelligence in decision and control.E. Computational intelligence approaches and linear models in case studies of forecasting exchange rates. [121] Rubin. R.D. 177. Proceedings of the FLINS’94. Proceedings of the FLINS2008.. D’hondt...A. Fuzzy Logic with Engineering Applications. 5292– 5315. 2000. E. and D. World Scientific. 129.. [113] Ruan. [128] Ssali. D. Fuzzy logic and intelligent technologies in nuclear science..C. L.. 2004.. Martinez. and Kerre. P. [125] Sebastian. (Eds. Singapore..P. A. Coelho. Al-Madani.E. Springer.H. Microprocessors and Microsystems. Cybernetics and Systems: An International Journal. Lu. D’hondt.). Computational intelligence and decision trees for missing data estimation.. neural networks. H... 192–198. Proceedings of the FLINS 2000.. E. World Scientific. I.F. J.. ¸ [108] Riziotis. E.. .J. 1995. Proceedings of the FLINS’96.. 31.).. 21–25.. 33. 2008. D’hondt. 2008. (Eds. Proceedings of the FLINS 2006. (Ed.. 2007. [118] Ruan. J. (Eds. M. Govaerts. and Gupta. Springer. [115] Ruan. (Eds. and Kerre. P. 132–138 [123] Salman. P. and Kerre.). Fodor.. Intelligent systems and soft computing for nuclear science and industry.. M. Using Computational Intelligence in Fuzzy Engineering Design. Fuzzy logic and ı intelligent technologies for nuclear science and industry. [116] Ruan. World Scientific....E.. Marwala.. E.N. Pte. 2004. Nachtegael. ˙ Kaya. Journal of Dong Hua University (English Edition). Allan.. A¨t Abderrahim. D’hondt. 1997.. D. W. III. [110] Ross. Role of computational intelligence in sensory evaluation. and Kerre. (Eds. 45–59.E.. S.F. 1994.E.). [111] Ruan... D. J. Proceedings of FLINS the 2002. Singapore.. Cınar I. D.A. [122] Rudas. 363–371. D. Singapore. World Scientific. Singapore.. P. World Scientific. 26. A.G. Proceedings of the FLINS 2004. De Cock. A¨t Abderrahim. D.). L. [112] Ruan. P. Intelligent hybrid systems: fuzzy logic.. Singapore. 816–823. P. The role of computational intelligence in data mining.. [124] Santos. P.). M. Schleiffer. Proceedings of the 2005 IEEE International Conference on Information Reuse and Integration. Singapore. A. P... 1996. Costa Jr. 1997. Neural network training using genetic algorithms. Information Sciences.C. E. D’hondt.. Singapore. A.. Raj.. [117] Ruan. Particle swarm optimization for task assignment problem. P. J. D’hondt.S. 1998. H. [114] Ruan. Computational intelligence in photonics technology and optical networks: A survey and future perspectives. Ahmad. 2002. 2000. M. E.. Proceedings of the International Joint Conference on Neural Networks.. 2002. pp. International Journal of Computers.

..E. Computational Intelligence in Nuclear Engineering..L.V.T.L..C. K. Athanassios. T. 82–102. L. Computers in Industry.. V. Application of computational intelligence techniques in active networks. S. 4 (2). [144] Yang. A.. J. John Wiley & Sons. Small-signal stability assessment for large power systems using computational intelligence. Applied Acoustics. Yang. Zou.. 31. Coding and clustering of design and manufacturing features for concurrent design. Evolutionary programming based economic dispatch for units with non-smooth fuel cost functions. Anagnostakis. IEEE Transactions on Evolutionary Computation. Evolutionary programming made faster.K. [133] Uhrig. pp.. Bennett. A New Evolutionary System for Evolving Artificial Neural Networks. 112–118. E. and Liu. Tzafestas. 1996. L. Journal of Intelligent and Robotic Systems. pp. 139–153... [143] Xue. 17. C. Applied Intelligence. Robotics and Computer-Integrated Manufacturing.. 103–111... 2004. 2004..T. [145] Yao. L. Erlich.. 2004.. 1992. Computers in Industry. 2001.. Rezvani. 693–704.. D. 7–68.P. D. X. 1998. [147] Zadeh. Z. Tieu.. 11–23. [137] Wang. Neural Networks and Fuzzy Logic Applications in C/C++. [141] Wilson. 3 (2). I. 20 (1). Canada. 2000. 2005. 127–138. Computational Intelligence Techniques for Short-Term Electric Load Forecasting. [131] Teeuwsen.G.T.. Design candidate identification using neural network-based fuzzy reasoning. [136] Wang. [129] Stathakis. Pedrycz.R. 2005 IEEE Power Engineering Society General Meeting 3. 694–713.. 1997.. The concept of a linguistic variable and its application to approximate reasoning”. J. H. 201–207. 871–885. X. P. [132] Tzafestas. Intelligent Control System for Slag Foaming. and Fricke. 264–271... Information and Control. Fuzzy Logic Equals Computing with Words. and Maguire. 2001. H. 165–177.L. 1999. . 1994. 44 (8).A. 1975. A. 14 (1). 8. [135] Venugopal. D’Alessio. [140] Welstead. J. Statistical and computational intelligence techniques for inferential model development: a comparative evaluation and a novel proposition for fusion.. [138] Warne. Huang.W. Karr. Predicting and optimising the airborne sound transmission of floor–ceiling constructions using computational intelligence. Y. Dong. Kalenchuk. [148] Zadeh. 37 (2). F. 20 (3).. IEEE Transactions on Fuzzy Systems. 11 (1). 1996. 2661–2668. An Adaptive. S. E.. L. [139] Weber. Proceedings of IEEE International Symposium on High Assurance Systems Engineering 8.. Three-grade fuzzy synthetical decision for mechanism concept. 116–125. 2004. G.A. R. Computational Intelligence-Based Process Optimization for Tandem Cold Rolling. IEEE Transactıons on Neural Networks. 65. V. J. El-Sharkawi.. J. G. Soft Computing. Nannariello.. Engineering Applications of Artificial Intelligence.P.. 1997. 479–496.A. IEEE Transactions on Geoscience and Remote Sensing. M.A. 199–249. Materials and Manufacturing Processes. 2006...... Fuzzy Sets. 338–353. Naredran. [142] Xu. W. C. L. K. D. S. 2305–2318. lnf Sciences.. Machine Design and Research. and Hines. Prasad. 382–396. S. Neural network model for design retrieval in manufacturing systems.. 1965. Knowledge management for computational intelligence systems. 2005. R.K. 20 (2). D. 9–11... [149] Zadeh. 8 (3). D.. Comparison of Computational Intelligence Based Classification Techniques for Remotely Sensed Optical Image Classification. IEEE Transactions on Power Systems. [130] Sun. 8. Nuclear Engineering and Technology. [146] Yao. 34. [134] Vasilakos. 2005. 16 (5).D. Wu.Computational Intelligence in Complex Decision Systems 45 pp.

1987.. H..-J. 431–454. and D. Lin. Dordrecht. Computational Web Intelligence (CWI): Synergy of computational intelligence and web technology. Cınar I. [152] Zimmermann. H. Computational Intelligence and Environmental Planning. Cybernetics and Systems: An International Journal. Z. Boston. 2004. Zhang. Agent-Based Hybrid Intelligent Systems An Agent-Based Framework for Complex Problem Solving. C. Lancaster. T. [153] Zimmermann.46 C. Kahraman.. 2002.-J. pp. ˙ Kaya. 35. (Eds. IEEE International Conference on Plasma Science 2. decision making.-Q.. . Springer. 2004. and expert systems. Fuzzy sets.Y. Y.). 1104–1107. [151] Zhang. ¸ [150] Zhang.

as depicted in Eq. and Lu´s F. Basically. which considers previous solutions. this type of problems includes two main tasks. Atlantis Computational Intelligence Systems 2. Rita A. 2829-516 Caparica. Ribeiro. 1 corresponding author D. Once a decision is made (alternative selected) the problem is solved. 2.1007/978-94-91216-29-9_2. Computational Intelligence in Complex Decision Systems. a MCDM model includes two phases [3]: (a) classify the alternatives regarding each criterion and then aggregate the respective classifications to obtain ratings per alternative. Ruan. from a finite number. In this work we introduce multiple criteria dynamic decision making models (MCDDM) and discuss contributions to deal with the difficult problem of imprecision in dynamically changing input data. regarding a set of pre-defined criteria. we must consider a dynamic and adaptable decision model. the aim of Multiple Criteria or Multiple Attribute Decision Making problems [1] (henceforth called multiple criteria) is to find the best compromise solution from all feasible alternatives assessed by pre-defined criteria (or attributes). To illustrate the approach. Portugal Email: rar@uninova. Campus FCT/UNL. 2]. Sim˜ es ı o UNINOVA-CA3.1 Problem Statement In general. However. Pais. This type of problems is widespread in real life situations [1. © 2010 Atlantis Press/World Scientific 47 .Chapter 2 Uncertainty in Dynamically Changing Input Data Tiago C. rating the alternatives regarding each criterion and then ranking them. for situations involving reaching consensus or requiring several steps before reaching a final decision. A Multiple Criteria Decision Making (MCDM) problem is usually modeled as a decision matrix. a simplified example of autonomous spacecraft landing is presented. Usually.pt1 The main objective of multiple criteria decision making models is to select an alternative. DOI 10. (1).

A. Pais. and “involves the notion of feedback”. An ⎢ ⎢ x21 ⎢ ⎢ . ⎡ C1 C2 · · · Cm x11 x12 · · · x1m x22 · · · . and L. . dynamic systems usually have spatial-temporal limitations such as. m) and xi j is the evaluation (or the satisfaction) of alternative Ai regarding attribute C j . ⎦ · · · xnm where Ai is the alternative i (= 1. ⎥ . this process is called “dynamic evaluation process” because it involves aggregation of past and current ratings. the evaluation/selection process is more sophisticated because it must include information from previous iterations (historic data). One important requirement that both static MCDM methods [1] and MCDDM methods face is that data must be both numerical and principally comparable. C j is the attribute j (= 1. we need to transform and normalize the data to compare them. . Sim˜ es o (b) order the alternatives (ranking). Here we also address another important aspect of many multicriteria decision problems: uncertainty in input data. as well as selection of a sub-set of possible “good” candidates for the next iteration (historical feedback). Sometimes. Ribeiro. Henceforth.48 T. . When in presence of both qualitative and quantitative variables. . We will address four main aspects of any dynamic spatial-temporal decision problem: . Our main objective is to present contributions for handling the data preparation process in MCDDM models with imprecise input data. . ⎣ . According to Richardson and Pugh [5] system dynamics problems have two main features: “they involve quantities which change over time”. . specifically decision paradigms with several iterations. xn1 ⎥ x2m ⎥ ⎥ . In general. ⎢ . xn2 ⎤ (1) A1 A2 . In MCDDM a big challenge is how to represent dynamic input changing data. n). . R. . . Moreover. .F. . ⎥ . . when we are dealing with dynamic processes.. .C. where the highest rate usually corresponds to the “best” alternative to be selected. the search space varies for each step (iteration) and there are several steps (iterations) to reach a final decision [4]. In this work the context is Multiple Criteria Dynamic Decision Making (MCDDM) models. which need to take into account past/historic information (feedback) to reach a decision (or decision consensus). . All these aspects fall under a general umbrella of data preparation [6] and they constitute the basis for dynamic spatial-temporal decision models. there is a lack of confidence and accuracy on the available input data and we must devise ways of incorporating this uncertainty on the transformed input data.

The two main phases.1 we depict the conceptual design for the dynamic decision model. 3]. depending on criteria satisfaction and decision process phase (iteration). there are few contributions in the literature (see for example [9. when facing dynamic multiple criteria decisions. 11]). and they will be discussed in detail. The fourth aspect is mostly concerned with suitable operators to aggregate current and past information that should be considered at each iteration of the decision process. 3]. To better explain our proposed contributions to deal with uncertainty in dynamically changing input data during the temporal decision process. we use a simplified example of a site selection problem for spacecraft landing on planets [12]. respectively. which includes historic information to ensure a truly dynamic and adaptable decision algorithm. (IV) How to use feedback past information from the dynamic model to evaluate the alternatives at each decision point.Uncertainty in Dynamically Changing Input Data 49 (I) How to represent and normalize the input parameters within a dynamically changing input data. 2.1. 1. (III) How to consider different criteria importance. The latter will only be addressed here briefly because the authors already proposed some contributions for this process [8]. However. (II) How to deal with uncertainty in dynamically changing input data. Summarizing. 4. 6]. There are many methods and techniques to deal with static MCDM [7. shown in Figure 2. This architecture involves two important aspects: how to perform a correct data preparation when we are dealing with dynamically changing input data and how to deal with a complex evaluation process. The first three questions belong to known literature processes about data preparation and multicriteria processes [7. 1.2 Description of Methodology In this chapter we present an architecture for a multiple criteria dynamic decision model. the aim of single or dynamic Multiple Criteria Decision Making problems is to find the best compromise solution from all feasible alternatives. where several iterations are at stake and feedback from step to step is required. are: . 10. assessed by pre-defined criteria (or attributes) [7. This case study requires a dynamic and adaptable decision model to ensure robust and smooth site selection along the final descent. In Figure 2.

As mentioned. When dealing with MCDDM problems there are three important questions to answer: how to define and represent the variables? Is the raw input data extracted accurate and do we have confidence in it? Do we know if the search space includes all possible alternatives to be evaluated at . we need to transform and normalize the data to compare them. 2. R. For this phase we propose six steps that are detailed in Sec. Its main tasks are to identify the criteria and the alternatives of the problem and then to clean and transform the input data to obtain parameters/variables that can be manipulated to achieve a decision. the historic sub-set is created and is provided as feedback for the next iteration. (A) Phase A.50 T. both static and dynamic.1 Conceptual design of the dynamic decision model. is that data must be both numerical and principally comparable. in this work we focus on the first step.2. Pais. in presence of uncertainty. Sim˜ es o Figure 2. 2. Decision Evaluation Process with Feedback – This process starts with rating the current alternatives (assessed by criteria) and then performs evaluation of alternatives. considering rating value plus the historic information available. This process goes on until a stoping criterion is reached. Data preparation in dynamic environments – This process deals with transformation of inputs for the decision model. Finally.C. Data Preparation in Dynamic Environments Here we discuss our proposed data preparation steps for handling dynamically changing input data. one important requirement of MCDM methods [1].1. and L. Ribeiro.1 Phase A. This work also addresses an important aspect of many multiple criteria decision problems: uncertainty in input data.2. As mentioned. (B) Phase B.A. When in presence of both qualitative and quantitative variables.F. hence more emphasis is put on Phase A in the remaining sections.

by normalizing and representing the input variables with fuzzy membership functions.. besides allowing us to deal. i. representation and normalization) all candidate alternatives with membership values of zero. Parameters identification – This process is characterized by identifying the input variables (problem understanding). This process corresponds to including some degree of uncertainty (either given from experts or from statistical processes) when there are doubts about the correctness of collected data. WPM and TOPSIS. Weighting criteria – we should also consider the assignment of relative importance to criteria (weights). such as MaxiMin. gaps in collect data etc. • Step 2. number of alternatives to be evaluated with the defined criteria. This method simply discards an alternative if it has a criterion value below a threshold.Uncertainty in Dynamically Changing Input Data 51 each step? Our proposed data preparation steps for MCDDM with uncertainty are: • Step 1. with imprecision in data. • Step 5.). Cleaning – After defining the fuzzy set functions. • Step 6.. in a simple fashion. • Step 4. selects dominating alternatives – better or equal in each criteria – and from this set trade-off analysis can be performed).. as well as identifying the set of candidate alternatives [3].g. Fusion – transformed and cleaned data from each criterion is integrated using a conjunctive method [7] to define the general search space. in any criterion. either per iteration or for a set of iterations. This process is usually called fuzzification [2] and involves choosing the topology for the fuzzy sets representing the input variables. per iteration.e. For this step we suggest using weighting func- . any other constraints that alternatives are required to obey will be used to further clean the data. In a multiple criteria decision making environment these correspond to determine criteria for assessing candidate alternatives. • Step 3. A possible solution could be to use the previous value when there is missing data [13]. The Conjunctive method is an outranking method [7] so it does not give a score for each alternative like most common scoring methods. for each criterion (i. Moreover. thus further reducing the set of candidate alternatives. Transformation – Data is transformed from a raw state into data suited for decision support. With a fuzzy logic approach we can guarantee normalized and comparable data definitions. are discarded. To obtain the final score a Pareto optimal strategy can be used (i. A filtering function is presented that can handle this type of intrinsic uncertainty in input data. This process should elicit criterias importance. Filtering uncertainty – Data should be filtered to consider both lack of accuracy and/or lack of confidence on the collected input data. WSM. Another aspect of cleaning is how to handle missing input data (e.. missing records.e.e.

besides using rating values to rank alternatives we also need to include historic information.2 Phase B. Sim˜ es o tions [14. In summary. However. The historical set is refreshed at each iteration. The notion of feedback is introduced to express that at each step (iteration) we build a sub-set of good potential alternatives (historic set) that is re-fed into the system. This feedback continues until a final decision is reached (consensus) or there are no more iterations.A. 2. we depict the proposed data preparation process with the six steps and their inter-relations. More details about each step and proposed formulations are discussed using the illustrative example to improve clarity of contributions. Decision Evaluation Process with Feedback In a simple MCDM model this process would be a simple ordering of alternatives regarding the respective rating value. In Figure 2. and L. a dynamic decision process with feedback includes three processes for its evaluation of candidate alternatives.2. Pais. R. usually denoted ranking of alternatives [3]. 15]. thus ensuring that best classified alternatives. using an aggregation method such as the weighted average and weighted with crite- . have a better classification in the next iterations. Ribeiro. about preliminary decisions taken along the decision process (iterations).1: (1) Rating process. because they can penalize or reward satisfaction levels per criterion. Figure 2. during a certain number of previous iterations.2. in a dynamic model. as input for the next iteration. as shown in Figure 2.52 T.F.C. regarding all criteria. During iteration t each alternative is classified.2 Data preparation process.

⊗ is the product operator. if we had 262 144 alternatives it would be quite inefficient to pass all data sets from one iteration to another. 16. There are many aggregation methods that can be used for this purpose [7. The dynamic evaluation algorithm for combining historical information (H(t − 1)) and current rating values (R(t)). W (aci j ) represents the weighting function for accuracy and confidence (topics detailed in (see Sec. henceforth called evaluation process. 1]. For instance. 11] that could be considered to deal with this evaluation process. using weight generating functions [14.W (acin ) ⊗ acin (2) is the aggregation operator (in our case Weighted Sum Method). During iteration t. . We refined the uninorm.2). This combination process. The proposed formulation for the rating process in our MCDDM is: ri = where W (aci1 ) ⊗ aci1. the previous historic set is combined with the rated sites at iteration t (obviously there is only combination if the same site exists in iterations t and t − 1). is based on the uninorm operator [16]. we propose to use specialized full-reinforcement operators [16. for a specific MCDDM we should define an appropriate size for k (best classified alternatives) and then record their respective rating to act as feedback for the next iteration. defining a hybrid operator [8] because we needed an operator with a compensatory nature outside the T-norms and S-norms [18] interval. but. ensures a continuous refreshment of the historic set for the next iteration (feedback). Hence. 15]: (2) Evaluation Process. The rating process deals with aggregation of all contributions per alternative regarding each criterion. The Historic process defines the set k of rated alternatives that are worthwhile considering in the next decision iteration. The set k will record the most relevant information at the end of each iteration. . . 7. We will not discuss more details about this phase. The size of k is context dependent.3. During the decision process. Noteworthy that we will only mention one historic set because T = 2 and the last historic set is H(T − 1) (= H(1)). As mentioned we propose to compute the weights to express the relative importance of each criterion. . because our focus is on data preparation in environments where input data changes dynamically. 12]. 2. We chose the uninorm operator because it presents full reinforcement behavior [17] and this is what we are looking for in our adaptable and dynamic decision process based on historical data. There are a wide variety of operators and techniques [18. . (3) Historic process. we should compute these values for each alternative (per iteration).Uncertainty in Dynamically Changing Input Data 53 ria relative importance. 1. as pointed in previous works [8. 17] because it provides flexibility within the search space.

light level and terrain roughness. In previous works. evaluated with respect to a set of requirements (i. • Site selection that chooses a suitable landing site based on available hazard maps. 2.e. mission..The context of this case study.. propulsion and guidance constraints . Sim˜ es o from iteration to iteration (i. a final decision about the best place to land is achieved [19. R. The objective of this multicriteria dynamic decision making case study was to decide which was the best target site. 20.. This example illustrates interesting contributions for an efficient data preparation process.1 Background about Landing Site Selection Hazard avoidance includes three separate critical functions [12]: • Hazard mapping that estimates ground features based on imaging sensor data (camera or Lidar).A. • The site should be reachable with the available fuel. and creates hazard maps.3. Pais. and to dynamically adjust that choice as better information becomes available along the descent (by performing one or more re-targetings). decision step). Of particular relevance to an autonomous hazard avoidance system is the ability to select landing sites in real-time as the spacecraft approaches the planet’s surface.e. The goal of the site selection process is to provide an adequate target-landing site. we have shown good results of using a Fuzzy Multiple Criteria Decision Making approach to solve this problem [8. 2. during the spacecraft final descent. 12]. and L. • The site should be visible from the camera along the descent phase. we have lack confidence on available data and also have doubts about their accuracy). However. criteria): • The site should be safe in terms of maximum local slope.e.F. . and contains some intrinsic imprecision (i. and also summarizes the decision evaluation process with feedback [8. 12] to wrap-up the complete MCDDM approach. 12].3 Example in Practice The example used in this section is a simplification of a real case study [12]. for spacecraft landing on planets. • A robust guidance to reach the selected target. Ribeiro.C. After several iterations.54 T. we never discussed important aspects related with dynamically changing input data and how the data preparation process ought to be done.

g.1 Step 1. in the 8th iteration of the CRATERS dataset 2. . (1) Slope Map – provides the minimum angle to be feared by each pixel. in kg. (4) Fuel Map – provides the values. the spacecraft is continuously mapping and evaluating the terrain below it. Data Preparation in Dynamic Environments 2. that are necessary to reach the spot corresponding to each pixel. each with a matrix structure (512 × 512). It then feeds that information to the site selection component.3.Uncertainty in Dynamically Changing Input Data 55 In the final phase of descent.3 Representation of the Shadow criterion for all sites. Identification of parameters To illustrate the approach let us start by identifying the criteria (step 1 of data preparation Sec. which will in turn provide the navigation system with the landing coordinates. starting at around 2. We identified six inputs (hazard maps).2 Phase A.2. according to a set of criteria (e. 2.7 km above the surface. 0 240 100 210 180 200 150 120 300 90 400 60 30 500 0 100 200 300 400 500 0 Figure 2.3. At all times there is a degree of uncertainty with the evaluations of each site.3 depicts the shadow criterion).. in degrees. (3) Shadow Map – provides the values of light in a grey scale (0-255).2). Figure 2. that decreases (non-monotonically) as the spacecraft approaches the surface.2. (2) Texture Map – provides the variance value for each pixel.

C. meaning: 0 if the site is not reachable and 1 if the site is reachable.e. Ribeiro. this process tries to answer the question on how to transform input data to be appropriately usable in the decision process. the site. if criterion value is low our membership .. As mentioned. (6) Distance Map – provides the distance (in meters) between each pixel to the current target. and L. All other maps are provided in the same fashion except Distance. which is calculated. in the 8th iteration of the CRATERS dataset 2.4 Representation of the Texture criterion for all sites. Figure 2. Transformation Now let us deal with Step 2.3. An example of a Shadow map can be observed in Figure 2.3 and a texture map is shown in Figure 2. Hence. In our case we are trying to choose the best site for landing. R.e. our membership values have to be monotonically increasing towards the best criterion value.A.2. data transformation.4.2 Step 2. i. This transformation is two-folded: how to represent the variables and how to normalize the data. Pais. i. which best fulfils all criteria.F.. An important problem on how to fuzzify variables [2] is to choose the best topology for membership functions because we have to take into account our objective in the decision model.56 T. Sim˜ es o (5) Reachability Map – provides values between 0 and 1.

2552] but we consider the upper bound to be the maximum value of texture map (see Figure 2. to define the best morphology for criteria representation. Further. and α defines the range for the function plateau. For our example we used fuzzy sets [2] to represent and normalize our input variables. The variable domain is [0. After. The trapezoidal. is to ensure that data is both numerical and comparable. for representing the Texture map criterion the Gaussian membership function seems the most adequate function because some values.4) for all iterations until the current one. trapezoid and gaussian.5 shows the membership functions for iteration 23 of a sampling scenario. . we compared three types of membership functions more commonly used: trapezoidal. For example. Further. Figure 2. triangular and gaussian membership functions were build using the Equations 3. Moreover. triangular and Gaussian. have a too high membership value for triangular and trapezoidal functions when compared with the Gaussian and this can create “bias” in the decision results. respectively. upperBound]. consider that x belongs to the interval [0. upperBound − x Triangular(x) = upperBound −x2 Gaussian(x) = exp 2σ 2 (3) (4) (5) where upperBound2 . the membership functions for texture represent “Low Texture Values” because it is what we want to maximize. As mentioned. for instance 40. (6) 2 log(β ) and β is a parameter that satisfies the following condition: Gaussian(upperBound) = β . Gaussian penalizes high variable values and this is desirable because texture is a problem when values are high. 4 and 5. when in presence of both qualitative and quantitative variables.Uncertainty in Dynamically Changing Input Data 57 value is high and vice-versa. we compare them and select the most representative for usage with our spatio-temporal changing input data. σ2 = Concluding. Another important aspect to deal with in this step. To select which was the best topology for membership functions we used available test data to build three well-known functions (from the input data): triangle. we need to transform and normalize the data to compare them. ⎧ ⎪1 ⎨ if x c Trapzoidal(x) = upperBound − x ⎪ if x > c ⎩ upperBound − c where c = α × upperBound.

7]. 1 0. this aspect is not further discussed here. which is quite important since there are around 262.3.3 Step 3. Hence.58 T. R. which have membership value of zero.C. . Sim˜ es o 2. hence our choice for the proposed approach. Cleaning In this illustrative case we start by setting to Zero all pixels in the 512 × 512 input matrix. Moreover. 2.3.5 0. After.000 original alternatives (image pixels corresponding to sites). Ribeiro. we used the the conjuntive method [7] with the min operator. this process allows reducing the number of alternatives (search space).2. and L. Conjunctive methods are normally used for discarding unacceptable alternatives [21. in the texture example all alternatives that have texture value higher than 68 are discarded.1 0 0 10 20 30 40 50 60 70 Figure 2.7 0.3 0.A. as other dynamic decision problems might face. With this method we can easily integrate all transformed and cleaned criteria (hazard maps).9 0. we proceed with other types of “cleaning”. Since for our case study we do not have missing records there is no need to define a strategy to fill the gaps.8 0.5 Membership functions for “Low Texture”.F.4 0. Pais. and obtain the set of 3-D coordinates of possible candidate alternatives.2 0.6 0. Fusion For fusing the different input parameters to obtain the common set of candidate alternatives for target landing. for instance.4 Step 4. corresponding to the 512 × 512 matrices.2.

In this case we propose a filtering function that combines metrics to deal with both lack of accuracy . whereas the black areas are those alternatives that have been discarded on the fusion process. The objective of this task is to deal with intrinsic imprecision/uncertainty found. (c) If there is one value lower than the minimal acceptable level. The procedure for this method is as follows: (a) Decision maker specifies a minimal acceptable level for each criterion.2.3. (b) For each alternative determine if all criterion values exceed the respective minimal acceptable threshold specified by the decision maker. the respective alternative is rejected. It can be observed with these two steps we clearly reduced the number of alternatives to be evaluated. in the input values.6 we depict the final set of alternative pixels that remained after the cleaning and fusion processes took place.5 Step 5.Uncertainty in Dynamically Changing Input Data 59 Figure 2. Filtering Uncertainty Step 5 deals with filtering uncertainty from data. 2.6 Image of all alternatives. Formally: Ai is acceptable ⇔ ∀ j : xi j > l j where l j is the minimal value for the criterion C j In Figure 2. at each iteration.

20 + 5]. respectively) will be taken into account in the decision model using the following expression: Filter(xi j ) = wc j × 1 − max x∈[a. The confidence value wc j = 0.9. and represent by aci j = Filter(xi j ). For each xi j we therefore obtain a filtered value that we call accuracy&confidence membership value. b] is defined as follows: ⎧ ⎨min(D) if xi j − ai j min(D) a= ⎩x − a ij i j if xi j − ai j > min(D) ⎧ ⎨xi j + ai j if xi j + ai j max(D) b= ⎩max(D) if x + a > max(D) ij ij (8) where ai j is the accuracy associated to criterion j for site i. b] where ai j = 0. the process whereby they are obtained may be made more reliable.34. Sim˜ es o and lack of confidence in the collected input values.F. it allows dealing with two types of imprecision: lack of accuracy in the data and lack of confidence in the data. for all sites i. Moreover. We believe this is a simple and easy way to tackle the problematic of how to handle imprecision in data in dynamic decision . The accuracy is given as percentage of criterion value and confidence belongs to the unit interval. and [a. Ribeiro. an accuracy of 90 % for slope means that each slope value belongs to the interval [a. Though the criterion may always have a degree of inaccuracy associated with values read. The membership function representing the fuzzy set “Low Slope” is shown with the circle shaped markers. xi j is the value of jth criterion for site i. In this case. We can observe the result of using the Filter function in Figure 2.60 T. For instance. and shown with the triangle shaped markers is the resulting function after using the Filter.7.52 and a filtered (with accuracy&confidence) membership value of 0.C. and D is the variable domain. the decision maker is indicating there will be room for a 20 % improvement in the determination of xi j . In the example. and L. this formulation enables dealing with imprecision in the collected input hazard maps.9 confidence value means that we have a confidence on the [a. b] interval of 0. an accuracy value of ai j = 5 is used.7 we can observe that for a texture value of 20 we have a membership value for “Low slope” of 0. the accuracy and confidence parameters (ai j and wc j .9 × xi j . Formally. For example. in Figure 2.80 indicates in general how confident one is with the quality of the process that generates the input values. R.b] μ (x) − μ (xi j ) × μ (xi j ) (7) where wc j is the confidence associated to criterion j. As it can be observed.A. That means that a value read as 20 may actually come from anywhere in the interval [20 − 5. μ is a membership degree in a fuzzy set. Pais. On the other hand. a 0.

(9).52 0. variance: 303.6 0. and resulting criteria weight 0. the linear weight generating function of Equation (10) is applied. Weighting Criteria This step is the final one for completing the data preparation in MCDDM. However.4 0. All other variables and parameters have the same meaning as in Eq.00.33) y * weight 0. in our approach to deal with accuracy we assumed that our decision maker has a pessimistic attitude towards the problem.85] 50 Figure 2.00.2. .Uncertainty in Dynamically Changing Input Data 61 1.80. confidence: 0.80 weight(y. close to zero an optimistic attitude).3. Hence. of each alternative for each criterion. Filterα (xi j ) = wc j × 1 − α max x∈[a. (7). 1] is a parameter that reflects the attitude of the decision maker (α value close to one represent a pessimistic attitude. models. membership function: low texture fuzzy_gaussian (mean: 0.2 0. Hence. 52. 2.29) [in]accuracy: 5. Given the final membership value. Equation (7) can be seen as a particular case of a more general approach described by Eq. The different stages through which an alternative with an initial evaluation of 20 on the criterion passes is also shown. beta: 0.0 membership value.34 0.7 Membership function representing the set of values with low texture.23 40 30 20 10 membership functions' range: [0. alpha: 0.b] μ (x) − μ (xi j ) × μ (xi j ) (9) where α ∈ [0. we propose to compute the weights to express the relative importance of each criterion. before and after applying the accuracy and confidence factors.00 0. It deals with defining the relative importance for each criterion. to ensure that a poorly satisfied criterion is not “emphasized” by a pre-determined high importance.67 0.6 Step 6. and the criterion’s final contribution to quality calculated.00. In dynamic processes it is rather important to consider the satisfaction level.8 criterion: texture.

The logic of these weighting functions (see Eq. . (11)) is that the satisfaction value of a criterion should influence its assigned relative importance. The β parameter provides the slope for the weighting functions.62 T. if the car is quite expensive the final decision result should be less than the simple multiplication of weight and satisfaction value. . (b) Important (I= 0. . 1] and aci j is the accuracy and confidence membership value of jth criterion for site i.8). . Sim˜ es o using the following linear weight generating function [14. In Figure 2. Ribeiro.A. which will depend on the criterion at hand.C. (d) Low importance (L= 0. Note that aci j corresponds to the filtered value of xi j from Eq.4). 15] : L j (aci j ) = α j 1 + β j aci j 1 + βj (10) where α j . (b) Medium (M= 0. In this work this parameter has the following values: (a) High (H= 1). The α parameter provides the semantics for the weighting functions as follows: (a) Very Important (VI= 1). Pais. (7). and L.6667). R. acim ).F. The definition of these weighting functions morphologies is given by the parameters α and β . (e) Very Low importance (VL= 0.6). (10). (c) Low (L= 0. and L j is a linear weight generating function presented in Eq. (d) Null (N= 0). the line with the square shaped markers depicts the behavior of the weight generating function for the texture criterion. For example.3333). depending on the altitude of the spacecraft in relation to the surface. W (ac) = L j (aci j ) k=1 ∑ Lk (acik ) m (11) where ac = (aci1 . β j ∈ [0.2). if we are buying a car and the price is a “very important” criterion. (c) Average importance (A= 0. with the logic that higher values of β means higher slope. . For our case study we considered that the relative importance of criteria has different morphologies for each criterion.7.

we have a somewhat low membership value of . for high altitudes (> 1000 m. In Figure 2.8) with Low penalty (β = 0. Table 2.8 1.0 The rational above is expressed by the α parameter.33) 0.Uncertainty in Dynamically Changing Input Data 63 For example. we depict the proposed values for α and β parameters for this phase.6 0.6 membership value Figure 2.4 0.667 Slope 1 0. reachability and slope are all “very important” in this phase. b: 0.33) shadow (a: 0.8 Weighting functions.7 we see the weight generating function for the texture criterion.00. a lower satisfaction for criteria should be much more penalized in the case of reachability. b: 0.25) reachability (a: 1. b: 0.8 0. < 2000 m.80.00) texture (a: 0. 0. b: 0.0 0.00) scientific interest (a: 0. b: 0.40. 250 historic set size). as depicted in the figures for each phase.6 0.333 ScIn 0.8). In the example shown. a little more for fuel and less for slope.25 1.1 α and β parameters. b: 0.4 0 Text 0.2 0.8 criterion importance (weight) 0. and their respective plots (Figure 2. the rationale for “very important” is that although fuel. when xi j = 20.333 Shad 0.20.4 0.111 Dist 0. For example.0 0.00.60. Fuel Reach 1 0. This is an Important criterion (α = 0. b: 0.0 slope (a: 1.67) distance (a: 0.11) fuel (a: 1.33) for lower satisfaction values. while the β parameter provides more or less penalties for lower satisfaction values in the criteria.2 0.2 0 α β 1 0.00.

avoids having to define different β parameters for each weighting functions.64 T.F.. having in mind the evaluation alternatives when feedback is involved (phase B in Sec. by combining R(T ) with H(T − 1) with a hybrid uninorm operator [8]. R. in this section we just present a summary of required activities to solve the illustrative example. Details about contributions for this topic can be seen in our previous works [8. 2. it means the historical information for a specific site “remembers” its past behavior plus the actual rating.3 Phase B. The K size changed 3 times in our example because it depended on the spacecraft altitude (distance from surface). This set K constitutes the feedback information that is made available for the next iteration. When there are no more iterations the decision process stops and the best alternative is the one with the highest value. and L. weighted with their relative importance. from the evaluated alternatives. down from the criterion’s base importance of 0.3. (iii) Historical process – This process determines a subset of good alternatives. 12]. There are many methods that can be used for this aggregation but a simple and sound method is the Weighted Sum rating method [1] and this is the one we used in our example. (ii) Dynamic evaluation process – This process refers to the aggregation of historical information with the alternatives rating and then their respective ordering to obtain a set of ranked alternatives.2. The three steps for the evaluation process in our example are: (i) Rating process – This process refers to the classification of each alternative regarding all criteria. We considered different sizes for (K) depending on different phases of the decision process. 2. 20. [22].8. More details about this weighting process in the case study can be seen in Ref. .A. This process is performed at each iteration of the decision process.67 is attached.C. This option simplifies the weight assignment process but it is less accurate.34 to which a weight of L j (aci j ) = 0. An important point to highlight is that the parameter values proposed for the phases (or iterations) can be tuned for any other dynamic decision applications. Since this process takes place after evaluation. Ribeiro. Sim˜ es o aci j = 0. i. Decision Evaluation Process with Feedback Since our scope is dynamically changing input data.2). Pais. to be considered as feedback information from iteration to iteration.e. A simplification of the weighting functions could be to define “a priori” a slope for each importance.

P. In addition. each with several contributions on how to perform them. 432–432. since both fields require several decision steps and feedback information. requiring feedback from iteration to iteration. L. Magdalena. Reynaud.Computational Intelligence in Complex Decision Systems 65 2. W. (2004). Introduction to System Dynamics Modeling with DYNAMO. Ribeiro. In eds. and J. Methods and Applications. 78(2). Psychological Review. A. R. Richardson and A. Our scope was spatial-temporal multiple criteria decision problems.-L. SpringerVerlag. and S. We believe. (1996). [8] T. Cios. vol. [4] J. Proceedings of the 12th International Conference on Information Processing and Manage- . Ross. Fuzzy Multiple Attribute Decision Making. detection and isolation (FDI) problems. Decision field theory: a dynamic-cognitive approach to decision making in an uncertain environment. Number 375 in Lectures Notes in Economics and Mathematical Systems. Data Mining Methods for Knowledge Discovery. 155–181. Busemeyer and J. W. Productivity Press. Ribeiro. Swiniarski. Fuzzy multiple attribute decision making: a review and new preference elicitation techniques. Multi criteria Decision Making Method: a Comparative Study. 100. (1998). Finally.-J. Kluwer Academic Publishers. Kluwer Academic Publishers. Hwang. both at the level of better operators or new search methodologies to reduce the decision process are foreseeable. The focus of the work was on the first phase. Portland. John Wiley & Sons. for which we proposed six steps. and R. [7] S. Pedrycz. [2] T. [3] R. these are interesting contributions to improve decision support approaches for MCDDM problems. dealing with uncertainty in dynamically changing input data was a challenge and future contributions to improve the evaluation process. (1993). Verdegay. Y. L. C. (2002). Townsend. 44. Potential applications of this work could be on medical diagnose decision support and/ or fault. divided into two main phases: data preparation and dynamic evaluation with feedback. We specifically proposed new contributions for dealing with uncertainty in dynamically changing input data and also for dealing with dynamic changes in criteria importance. Oregon. Ojeda-Aciego. comparison studies between different methods and technologies may also prove to result in improved performance and accuracy for the selection of alternatives in presence of multiple criteria in dynamic decision processes. Triantaphyllou. (1992). Fuzzy Sets and Systems.4 Future Research Directions We presented a general architecture for dynamic multiple criteria decision problems. Applied Optimization Series. Devouassoux. (1981). data preparation. L. Fuzzy Logic with Engineering Applications. Pugh III. Bibliography [1] E. [6] K. Chen and C. M. Pais. Dynamic ranking algorithm for landing site selection. [5] G.

http://dx. 111– 120. Portugal (September. (2008). Y.C. Coimbra. 608–613. A. J. 28 (6). Melbourne. Aggregation with generalized mixture operators using weighting functions. R.org/10. 2004). Rybalov.gimac. T. pp. D. In GNC 2008: 7th International ESA Conference on Guidance. IEEE Transactions on Systems. R. World Scientific Proceedings Series on Computer Engineering and Information Science. Journal of Mathematical Psychology.F. M. D’hondt. Handling missing values in solar array performance degradation forecasting. Kluwer Academic Publishers. Townsend. Singapore (Aug. Ribeiro. Hazard avoidance developments for planetary exploration. 1. R. A. Pais. Devouassoux. Busemeyer. and L. Pais. Social selection based on multidimensional criteria. ISBN 978-981-279-946-3. P. 10(1). R. C. Navigation & Control Systems. Mdft account of decision making under time pressure. Contributions for dynamic multicriteria decision making models. Sept-Oct. Jean-Marius and S. R. Ruan. Ribeiro. and T. Rybalov.pdf. H. 108(2). Jonniaux. Portugal. Dawes. Man and Cybernetics. Devouassoux.es/ipmu08/proceedings/papers/080-PaisEtAl. R. and A. (1998). Psychol Rev. Y. and S. Ribeiro. R. ISSN 0165-0114. Diederich. Psychonomic Bulletin & Review. Mart´nez. A. Fuzzy Sets Systems. 104–109. 145 (2). Lisbon. Pais. Yager and A. 68(1). T. (2001). Computational Intelligence in Decision and Control – Proceedings of the 8th International FLINS Conference. and E. Ribeiro. Integrated vision and navigation for a planetary lander. (2003). A. and J. 757–769. Reynaud. Journal of Abnormal and Social Psychology. Pereira and R. Dynamic stochastic models for decision making under time constraints. (2003). . Sim˜ es o [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] ment of Uncertainty in Knowledge-Base Systems (IPMU). (1964).uma. In International Fuzzy Systems Association World Congress – IFSA. Regions rating for selecting spacecraft landing sites.1016/0165-0114(95)00133-6. A. Part B. 370–92. (2009). J. G. Roe. E. Montero. J. C. Viana. Pereira.66 T. (1996).. In 49th International Astronautical Congress. European Journal of Operational Research. Diederich. N. World Scientific.doi. ISBN 978-84-612-3061-7. Fuzzy Sets and Systems. R. (1997). 260–274. pp. A. 157–166. Australia. 2008). Zimmermann. Generalized mixture operators using weighting functions: A comparative study with wa and owa. In eds. Marques Pereira. 43–58. Pais and R. Strandmoe. Multialternative decision field theory: a dynamic connectionist model of decision making. Donati. S. Yager and A. (2001). ı Kerre. vol. 41(3). R. (2008). L. (2003). In Proceedings of the 15th Mini-EURO conference on Managing Uncertainty in Decision Support Models (MUDSM 2004). 329–342. Uninorm aggregation operators. URL: http://www. T. Ribeiro. 1039–1044. 80(1). R. Full reinforcement operators in aggregation techniques. Ribeiro and R. Ribeiro.A. Fuzzy Set Theory–and Its Applications. 137(1). R. Lu. (1998). Reynaud. A.

. linear programming (LP) has been one of the operation research techniques. D. which gives the conceptual and theoretical framework for dealing with complexity. Atlantis Computational Intelligence Systems 2. Guo [10. 240-8501 Japan E-mail: guo@ynu. 3. Two different possibility distributions. Zadeh [15]. Tanaka and P. 11]. none of these methods is suitable for an overall analysis of the effects of imprecision in parameters. But the strict requirement of LP is that data must be well defined and precise.Chapter 3 Decision Making under Uncertainty by Possibilistic Linear Programming Problems Peijun Guo Faculty of Business Administration. Bellman and L. Yokohama National University. © 2010 Atlantis Press/World Scientific 67 . i. H. Fuzzy solutions are used to approximate the feasible region of decision variables left by the given fuzzy satisfaction levels. However. L. Hodogaya-Ku.A.1 Introduction For some considerable time.1007/978-94-91216-29-9_3. It can be said that a fuzzy solution associated with an upper possibility distribution leaves more rooms than the one associated with a lower possibility distribution. which is often impossible in real decision problems.A.jp A decision making problem in an upper decision level is described in this chapter by a set of fuzzy satisfaction levels given by decision makers. Another way to handle imprecision is to model it in stochastic programming problems according to probability theory. with the help of sensitivity analysis and parametric programming. imprecision and vagueness R. Ruan. Computational Intelligence in Complex Decision Systems.ac. Yokohama.e. DOI 10. The traditional way to evaluate any imprecision in the parameters of an LP model is through post-optimization analysis. symmetrical triangular possibility distributions and exponential possibility distributions are considered and their upper and lower possibility distributions are obtained. Zadeh [1]. Tokiwadai 79-4. which has been widely used and got many achievements in both applications and theories.E. A third way to cope with imprecision is to resort to fuzzy set theory.

i = m1 + 1. Zimmermann [14]. Mathematically. i. This means that in an uncertain environment.68 P. [9] proposed several approaches to obtain various fuzzy solutions to enable a decision maker select one preferable fuzzy decision among the obtained ones. For the lower level decision problems. .e. . . As an extension of that idea. fuzzy decisions are often required in the upper level because giving some rooms for lower level is necessary in real decision problems. H. . symmetrical triangular possibility distributions and exponential possibility distributions. Rommelfange [7]. H. Sakawa [8]. It can be said that a fuzzy solution associated with an upper possibility distribution leaves more rooms than the one associated with a lower possibility distribution.. The fuzzy decisions obtained in the upper level will be sent to the lower level where crisp decisions will be made within these fuzzy decisions. (1) . Inuiguchi et al. Fuzzy solutions are used to approximate the feasible region of decision variables left by the given fuzzy satisfaction levels. . a kind of decision problem in an upper decision level is described by a set of fuzzy satisfaction levels given by decision makers. ······ bi1 x1 + · · · + bin xn ≈+ Ci . On the other hand. Guo Generally speaking.-J. Guo et al. From the bottom to the top. and LP technique is used to obtain the largest possibility distribution of decision variables. 3].. By contrast. Assuming fuzzy decision variables are governed by two different possibility distributions. In this chapter. m. Tanaka et al. the objectives of decisions become large and complex because of the inherent fuzziness of decision problems. m1 . 3. M. Asai [13] initially proposed a possibilistic linear programming formulation where the coefficients of decision variables are crisp while decision variables are fuzzy numbers. decision makers need to make a crisp decision under a fuzzy environment as widely researched by conventional fuzzy programming problems M. their upper and lower possibility distributions are obtained with different fuzzy degrees. [4. H. P. a kind of upper level decision problems can be described as follows: bi1 x1 + · · · + bin xn ≈− Ci . Guo et al. Tanaka and K.2 Fuzzy Decisions in Possibility Linear Programming Problems Decision problems are often described in term of a multilevel functional hierarchy. . the coefficients of decision variables are fuzzy numbers while decision variables are crisp ones. i = 1. 5]. Tanaka et al. . a crisp decision is made to meet some benefit criteria. [2. This chapter aims to obtain possibility distributions of fuzzy decision variables to reflect the fuzziness of decision problems. H. in fuzzy linear programming models.

Ci is a fuzzy number (i = 1. As a result. (2) (3) (soft equal) become “=” (hard equal) and correspond- ingly Ci becomes a crisp value in (1). Because the feasible regions left from possibilistic constraints are conflicting and partially inconsistent in nature. each possibilistic constraint leaves some feasible region for decision variables. However. Each fuzzy number Ci in (1). . The possibility distribution of A is defined (5) . .Decision Making under Uncertainty by Possibilistic Linear Programming Problems 69 where the symbols ≈− and ≈+ .1 Triangular Possibility Distributions of Fuzzy Decision Variables Assume that Ci is a symmetrical triangular fuzzy number denoted as (ci . if “Yi ≈− Ci ” and “≈+ ” “≈− ” and “≈+ ”. . can be regarded as a fuzzy goal from each department of a corporation given by its manager with only considering the benefit of his own department where the center of fuzzy number is an ideal point and the spread of fuzzy number represents some tolerance. for example. . . represent two kinds of “approximately satisfy”. The feasible region of fuzzy constraint is characterized by a symmetrical tri- angular fuzzy decision vector A = [A1 . . di )T where ci and di are its center and spread with the condition ci − di x = [x1 . the problem considered now is to obtain an integrated feasible region of decision vector. a reasonable plan should be feasible for constraints from all of such departments. . . . let us consider how to obtain the feasible region from possibilistic constraints described in (1). In what follows. . . n) and bi j is an associated crisp coefficient. j = m1 + 1. In other words. . Let us rewrite the fuzzy constraints (1) as follows. a − r x i i i ri ΠAi (xi ) = ⎩ 0. the right-hand sides of the model (1) describe the desires of decision makers and the left-hand sides are some technical conditions to achieve them. . i = 1. defined later. m1 . which satisfies all of possibilistic constraints. A2 ]t . m) representing a fuzzy satisfaction level offered by decision makers. and x j is a decision variable ( j = 1. . . . b j1 x1 + · · · + b jnxn ≈+ C j . . . . . . ri > 0 (4) ith 0. where the subscripts i and j correspond to symbols Obviously. respectively. . bi1 x1 + · · · + binxn ≈− Ci . m. otherwise ai + ri . we can rarely obtain the feasible solutions for the case of m > n.2. xn left from the ]t as follows: ΠA (x) = ΠA1 (x1 ) ∧ ΠA2 (x2 ) ∧ · · · ∧ ΠAn (xn ) where ⎧ ⎨ 1 − |xi − ai | . 3. . .

. . m1 . . . . m1 . where center yi and spread si are bti a and bti r. . . . x2 . . . i = 1. rn ]t (ri 0. (7) Constraint Yi ≈− Ci is explained as “Yi is a little bit smaller than Ci ”. . xn ]t becomes a possibilistic vector A = [A1 . ai and ri are the center and spread of fuzzy number Ai . where h is a predetermined possibility level by decision makers. (8) (9) (10) 0. . A is simply denoted as (a. m1 . . bti r)T .1 Explanation of Yi ≈− Ci with degree h. with a = [a1 . . i = 1.1 which shows that “Yi is a little bit smaller than Ci ”. It can be understood that the higher h is. . . . A graphical explanation of Yi ≈− Ci is given by Figure 3. respectively described in (8) and (9). i = 1. . . an ]t (ai 0. defined . yi − (1 − h)si yi + (1 − h)si yi ci − (1 − h)di. . and bi = [bi1 . A2 ]t . . i = 1.70 P. . denoted as Yi ≈− Ci . but the center of Yi is confined to be larger than the left endpoint of Ci shown by the dotted line described in (10). the constraint Y j ≈+ C j is explained as “Y j is a little bit larger than C j ”. . r)T . . The left-hand side of the fuzzy constraint (2) or (3) is written as Yi = bi1 x1 + · · · + binxn . . Similarly. respectively. defined by the follow- ci − (1 − h)di. . . bin ]t ing inequalities. . the possibility distribution of fuzzy set Yi can be obtained by the extension principle as ΠYi (y) = (yi . . i = 1. . . . Guo Figure 3. . . respectively. n). . . si )T = (bti a. . . means that the left and right endpoints of the h-level set of Yi are smaller than the left and right ones of Ci . . ci + (1 − h)di. . n) and r = [r1 . the stricter constraints are. (6) Since x = [x1 .

.. m1 . . . . . . . s. m1 . i = 1. t. . m1 . . . m1 . j = m1 + 1. . c j + (1 − h)d j . . It implies that the center vector must exist in the region formed by (20) and (21). . . m1 .Decision Making under Uncertainty by Possibilistic Linear Programming Problems 71 by the following inequalities: yi − (1 − h)si yi + (1 − h)si yi ci − (1 − h)di. . . . . m1 . . i = 1. and fuzzy constraint (3) can be rewritten as: btj a − (1 − h)btj r btj a + (1 − h)btj r btj a c j − (1 − h)d j . . . j = m1 + 1. . it is easy to understand that the center vector a should comply with the following constraints. (14) (15) (16) ci − (1 − h)di. . . It follows from the above definitions that fuzzy constraint conditions leave some feasible region for decision variables described by inequalities (8)-(13). i = 1. . . Considering (7). ci + (1 − h)di. . . . The following LP problem is used to find out a candidate of the center vector.. . . i = 1. . . . i = 1. max g(a) a (22) (ci − (1 − h)di)(1 + β ). ci (1 − β ). .. . . i = 1. . . (20) (21) c j + (1 − h)di. fuzzy constraint (2) can be rewritten as: bti a − (1 − h)bti r bti a + (1 − h)bti r bti a ci − (1 − h)di. . (11) (12) (13) ci + (1 − h)di. i = 1. ci + (1 − h)di. . . m. m. . . j = m1 + 1. i = 1. m. m. . . . . . . bti a bti a btj a btj a β a 0. . m. j = m1 + 1. . . (17) (18) (19) c j + (1 − h)d j. j = m1 + 1. where the objective function g(a) is given by decision makers to characterize their preference for selecting the center vector. m1 . m1 . . m. . c j (1 + β ). ci − (1 − h)di cj btj a bti a ci . (c j + (1 − h)d j )(1 − β ). i = 1. From constraints (14)-(19). 0. m1 .. which is regarded as a reference point in the feasible . . j = m1 + 1.

. m. t. ci + (1 − h)di − bti a0 . Parameter β 0 is used to reduce the original feasible region to guarantee the obtained center vector inside the region formed by (20) and (21). j = m1 + 1. j = m1 + 1. where w is the weight vector of the lower spread vector rl . . . m. with the condition Πu Πl . . m1 . . . . t. . upper and lower possibility distributions denoted as Πu (x) and Πl (x). c j + (1 − h)d j − btj a0 . . After determining one center vector. i. The upper possibility distribution of A corresponds to the upper possibility distribution of Y and the lower possibility distribution of A corresponds to the lower possibility distribution of Y . . −(1 − h)bti rl (1 − h)bti rl −(1 − h)btj rl (1 − h)btj rl rl 0. m. .. . . .2 gives a graphic explanation of upper and lower possibility distributions of Y restricted by the possibility constraint Y ≈− C. c j − (1 − h)d j − btj a0 . c j − (1 − h)d j − btj a0 . With considering a center vector a0 . j = m1 + 1. i = 1.e. m. . Pawlak [6]. .. . the problem for finding out the lower spread vector of A is formalized as the following LP problem. max wt ru ru (24) ci − (1 − h)di − bti a0 . . respectively. the problem for finding out the upper spread vector of A is formalized as the following LP problem. ci + (1 − h)di − bti a0 .. s. Similarly. . min wt rl rl (23) ci − (1 − h)di − bti a0 . −(1 − h)bti ru (1 − h)bti ru −(1 − h)btj ru (1 − h)btj ru ru 0. . i = 1. s. . . . we investigate two kinds of possibility distributions of fuzzy vector A. Guo region of decision variables. . m1 . which have some similarities with the concept of rough sets Z. Figure 3. . . . m1 . . denoted as a0 . m1 . i = 1. i = 1. . j = m1 + 1. . .72 P. c j + (1 − h)d j − btj a0 . It shows that the region covered by the lower possibility distribution of Y is completely included by the h-level set of the given possibility information C while the region covered by the upper possibility distribution of Y completely includes the h-level set of the given possibility information C under the definition of Y ≈− C. The upper and lower possibility distributions of Y are the upper and lower bounds of Y restricted by Y ≈− C for a fixed center of Y .

j = m1 + 1.rl (25) ci − (1 − h)di − bti a0 . . . −(1 − h)bti rl (1 − h)bti rl −(1 − h)btj rl (1 − h)btj rl −(1 − h)bti ru (1 − h)bti ru −(1 − h)btj ru (1 − h)btj ru ru − rl rl 0. . ri )e . . Guo et al. ci + (1 − h)di − bti a0 . . . 0. i = 1.2 Illustration of upper and lower possibility distributions in Y ≈− C. m. . . . j = m1 + 1. . . . c j − (1 − h)d j − btj a0 . m. . i = 1. .Decision Making under Uncertainty by Possibilistic Linear Programming Problems 73 Figure 3. c j + (1 − h)d j − btj a0 . the problem for obtaining the upper and lower spreads of A simultaneously is introduced as follows P. ci + (1 − h)di − bti a0 .2 Exponential Possibility Distributions of Fuzzy Decision Variables Assume that Ci is characterized by the following exponential possibility distribution ΠCi (x) = exp(−(x − ci )2 /ri ). . . s. . m1 . . respectively. . [5]: max wt (ru − rl ) ru . j = m1 + 1. assume that x = [x1 . .. . . simply denoted as (ci .. . Likewise. xn ]t is characterized by an exponential . With considering the inclusion relation between the upper and lower possibility distributions. m1 . m. . . c j − (1 − h)d j − btj a0 .2. i = 1. m1 . . i = 1. m1 . . j = m1 + 1. .. where ci (26) 0 and ri > 0 are the center value and the spread of Ci . c j + (1 − h)d j − btj a0 . m. 3. . . . t. . . ci − (1 − h)di − bti a0 .

. the fuzzy constraint conditions (2) and (3) can be rewritten as follows: bti a − bti a + bti a btj a − (− ln h)bti DA bi (− ln h)bti DA bi ci − (− ln h)ri . . Guo possibility distribution as ΠA (x) = exp −(x − a)t D−1 (x − a) . . defined by the following inequalities: yj − yj + yj (− ln h) · w j (− ln h) · w j cj + cj − cj + (− ln h) · r j . and DA is a symmetrical positive definite matrix.74 P. cj − (− ln h)r j . . bti DA bi )e . an ]t (ai matrix DA is called a possibility distribution matrix. . (− ln h) · ri . (− ln h) · r j . The constraint Y j ≈+ C j is explained as “Y j is a little bit larger than C j ”. Ishibuchi [12]: ΠYi (y) = (yi . Considering (6) and (27). As a result. wi )e = (bti a. respectively described in (29) and(30). where h is a predetermined possibility level by decision makers. Tanaka and H. bin ]t . . The constraint Yi ≈− Ci is defined by the following inequalities: yi − yi + yi (− ln h) · wi (− ln h) · wi ci − ci − ci + (− ln h) · ri . denoted as DA > 0. but the center of Yi is confined to be larger than the left endpoint of Ci described in (31). (− ln h)ri . (35) (36) (37) (38) (− ln h)btj DA b j . (32) (33) (34) (− ln h) · r j . the possibility distribution of the fuzzy set Yi can be obtained by the extension principle as follows H. Yi ≈− Ci means that the left and right endpoints of the h-level set of Yi are smaller than the left and right ones of Ci . . . A simply denoted as ΠA (x) = (a. where bi = [bi1 . Similarly. where a = [a1 . DA )e . ci − ci + (− ln h)ri . The (− ln h) · ri . (29) (30) (31) (28) (27) 0) is an n-dimensional center vector.

with the condition that Πu Πl holds. the vector satisfying (45) and (46) is probably qualified as a center vector. Du Dl > 0 should be satisfied. (44) From the constraints (35)-(40).Decision Making under Uncertainty by Possibilistic Linear Programming Problems 75 btj a + btj a where DA > 0. Since (45) and (46) are necessity conditions for a being a center vector but not sufficiency conditions.. Πu (x) and Πl (x) are represented as Πu (x) = exp −(x − a)t D−1 (x − a) .. the objective function for obtaining the upper possibility distribution of X is introduced as max Du i=1.m ∑ δi (bti Du bi ). u Πl (x) = exp −(x − a)t D−1 (x − a) . upper and lower possibility distributions denoted as Πu (x) and Πl (x). (43) 0 is a weight of the where m is the number of fuzzy constraint conditions of (1) and δi of X is introduced as min Dl spread of Yi .. (41) (42) The upper possibility distribution of X corresponds to the upper possibility distribution of Y and the lower possibility distribution of X corresponds to the lower possibility distribution of Y . Similarly. i.. we know that the center vector a should comply with the following constraint conditions: ci − cj (− ln h)ri btj a cj + bti a ci (45) (46) (− ln h)r j .. which imply the center vector must exist in the region formed by (45) and (46).e. cj + (− ln h)r j .m ∑ δi (bti Dl bi ).. respectively. the objective function for obtaining the lower possibility distribution i=1.. The following LP problem is used to find out a candidate of the center vector of Πu (x) and Πl (x) in the . (− ln h)btj DA b j cj + (− ln h)r j ... we investigate two kinds of possibility distributions of the fuzzy vector X. Because bti DA bi is the spread of fuzzy number Yi . l Since Πu Πl is required. (39) (40) Here.

Guo middle of the feasible region formed by (45) and (46). . constraints (45) and (46) are just necessity conditions for a0 being a center vector... The parameter β is used to guarantee the obtained center vector inside the region formed by (45) and (46)... bti a bti a btj a btj a ci − (− ln h)ri (1 + β ). (− ln h)ri . t. c j (1 + β ). As mentioned before.76 P.. (47) β where the objective function g(a) is used to characterize the preference of decision makers for selecting some center vector regarded as a reference point in the feasible region of decision variables. ci (1 − β ). Otherwise. a 0. After finding out the center vector a0 .. The obtained optimal solution of (47) is denoted as a0 ... 0. DA )e .. a0 should be checked whether there is an admissible set in the following constraint conditions. Assume that a0 is a valid center vector. ci − ci + (− ln h)ri . Thus.. (− ln h)btj DA b j (− ln h)btj DA b j cj − cj + (− ln h)r j .... max g(a) a s. . If there is an admissible set in (48)... Then. the possibilistic programming problems for finding out the upper and the lower possibility distributions can be formalized as the following two crisp optimization (− ln h)bti DA bi (− ln h)bti DA bi . other feasible solutions around it are investigated assuming that they are governed by a possibility distribution (a0 . cj + (− ln h)r j (1 − β ). (48) ... (− ln h)r j ... . The parameters β and h should be changed to obtain a suitable center vector.. a0 is qualified for a center vector. bti a0 − bti a0 + btj a0 − btj a0 + DA > 0.

. (− ln h)r j ... bti a0 − bti a0 + btj a0 − btj a0 + Du > 0. (− ln h)ri ...... Guo et al.m ∑ δi (bti Du bi ) . the matrices Du and Dl will be solved by a primary problem and an auxiliary problem as follows P. (− ln h)bti Dl bi (− ln h)bti Dl bi . [4]... (50) s... max Du i=1. (49) cj − cj + (− ln h)r j . (− ln h)ri ..... ci − ci + (− ln h)ri . (− ln h)btj Du b j (− ln h)btj Du b j .m ∑ δi (bti Dl bi ) ... bti a0 − bti a0 + btj a0 − btj a0 + Dl > 0... min Dl i=1.. In order to ensure the inclusion condition Πu Πl ....... (− ln h)r j .... cj − cj + (− ln h)r j ..... ci − ci + (− ln h)ri .Decision Making under Uncertainty by Possibilistic Linear Programming Problems 77 problems P. (− ln h)bti Du bi (− ln h)bti Du bi . [4]. Guo et al... t. t. . s... (− ln h)btj Dl b j (− ln h)btj Dl b j .

bti Du bi bti Du bi btj Du b j btj Du b j Du > 0. we set the spectral decomposition of Du as follows: Du = PΛPt . If the obtained matrix satisfies Du > 0 then it is a possibility distribution matrix. (btj a0 + ( (− ln h) · r j − c j ) /(− ln h).. . (bti a0 + ( .. . (52) where |di j | = |di j + − di j − | with di j + After finding out the positive matrix Du .1 is added to (51) by which (51) becomes an LP problem. we firstly solve the problem (51) without the constraint condition Du > 0. . n) 0.. Thus.. . it also means that a0 is qualified for a center vector. Guo et al. 2 s. [4])... ⎢ .. ⎥ .. ⎥. (53) where normalized eigenvectors are the columns of the matrix P and Λ is the diagonal matrix formed by the corresponding eigenvalues as follows: ⎤ ⎡ λ1 0 · · · 0 ⎥ ⎢ ⎢ 0 λ2 · · · 0 ⎥ ⎥ ⎢ ⎢ . .78 P. . In order to obtain Du easily.. .. Otherwise a sufficient condition for Du > 0 shown in Theorem 3. j j 0 and di j − (i = 1..1 (P.. (− ln h) · ri − ci )2 /(− ln h). solving (51) is also a procedure for checking a0 by (48). (− ln h) · ri + ci − bti a0 )2 /(− ln h).... . If a positive matrix Du can be obtained with the obtained vector a0 . t. ...m ∑ δi (bti Du bi ) . ⎦ . Guo Primary problem: max Du i=1. 0 0 · · · λn (54) . Given a real symmetric matrix [di j ] ∈ Rn×n .. j=i ∑ n (di+ + di−).. . Theorem 3. ⎣ . . a sufficient condition for this matrix being positive definite is dii > j=1.. (51) (− ln h) · r j + c j − btj a0 )2 /(− ln h). The optimization problem (51) is nonlinear because of the constraint condition Du > 0.

⎣ . (k = 1... btj a0 + (− ln h) · ri − ci /(− ln h).. [4]). respectively. . n). . we have Dl = PΦPt Definition 3.3 Numerical Examples In this section..2 (P. 2 2 (− ln h) · ri + ci − bti a0 /(− ln h).. 3..1.. n) ensure that Du where ε is a given small positive value. . (bti P)Φ(bti P)t (bti P)Φ(bti P)t (btj P)Φ(btj P)t (btj P)Φ(btj P)t bti a0 + . . (56) (− ln h) · r j − c j /(− ln h). ⎥ . 2 2 (− ln h) · r j + c j − btj a0 /(− ln h). ⎦ 0 0 · · · φn The matrix Φ is obtained by the following auxiliary problem P. two numerical examples are given to show how to obtain lower and upper possibility distributions of fuzzy decision variables. where Φ is a positive diagonal matrix as follows: ⎤ 0 ··· 0 ⎥ φ2 · · · 0 ⎥ ⎥ (55) .. After obtaining Φ. For a possibility level h. . . . ... Let us denote the feasible sets of decision variables obtained from upper h and lower possibility distributions with a possibility level h as Su and Slh ...Decision Making under Uncertainty by Possibilistic Linear Programming Problems 79 Assume that upper and lower possibility distribution matrices have the same eigenvectors. The h sets Su and Slh are defined as follows: h Su = {x ∈ Rn |Πu (x) h}..... . .. . h}. . ε Here ε φk λk .m ∑ δi (bti P)Φ(bti P)t . . Guo et al. (58) (59) Slh = {x ∈ R |Πl (x) n h Theorem 3. Thus Dl can be expressed as Dl = PΦPt ⎡ φ1 ⎢ ⎢0 ⎢ ⎢ . . Guo et al. Auxiliary problem: min Φ i=1.. Dl > 0 holds such that Πu Πl is satisfied (57) φk λk (k = 1. [4]. s. ⎢ . Slh ⊆ Su .... t. . ⎥. .

7.)T . 3 x1 + 4 x2 ≈− 34. 2 x1 + x2 ≈− 15.28.3 0.02]t . Guo et al. 9 = (9. r2 ]t .64.80 P. P. 3 = (3. 10. [5]). where 15 = (15. Example 3. x1 + 2 x2 ≈+ 13.5..58]t . The upper and the lower possibility distributions of fuzzy decision variables with h = 0.5 are shown in Figure 3.56. In Figure 3. The region of decision variables left by the possibilistic constraints of (60) is characterized by the possibility distribution of decision variables. The fuzzy numbers are all governed by exponential possibility distributions as follows: 17 = (17. Guo et al. a center vector was obtained by (22) with β = 0.)e .)T . 4. r)T with a = [a1 . 4. 13 = (13.1)e .)e .9)T . 1. 15 = (15.2. 2. x1 ≈− 5.. Let us consider the following possibilistic constraints.)T .5 the regions inside the two contours are feasible sets Su and Sl0. 1. x1 + 2x2 ≈+ 17.1 (P. x1 + x2 ≈+ 9.01 and g(a) = 2x1 + 3x2 as follows: a = [5. 2. 0.5. [4].5. Guo Example 3. rl = [0.)e . 4. x2 ≈− 4. It is obvious (60) that the solution obtained from the upper possibility distribution leaves more room than the one from the lower possibility distribution. 1.. (62) where x1 and x2 are decision variables.5. 10. (61) Assume that the possibility distribution of decision variables x1 and x2 is X = (a. (63) ..29]t . r = [r1 . Setting h = 0. 5.)T .3.5 = (4. 1]t as follows: ru = [0.5 .5.5 = (5. a2 ]t . x1 ≈− 3.. Let us consider the following fuzzy decision problem. 0. 2x1 + x2 ≈− 15... respectively. 34 = (34. Upper and lower spread vectors were obtained by (25) with w = [1.

3 as follows: a = [2. where a = [x1 .006 0. The feasible sets will be sent to lower decision levels where one specific solution will be chosen with considering the additional conditions.850. The upper and the lower possibility distributions of fuzzy decision variables with h = 0.3 Su and Sl0.5 (the outer is from the upper possibility distribution and the inner is from the lower possibility distribution). with δi = 1 as follows: Du = Dl = 1.298]t .357 5. 8.560 1.3 are shown in Fig. The fuzzy so- lution obtained from the upper possibility distribution leaves more room than the one from the lower possibility distribution.357 1. (65) and upper and lower possibility distribution matrices were obtained by (51). DA = d11 d12 d12 d22 .4 of which the 0. respectively. (56) and (57) The feasible region of decision variables left by the four fuzzy constraints (62) is characterized by the possibility distribution of decision variables.Decision Making under Uncertainty by Possibilistic Linear Programming Problems 81 0. (66) . (64) These four fuzzy constraints leave a feasible fuzzy region for decision variables.3 regions inside the two contours are feasible sets Su and Sl0.3 . DA )e . x2 ]t . a center vector was obtained with h = 0. .003 0. Setting β = 0. Decision variables x1 and x2 are governed by an exponential possibility distribution X = (a.405 1. 3.05 and g(a) = x1 + 2x2 in the LP problem (47).022 .5 Figure 3.006 1.

Upper and lower possibility distributions are obtained to represent solutions with different fuzzy degrees in decisionmaking procedure. “≈+ ” and “≈− ”.82 P.e. the possibility distributions of fuzzy variables with symmetrical triangular and exponential distribution functions are studied to reflect the inherent fuzziness in this kind of decision-making problem. . 3. i. These fuzzy satisfaction levels can be represented by two kinds of soft equal. An upper possibility distribution reflects the largest distribution and a lower possibility distribution reflects the smallest distribution of fuzzy decision variables for some given center vector.. Guo Figure 3. which leave some feasible region for decision variables.4 Conclusions For upper levels of decision-making systems. In this chapter. decision makers can always set some fuzzy satisfaction levels based on their experience. Decision makers in lower levels can make some more detailed decisions within the region of decision variables obtained by the proposed methods with considering some practical situation. The obtained fuzzy decision will be sent to lower levels of a decision-making system. The fuzzy decision based on an upper possibility distribution leaves more room than the one based on a lower possibility distribution.4 Possibility distributions of fuzzy decision variables (the outer is from the upper possibility distribution and the inner is from the lower possibility distribution).

8 (1975) 199–249. Journal of the Operations Research Society of Japan. International Journal of Approximate Reasoning. . Information Science. 1 (1978) 45–56. Fuzzy solution in fuzzy linear programming problems. 17 (1970) 141–164. Zadeh. and H. P. Slowinski and J. Asai. Upper and lower possibility distributions of fuzzy decision variables in upper level decision problems. Stochastic Versus Fuzzy Approaches to Multiobjective Mathematical Programming under Uncertainty. Guo. [2] M. Ishibuchi. A.-J. [10] H. (1990) 45–68. Tanaka. Management Science. and M. [6] Z. Tanaka and P. Guo. 92 (1996) 512–527. Tanaka and H. Fuzzy linear programming and applications. Tanaka and K. Fuzzy Sets and Interactive Multiobjective Optimization (Plenum Press. Rommelfange. Vol. and Cybernetics. [5] P. Tanaka and P. [15] L. Tanino. Fuzzy Sets and Systems. Fuzzy programming and linear programming with several objective functions. H. Dordrecht. T. Possibility distributions of fuzzy decision variables obtained from possibilistic linear programming problems. E. and H. Ichihashi. Zimmermann. Fusion of multi-dimensional possibilistic information via possibilistic linear programming. 111 (2000) 71–79. Eds. 111. in: R. [9] H. Fuzzy Sets and Systems. International Journal of Computer and Information Science. pp. Tanaka. [11] H. The concept of a linguistic variable and its application to approximate reasoning – I. Pawlak.A. Bellman and L.-J. (Physica-Verlag. Guo. Teghem.Computational Intelligence in Complex Decision Systems 83 Bibliography [1] R. Portfolio selection based on upper and lower exponential possibility distributions.29–45 (2000). Evidence theory of exponential possibility distributions. 8 (1993) 123–140. [4] P. 1999). 1993). Inuiguchi. 113 (2000) 323–332. Zimmermann. Tanaka and H. [13] H. Sakawa. Inuiguchi. No. T. [7] H. 11 (1982) 341–356. Fuzzy Sets and Systems. [14] H. New York. 14 (1984) 325–328. Kluwer Academic Publishers. Fuzzy Sets and Systems. European Journal of Operational Research. Decision making in a fuzzy environment. Guo. Guo. [8] M. Fuzzy Programming: A survey of recent developments. and H. Possibilistic Data Analysis for Operations Research.. [3] M. Heidelberg. Zadeh. 44 (2001) 220–229. 1. Zimmermann. IEEE Transactions on Systems. [12] H.-J. European Journal of Operational Research 114 (1998) 131–142. Man. Rough set. H. Sakawa: Membership Function Elicitation in Possibilistic Programming Problems. Tanaka. Entani.

The area is related to realistic and interactive experiments in real time computational systems.ufpb. visual and experimental approaches. Ruan. In the near future. From correctness degree the user will be qualified to perform the real task. This aspect makes the experimentation an important stage in the learning process.ufpb. These new resources have as goal to enlarge the information access and allow that it could be obtained outside educational environments and in moments conD. An evaluation system must monitor the user’s interactions with the VR application and be able to verify in real time the correctness degree of those interactions. methodologies for evaluation will be used to verify performance of multiple users in collaborative VR environments. Recently. the acquisition of knowledge can depend on the comprehension of physical phenomena. Computational Intelligence in Complex Decision Systems.br/∼labteve 1 liliane@di.Chapter 4 Intelligent Decision Making in Training Based on Virtual Reality Liliane dos Santos Machado 1 and Ronei Marcos de Moraes 2 Laboratory of Technologies for Virtual Teaching and Statistics. 2 ronei@de. Technological advances provide new resources for learning to increase or facilitate the assimilation of contents. as skills evaluation related to dexterity and reasoning power in critical procedures.br Virtual reality (VR) is an evolving area for simulation and training applications.br. This chapter includes the computational decision making based on statistical and fuzzy methodologies to perform user skills evaluation. © 2010 Atlantis Press/World Scientific 85 .1007/978-94-91216-29-9_4. DOI 10.1 Training Aspects The acquisition of knowledge can happen in several ways. In spite of the oral approach be the more traditional. 4. Federal University of Paraiba – Brazil http://www. Atlantis Computational Intelligence Systems 2. The transmission of knowledge includes oral.de.ufpb. publications in the literature show the use of continuous evaluation methods to measure the evolution of the user’s performance. Those systems generate data and information that can be used to intelligent decision making.

This feature allows the training in safe conditions. . dos Santos Machado and R. It is important to observe that all these training methods have as goal provide to apprentice the opportunity of development of psychomotor capabilities. the communication between apprentices and a tutor can be synchronous or asynchronous. When the simulation occurs through the Internet. Its challenge is to explore the five human senses. the apprentice also has physical contact in the assisted training. one of its main advantages is the possibility of simulation of micro and macroscopic realities. and exchange information in real-time. This chapter presents the main concepts necessary to the comprehension of virtual reality and systems that use it. listen or watch a lesson and submit their doubts to the tutor.86 L. but a tutor aids them during the tasks execution. became possible to explore senses of reality and presence through the networks access and the availability of faster processors. the training is known as Web training. Thus. through the use of devices and algorithms. A computer program to simulate the tasks is used in the computational training. as well as those than could present risks in the real world. the results of actions can be observed instantly. The synchronous communication occurs when apprentices and the tutor share the same environment. Generally. real or virtual. Virtual reality is a recent research area that unites concepts of several other areas in order to digitally provide realistic environments. Marcos de Moraes sidered convenient by the apprentice. the computational and Web training can use virtual reality to simulate real tasks. Additionally. These applications can be described as applications based on virtual reality or Web applications. offers reuse and can be continuously available. Those will be necessary to understand how information from virtual reality applications can be used to intelligent decision making and which methods have been and could be used for this end. In the physical training the apprentice has direct contact with the subject contents and can handle and modify properties. the answer will only be obtained after some time. It results in applications endowed of three-dimensional content. interactivity and/or real-time communication. computational and Web training. in the asynchronous communication the apprentices read. For training purposes. to simulate environments or events. In the context of the new resources. Particularly. In an opposite way. it decreases or eliminates the consumption of materials. assisted. the training methods can be decomposed in four main categories: physical. In a similar way. Additionally. Thus.

but keep its application restricted to games and military researches with flight simulators. interactivity and involvement. electronics. Virtual Reality can be defined as a science that unites knowledge of several areas. vibrations and reaction forces. The use of interaction devices and immersion is directly associated to cognitive aspects of human being and can explore visual. Through the following decade. motor and smell stimuli. VR systems react in real-time their users’ actions and demand fast processing.2 Virtual Reality In 1991. This new technology referred to artificial worlds.1 Virtual reality: the union of knowledge of several areas to provide immersion. Examples of input devices are the tracking sensors. In those environments. Technological advances allowed it to be used in several other areas and to receive prominence as an important resource for training purposes. Figure 4. auditory. as computer science. users have stimulated their several senses by the use of special devices and have the sensation of immersion in a virtual environment. in order to offer computational systems that integrate characteristics of immersion and interactivity to simulate real environments (Figure 4. sounds. Those devices can be separated in two main categories: input and output. physics and cognition. tactile.1). responsible by detect the position of an object or part . beside others. created by computers. and electronic devices and promised to transform the society. VR acquired fame from fiction movies. Input devices send to system information related to user movements and interactions.Intelligent Decision Making in Training Based on Virtual Reality 87 4. The information is used to the processing of new graphic scenes. Rheingold published a book about a new and revolutionary technology called Virtual Reality (VR) [48].

temperature and force resistance. three-dimensional systems that present sound according to user position. the interaction devices. Figure 4. inside or outside the projection plane (monitor or screen). sending for them the answers of the system. dos Santos Machado and R. presentation of sounds and vibration of devices.88 L.2 shows a tracking device attached to glasses to move an object in the virtual world according to users’ head movements. keyboard and joystick. Other output devices are: aural displays. Visual devices are commonly the most used and can provide mono or stereoscopic visualization.2 Tracking device send users’ head movements and allows changing the point of view. When stereoscopic devices are used. The effect is obtained through the use of projection algorithms and glasses or head-based displays able to present different images to each observer’s eye. users can observe floating in the space the objects of the virtual environment. and the biological sensors that use voice recognition and muscular electrical signals. The reactions can occur as a change in the point-of-view of the graphical scene. and the data gloves. as the traditional mouse. and haptic devices. objects deformation. In spite of the restricted interaction possibilities. Output devices aim to stimulate users’ senses.1 Interaction Devices Interaction devices allow the user-application communication and can capture user actions in real-time to be used as source for the processing of application reactions. aural. Marcos de Moraes of the body in which is attached.2. mouse and keyboard are the most economically accessible devices and can be used in . and haptic displays. 4. as equilibrium. Due to their function. devices that explore the sense of touch and proprioperception. Figure 4. In this category are the visual. the output devices are important to provide the sensation of presence and immersion in the VR environment.

In general. They can capture users’ head movements and provide images according new points of view of the virtual environment.2 Haptic Devices Haptic devices incorporate sensors and actuators to monitor user actions and provide touch and/or force feedback. A category of interaction devices allows both input and output of data and the computer can use them to receive and send information from/to users. Device Touch screen monitor Mouse Data glove Data glove with exoesqueletum Haptic devices Tracking device Active X X X X X X Reactive X X X conjunction with others in the interaction process. VR applications use to integrate interaction devices whose shape or manipulation mode is similar to tools used in real procedures. Technological limitations and high costs still prevent the use of devices that explore the other senses. Touch screen displays are an example of this category of device since they can present information (output) to users and also send users actions for computer. The devices can provide bi-directional communication between user systems when they support the transfer of actions and reactions. In fact. tracking devices can be attached to any real object that. datagloves allow computers to identify users’ fingers movements and can be used to tell the system that users are touching or grasping an object. In a complemental way. . In the last case.Intelligent Decision Making in Training Based on Virtual Reality 89 Table 4. Haptic devices also can capture actions and provide reactions. when handled by users.1 presents some examples of interaction devices that can be used in VR systems.1 Examples of interaction devices. As example. in order to achieve high levels of realism. The perception of touch is related to the nature of the contact with an object. hearing and tactile are the human senses most explored by VR. However. 4. as texture or temperature. a tracking system is attached to datagloves to capture hand position also. From the examples presented. Table 4.2. Other example is the set head-based displays with tracking devices attached. actuators are necessary to send reactions to users. is important to mention that sight. the perception of forces is related to the sense of position and movement associated to interaction with objects. could inform the movement to the VR application.

The PhantomTM – Personal Haptic Interface Mechanism – is one of those devices. support and react with different amplitudes. as scalpels. as example.3). haptic devices allow movements with several degrees of freedom. They support six-degree-of-freedom movements (translations and rotations) are suitable for applications with stereoscopic visualization. . as some joysticks. In general. hammers and needles. exoskeleton of datagloves. Available in different versions (Figure 4. most of the time. This feature makes them portable.90 L. On the other hand. offer manipulation in a restricted space and use different technologies. Most of haptic devices for legs or arms provide force feedback. the body-based haptic devices use a connection point of the own device to provide force feedback. A group of ground-based devices uses a type of mechanical arm to provide touch and force feedback. need to be physically attached to a platform or stable surface to provide force feedback. the device uses potentiometers to offer force feedback with different intensities in three (translations) or six (translations + rotations) degrees. A base in which the mechanical arm is attached and whose extremity is similar to a pen composes it. Ground-based haptic devices. Its popularity is due to its shape and manipulation way that allows the intuitive simulation of hand tools. Marcos de Moraes Haptic devices that provide force feedback can be separated in two main categories: ground-based or body-based. dos Santos Machado and R.3 Hapic device used to simulate a needle. The reason they do not usually provide touch sensation is that hands are more suitable and sensitive to this perception. Figure 4. Those devices used to have their cost related to the amount of force supported and returned.

It includes simulators whose goal is allowing practice in environments that imitate the reality of a procedure. . Since the 90’s. Nowadays. there aren’t risks involved. The system is composed by a replica of the machine cockpit with specific software that reproduces forests with trees of several diameters and heights. Users can practice in those environments and be technically and psychologically prepared to perform real tasks. During a training execution in a VR system.g. information from the simulation (user actions) can be collected and processed to evaluate users and assist decision making related to their abilities. Since the harvest is an action that cannot be repeated without loss. In order to teach motoric operating of machines for timber harvesting. the system can teach the best method to harvest and allows repetition of the method. Generally. The VR system is composed by intelligent methods that present new levels of difficulty and can use past training data to provide continuity in the learning process. In those virtual environments users feel visually as well as physically involved in scenes with features similar of the real world. bucking simulator). when automating basic functions and training cognitive skills (e. Other possibility is the use of information to change in real-time the simulation. and in later stages of training. the cockpit is set up in motion-platforms that can answer to user actions according to interaction movements. Such simulators use log files for evaluation in order to provide feedback for operator (trainee) and to allow the improvement of productivity and economy in operation. This kind of simulator has been used in two steps of training: in the beginning for training motoric skills. flight simulators are so precise and affordable that new pilots need to perform several hours of training in those systems before pilot a real aeroplane. including the update of experienced operators. Those simulators are constructed from a replica of a real cockpit of specific planes. there isn’t degradation of materials and the procedure can be repeated several times. increasing or decreasing the complexity as well as modifying their parameters according to user evolution.3 Training in Virtual Reality Systems Training is one of the most promising areas in the context of VR applications. flight simulators have been used in the training of new pilots in military and civil aviation. Sweden groups developed a simulator to allow new workers to acquire skill and get used to the machines [16]. In a similar way. can interact with their objects and observe the consequences of their actions. simulators can also offer training of new techniques through the presentation of different and/or generic situations (detached from particularities of real cases). As advantages: no one can be injured.Intelligent Decision Making in Training Based on Virtual Reality 91 4.. so much it was necessary.

placing. Marcos de Moraes Also for military purposes. the simulation of them in virtual environments can help to reduce such situations. Health is probably one of the most benefited areas of training systems based on VR. Since the task must be quickly performed. Additionally. dos Santos Machado and R. bone marrow harvest. Another training possibility in VR simulators is related to microscopic or macroscopic situations. Gao and L´ cuyer [15] developed a simulator for training and protoe typing the telemanipulation of nanotubes to help students or new operators to understand the phenomenon. In this context. In some cases. Additionally. gynaecology exam and rehabilitation. minimally invasive procedures are particularly explored because of its feature of be performed with small cuts in patient body. laparoscopy. VR systems for medical training include three-dimensional visualization. Nanotubes manipulation is an example of this kind of situation. there isn’t visual information available for the physician. models with deformation properties and haptic devices for touch and/or force feedback. The experience in those virtual battlefields can minimize psychological impact of soldiers in real battles and training them to deal with surprise events and with fast decision making. The bone marrow harvest for transplant in children is an example of that. the repetition of training in VR simulators is helpful when specific materials or procedures cannot be performed several times with plastic models or guinea pigs. deforming and connecting nanotubes. Laparoscopic procedures in virtual environments have been also developed due to the complexity of the tasks and the difficulty to understand how to manipulate the tools inside . In general.92 L. the system includes haptic manipulation and interactive deformation of the tubes. realistic images obtained by tomography or magnetic ressonance imaging (MRI). Some categories with VR-based simulators are: suture practice. In order to provide a valid simulator. among others. It is applied to several procedures and requires the operations of characterizing. the physician dexterity will result in lesser discomfort for patient and a faster recovery. The simulator of this procedure [25] provides stereoscopic visualization only of the external body and the apprentice must identify physico-elastic properties of tissues when use a virtual needle to harvest bone marrow. Those systems use haptic devices to simulate medical tools and propose to help apprentices to acquire dexterity to execute real tasks. dexterity is a key factor. In those procedures. Due to the fact that bad performed procedures can result in physical and/or psychological injuries. The games are composed by intelligent modules that present virtual enemies and challenge situations. VR have used the games approach to prepare soldiers for the battlefield. new manipulation tasks can be simulated in order to find optimal solutions.

they depend on devices to monitor user movements.4). the system must supervise user interactions to evaluate the training performed in simulators based on VR. torque. velocity. Then. temperature. in real-time. haptic devices also can do the same. The systems can evaluate users and rehabilitate cognitive and meta-cognitive processes as visual perception. the velocity of the movement and its acceleration. The system presents two modules with phases for observation and palpation. manipulates haptic devices and observes and feels the results of his actions. Other possibility for medical simulators is the training of stages of pathologies. In this context. resistance. sampling rates vary according to the device chosen. Those devices can collect movements in three or more degrees of freedom in constant intervals of time. Physiotherapy and rehabilitation have used VR to teach users the right form to perform movements. Such information allows VR system to calculate object position in the virtual scene. both supervised by an evaluation system that informs user performance. In order to achieve this. smells. but in a limited physical space. velocity and shape (Figure 4. However.Intelligent Decision Making in Training Based on Virtual Reality 93 human body. besides the direction of contact and its intensity. angle of visualization. as variables and functions responsible for feedback. The SITEG system is dedicated to training the identification of cervix pathologies in women [23]. In a similar way. Theoretically. Practically. including rare cases. Spacial position can be captured by tracking devices attached to some part of users’ body or to an object manipulated by them. Important features in those systems are graphics realism and intuitive interaction: both must provide immersion to user and make him stay concentrated in the movements performed [61]. forces. Particularly. several data for decision making. In both examples. temperature. In laparoscopic VR-based simulators the user visualizes a representation of the tweezers. the system can collect information related to position. Its goal is to offer a virtual environment with different stages of cervix inflammation. this information can also be utilized to verify the degree of accuracy of users when performing the simulation. it can be done in real-time to modify simulation aspects. acceleration.4 Collecting and Using Data from VR Simulators The collect of data during interaction in a VR-based simulation can provide. The positioning of objects manipulated by users can allow the identification of collision with other objects of the virtual scene. 4. sound. attention and memory [50]. devices must execute this task and some types of data cannot be collected nowadays due to technological restrictions. position vectors with last and present position are used to determine the collision velocity .

As example.4.94 L. increase or decrease of temperature. deformation of objects. dos Santos Machado and R. change of objects (as cuts or divisions) and smells. 4. Reactions from data processing collected from interaction can be send to users as force feedback. Figure 4.4 Interaction devices allow the collection of information from simulation. However. The intensity of feedback will depend on device capabilities.5 show how data collected from VR simulators can be used to evaluate users. and intensity. the continuity of reaction is an important factor in the levels of immersion. Marcos de Moraes Figure 4. Because users can have forgotten some decisions and actions done .5 Evaluation systems collect data from VR-based simulations to evaluate user performance.1 Online and Offline Evaluation Training procedures can be saved (video) for posterior analysis and used to inform users about their performance. it will be fundamental to decision making for evaluation purposes in simulators in which cuts and objects deformation is possible. Figure 4. sound. a new point of view of the scene.

online methods are linked to VR system and can offer results of performance immediately after the execution of the simulation. some performances cannot be classified as good or bad due to procedure complexity. Other methods for training evaluation can be found and the next sessions will present some of them. Several methods have been proposed to offline and online evaluation. However. number of mouse movements. answer accuracy.1 Using Classical Logic An intelligent evaluation using classical logic refers to a computational system that made an evaluation task in a similar way of one or more than one expert. and others. In opposite. they were separated in four sections: evaluation based on expert systems. users are evaluated according to the speed of answers. number of cuts. and evaluation based on Bayesian models. It is possible to find several kinds of simulators based on virtual reality and also several evaluation systems. Gande and Devarajan [14] used an Instructor/Operator station to monitor users movements to increase or decrease the difficulty degree of a simulation.5. the existence of an evaluation system coupled to a simulation allows fast feedback that can be used improve user performance. The station Instructor was prepared to evaluate the performance of a medical trainee based on predefined specifications and provided an effectivity report. The GeRTiSS System [1] was capable to create a report. Basically. the main problem in this kind of evaluation is the interval between training and evaluation.5 Evaluation Based on Expert Systems 4. with information about time of intervention.5 showed the independence between simulator and evaluation systems in spite of their simultaneous action. 4. In those systems. Offline methods evaluate users in a system independent of the VR system and data must be collected and stored for posterior use. Thus. online methods must present low computational complexity in order to not compromise the simulation performance. Because they are coupled to the training system. The surgeon could use that information to improve his performance by himself. Figure 4. they must also have high accuracy that doesn’t compromise the quality of evaluation. evaluation based hidden Markov models. evaluation based on mixture models.Intelligent Decision Making in Training Based on Virtual Reality 95 during the training. correct/wrong decisions and learning. The choice of the variables to be monitored depends on the relevance of each one to the training subject. the . In that system. Additionally. cauterisations. Other example is the WebSET System [18] that can simulate scenarios in which the users are trained.

4. among others.96 L. However. The architecture formed by knowledge base and inference system is named “expert system” [49]. The rule above is equivalent to a set of rules with only one conclusion. as for example rules. That means the system should never conclude “yes” and “no” from the same information. Each rule is related to a variable of interest and experts can have their own opinion about a specific fact. . belief and plausibility [53]. for example. some coefficients of certainty can be introduced to measure that uncertainty in the conclusion of rule. as the following: IF (<condition 1> OR <condition 2>) AND (<condition 3>) THEN <conclusion 1> IF (<condition 1> OR <condition 2>) AND (<condition 3>) THEN <conclusion 2>. A rule can be composed by conjunctions and disjunctions and can take more than one conclusion simultaneously.2 Using Fuzzy Logic As mentioned before. the inference system must activate rules in a knowledge base. In expert systems. However. The control of rules utilization is done by an inference system. the architecture formed by knowledge base and inference system is named “expert system”. Using rules and facts. which is able to evaluate the data received and to generate conclusions. in intelligent evaluation systems the expert knowledge is stored in a knowledge database using a logical representation as. Marcos de Moraes knowledge from one or more experts is stored in a knowledge database using a logical representation. dos Santos Machado and R.5. Then. For example: IF (<condition 1> OR < condition 2>) AND (<condition 3>) THEN (<conclusion 1> AND < conclusion 2>). When variable values are measured by an external subsystem and sending to the expert system. In those situations. there are cases in which features cannot be measured with the precision requested or the expert recognizes that the knowledge is not valid in specific situations. using facts that are obtained from beginning or after deduction process [8]. and possibilities [12]. knowledge about a procedure is modelled by rules as: IF <condition > THEN <conclusion>. rules. it is important the prevention of conflicts among those opinions in order to keep the knowledge consistency. which can be treated by probabilities [49]. new facts can be obtained using one or more reasoning ways.

regions of tissue and bones in where is possible to find bone marrow harvest. described by the following fuzzy sets: you need much more training. triangular and pi-functions and the fuzzy inference system used is Mamdani-type [28]. μA (c)=0. Machado et al. 1]. When there is subjectivity in the knowledge base. [24] used five performance classes. values of variables are collected by a haptic device and sent to expert system for evaluation. where [0. its representation can be done by fuzzy models [65]. In case of bone marrow harvest. Therefore. We say that a fuzzy set A of X is “precise” when ∃ c∗ ∈ X such that μA (c∗ )=1 and ∀ c=c∗ . where. 1}. you need training. experts can define. The system must analyse each step performed by users and classify users. express the presence and absence of a in relation to A.1] can be any bounded scale. In fuzzy expert systems. your training is good or your training is excellent. in an imprecise way. Several types of membership functions can be used as trapezoidal. Position y are coordinates which the needle touch the patient body. μA (a) = 1 and μA (a) = 0 respectively.Intelligent Decision Making in Training Based on Virtual Reality 97 In (classical) set theory. To evaluate a user. fuzzy rules model the knowledge about a procedure. for a given a ∈ X. human experts can only evaluate the procedure through their external impressions of the patient’s body. A fuzzy set A on X is characterized by its membership function μA : X → [0. μA (c) ∈ {0. each subset A of a universe X can be expressed by means of a membership function μA : X → {0. according to predefined performance classes. using a fuzzy expert system coupled to a surgical simulator based on virtual reality. A fuzzy set A will be said to be “crisp” when ∀ c ∈ X. The fuzzy rules of expert system are modeled by membership functions according to specifications of experts. you need more training. Marrow harvest shows the success or failure of trainee to harvest bone marrow and Trainee class is the classification of trainee. A fuzzy set [63] is used to model an ill-known quantity. 1}. at the end of training. Position needle is the angle of needle input to body of patient. this procedure is blind made. using fuzzy sets. . An example of rule for this expert system is: IF Position x is left center AND Position y is up center AND Position needle is acceptable AND Marrow harvest is yes THEN Trainee class is you need training where: Position x. The evaluation system proposed by Machado et al. [24] has privileged information because it can obtain external and internal data about patient’s body. to identify if more training of bone marrow harvest procedure is or not necessary. For example. because it is only what they can see.

The first corresponds to the visual stage of the exam and the second corresponds to the touch stage. After the visual exam. gynaecological exam is one of the most important exams to female health and allows detecting pathologies that can evolve to cervix cancer. what result in an incomplete training. Some medical exams require sequential steps and different knowledge for each one. The Simulator for Gynaecological Exam (SITEG) [26] allows the training of gynaecological exam and simulates different phases of pathologies. In the visual exam. The two stages of a real exam were divided to compose a visual and a touch exam. The doctor uses a lubricated glove to touch vagina walls and cervix to identify any wounds or lumps. as presented in Figure 4. When touch the cervix. The evaluation system for SITEG is based on two fuzzy rule-based expert systems. dos Santos Machado and R. For example. One example can be the patient’s discomfort when this exam is performed in the presence of a medicine student (apprentice). In the second stage there isn’t visual information. knowledge modelling of a complex procedure can demand a particular approach. The simulator presents randomly normal. A haptic device is used to allow touch and force feedback in the touch examination (Figure 4.2 Color Normal Herpes / HPV Inflamed rosy white with warts red Visual and haptic properties description Viscosity smooth with bubbles smooth Cervix elasticity Similar to an orthopedic rubber very soft hard/tense Texture similar to buccal mucous membrane spongy and irregular similar to buccal mucous membrane 4.5.3 Combining Expert Systems In some cases. the user must identify the cervix colouring according to a diagno- . Other difficulty is related to the students’ absence of opportunity to practice all possible cases. In the visual stage.2.6). In the first stage is used an instrument named speculum to allow the visualization of the vagina walls and of the cervix to check colour and surface of these structures. as it is showed at Table 4.7. this kind of exam presents some difficulties. HPV or Herpes and inflamed cases to the user. each one according to the knowledge obtained from an expert. In general. the user will have only the external view of the vagina and must perform a touch examination perceive the texture and detect if there is wounds or lumps. the doctor will also feel its elasticity. Marcos de Moraes Table 4. the user must observe the vagina walls and cervix and notice their colouring.98 L. This occurs because the only way of training is by the observation and experimentation. The gynaecological examination is performed in two steps.

Herpes/HPV or inflamed. It happens because contradictions . This information will be used and stored by the second fuzzy rule-based expert system.7 Diagram of the evaluation system. In the second stage.6 Gynaecological examination performed in a VR-based simulator: (left) visual exam with a speculum (grey object) and (right) external view of the unidigital exam. In this case. Figure 4.Intelligent Decision Making in Training Based on Virtual Reality 99 Figure 4. herpes or HPV. the users perform the touch exam and emit again an opinion according to the possible diagnosis: normal. The evaluation system compares it to the information related to the case presented and stores it in a database that will be used by the second fuzzy rule-based expert system in the second stage. At the end of the first stage the user must provide to the system his opinion about the colouring and his diagnosis. inflamed. doubt}. the evaluation system will request from the user a final diagnosis that should be chosen among: D = {normal. Each colouring has its own fuzzy membership function: normal (rosy). Herpes/HPV or inflamed. Herpes or HPV (white) and inflamed (red). At the end of both stages. the “doubt” option is pertinent due to the combination of information during a real exam that allows doctors to define a final diagnosis. sis of normality.

. (b) user is almost qualified to execute a real procedure.7). performance is regular. (d) user needs more training to be qualified. Thus. is provided to user in the evaluation report (Figure 4. This system uses five classes of performance: (a) user is qualified to execute a real procedure.100 L.1 Gaussian Mixture Models There are cases in which statistical distributions of data cannot be assumed as Gaussian distributions (univariate or multivariate). [59] explanation about GMM algorithm and classification. position and torque extracted at d different interval of time. The evaluation system is also capable to decide if the training was successful by the combination of the two stages and the final diagnosis with the case presented by the simulator. but it can be better. performance is very poor. xT } be a set of T vectors. Internally. where each one is a d-dimensional feature vector extracted by T different information at virtual space.6. This section presents the GMM method for training evaluation. obtained by the simulator. This information can be applied forces. a possible solution is modeling them using mixture models [29]. Moraes and Machado [32] presented a solution using Gaussian Mixture Models (GMM) as the kernel of an evaluation system. the evaluation system can classify the user into classes of performance. the evaluation system executes in real-time the rules of each stage according to the case presented to users. (e) user is a beginner. x2 . . After. dos Santos Machado and R. Parameter estimation equations for training expert models are presented first. the GMM method for training classification is then described as a maximum likelihood classifier. This way. performance is good. including the users’ class of performance.6 Evaluation Based on Mixture Models 4. We follow the Tran et al. 4. . All these information. Due to all those features. Marcos de Moraes about what the doctor decides on the patient’s conditions between the first and second stage of the exam can occur. angles. It is important to mention that in cases of wrong diagnosis. Since the distribution of these vectors is unknown. performance is bad. . Let X = {x1 . (c) user needs training to be qualified. it is approximately . that assessment methodology can be used in continuous evaluation of training. the evaluation system is able to detect where the user made a mistake by the analysis of the degrees of pertinence of his information.

Σi ) k=1 ∑ wk N(xt . . (1) where λ denotes a prototype consisting of a set of model parameters λ = {wi . T. μk . the likelihood of the GMM is done by: p(X | λ) = ∏ p(xt | λ). i (2) To train the GMM. λ) = wi N(xt . c and satisfies p(i | xt . an auxiliary function Q is used as follows: Q(λ. Σi ) are the d-variate Gaussian component densities with mean vectors μi and covariance matrices Σi : N(xt . wi . Σi ) = exp{−1/2}(xt − μi ) Σ−1 (xt − μi )}/(2π)d/2|Σi |1/2 . λ) then p(x | λ) p(x | λ) [46]. . . are the mixture weights and N(xt . . Σi ) (4) where p(i | xt . μi . found the 1 t ∑ p(i | xt . c. p(xt | λk ) = ∑ wi N(xt . its direct maximization is not possible. . . The maximum likelihood estimation is widely used as a training method. . The setting of derivatives of Q function with respect to λ to zero. λ) log wi N(xt . . μi . μi . λ) is the a posteriori probability for performance class i. i=1 c t = 1. λ)] t=1 t=1 ∑ p(i | xt . λ)xt ]/[ ∑ [p(i | xt . Σi }. . λ) T The algorithm for training the GMM is described as follows: . In this algorithm. t=1 T (3) The aim of maximum likelihood estimation is to find a new parameter model λ such that p(X | λ) p(X | λ). μi . i = 1. . i = 1. these parameters can be obtained iteratively using the Expectation-Maximization algorithm [10]. . given by (1). those parameters are estimated such that they best match the distribution of the training vectors. μi . However. Σk ) c (5) The Expectation-Maximization algorithm is such that if Q(λ. For a sequence of training vectors X for a λ. λ) following reestimation formulas: wi = μi = Σi = T t Q(λ. λ)(xt − μi)(xt − μi ) ∑ p(i | xt . Since the expression in (3) is a nonlinear function of parameters in λ.Intelligent Decision Making in Training Based on Virtual Reality 101 modeled by a mixture of Gaussian densities as the weighted sum of c component densities. λ) = ∑ T c t=1 i=1 ∑ p(i | xt . Σi ). . λ) t t=1 t t=1 (6) (7) (8) t=1 ∑ [p(i | xt . μi .

. k = 1. and the covariance matrix following (6). λ ) according to (5) and compute the Q function using (4). 3. λ ) at random satisfying (5). . Update the a posteriori probability p(i | xt .. a classifier is designed to classify X into N classes of performance by using N discriminant functions gk (X). denote models of N possible classes of performance. . . N. . So. using the log-likelihood. computing the similarities between the unknown X and each class of performance λk and selecting the class of performance λk∗ if [59]: k∗ = arg max 1 k N gk (X) (9) In the minimum-error-rate classifier. the discriminant function in (10) is equivalent to the following [19]: gk (X) = p(X | λk ) (12) Finally.102 L.e. the decision rule used for class of performance identification is: Select performance model k∗ if k∗ = arg max 1 k N t t=1 ∑ p(xt | λk ) (13) where p(xt | λk ) is given by (1) for each k. the mean vector. Since p(X) is the same for all performance models. Given a feature vector sequence X. To provide GMM classification. p(λk ) = 1/N. let λk . Compute the mixture weight. 4. (7) and (8). relative to the value of the Q function at the previous iteration. the discriminant function is the a posteriori probability: gk (X) = p(λk | X) It can be used the Bayes’ rule p(λk | x) = [p(λk ) p(x | λk )]/p(x) (11) (10) and assume equal likelihood of all performances. several classes of performance λ are needed. otherwise go to step 2. i. Stop if the increase in the value of the Q function at the current iteration. . dos Santos Machado and R. Generate the a posteriori probability p(i | xt . . 2. is below a chosen threshold. . N. k = 1. . Marcos de Moraes 1.

Bezdek [4] defined as objective function: Jm (U. μi . each uit represents the degree of vector xt i c c and 1 t T . which is the distance in (14).Intelligent Decision Making in Training Based on Virtual Reality 103 4. The main advantage of their approach is the ability of recognition of different geometric shape distributions in X using covariance matrices and mixture weights Tran and Wagner [56].2 Fuzzy Gaussian Mixture Models In cases in which there is imprecision in variables measures or multiple ways of correct execution of a procedure. For 1 0 m uit 1. In those cases a fuzzy approach for GMM can be used.6. Σi )] 1 1 −1 = − log wi + log(2π)d | Σi | + (xt − μi ) Σi (xt − μi ). denoted by dit > 0. given by: 2 dit = − log p(xt . evaluation can use a fuzzy approach. From within-groups sum of squared error function. we have: i=1 ∑ um = 1 and 0 < it t=1 ∑ uit < T t (15) 1 is a weighting exponent on each fuzzy membership uit and it is called the degree of 2 fuzziness. it gives Jm (U. 2 2 Substituting (16) in (14). t t=1 i=1 2 ∑ um dit it c (14) where U = {uit } is a fuzzy c-partition of X. λ) = ∑ belonging to the ith class. Tran et al. They used a modification of Fuzzy C-Means algorithm to estimate the GMM parameters. as a kernel of another evaluation system. i | λ) = − log[wi N(xt . [59] proposed an approach named FGMM that was followed by Moraes and Machado [33]. and a measure of dissimilarity. λ) = − ∑ t (16) t=1 i=1 ∑ um log wi − ∑ ∑ um it it −∑ t c t t=1 i=1 t c t=1 i=1 c t c 1 log(2π)d | Σi | 2 1 −1 (x − μi ) Σi (xt − μi ) 2 t (17) ∑ um it Minimizing Jm is performed using Lagrange multiplier Method and it gives [58]: wi = μi = Σi = t=1 t t=1 ∑ um it ∑ um xt ∑ um it it t=1 t=1 i=1 t=1 t ∑ ∑ um it (18) (19) ∑ um (xt − μi) (xt − μi ) it t t=1 ∑ um it t (20) The algorithm for training a FGMM is the following [55]: .

μ i . set uit = 1. λ ) at the current iteration relative to the value of the Jm (U. compute the fuzzy mixture weights. Finally. λ ) done by (4) in GMM. 2 < c < T . to minimize Jm is not equivalent to maximize Q(λ . . λ ). 3. m > 1. λ ) function at the previous iteration is below a chosen threshold. . a classifier is designed to classify X into N classes of performance by using N discriminant functions gk (X). c. is obtained the decision rule for class of performance identification done by (13). In FGMM method. Generate matrix U at random satisfying (15). Marcos de Moraes 1. λ ) = ∑ in order to reduce computations Qm (U. k = 1. denote models of N possible classes of performance. 4. λ ) = −Jm (U. dos Santos Machado and R. Σ i ) it (22) According to Tran and Wagner [55]. Thus. (21) if dit = 0 for some t. there are two cases [55] to select the discriminant function: m is closed to one (inclusive) or m > 1 and m is not close to one. an alternative discriminant function must be chosen Qm (U. λ ) = Qm (λ ) = c i=1 t=1 ∑ t um log wi N(xt . Update matrix U using: dit dkt (2/(m−1)) −1 c k=1 ∑ . In the GMM algorithm. otherwise go to step 2. . uis = 0. . In the first case. For i = 1. for all s = t. the fuzzy mean vectors. to minimize Jm is equivalent to maximize Q(λ .104 L. i | λ ) (1/(m−1)) 1−m (23) i=1 Let λk . Compute the dit . Stop if the decrease in the value of the fuzzy objective function Jm (U. Then. N. following (16). the U matrix in Qm (U. . Given a feature vector sequence X. λ ) in (4) is maximized and thus the likelihood function p(X | λ ) is maximized too. Then. the model parameters λ are determined such that the function Q(λ . the decision rule used for identification in FGMM is the same that applied in GMM done by (13). (19) and (20). Choose c and m. 1 i c and 1 uit = t T . the decision rule used for class of performance identification is: Select performance model k∗ if k∗ = argmax 1 k N (24) t=1 ∑ ∑ t c − log p(xt . given the training sequence X. . by use of log-likelihood in discriminant function done by (12). λ ) is replaced by (21) and (16) t=1 ∑ ∑ t c − log p(xt . i | λ ) (1/(m−1)) 1−m (25) i=1 . and the fuzzy covariance matrices following (18). . . 2. In the second case. such as gk (X) = Qm (λk ).

the use of Relaxation Labeling (RL) after Gaussian Mixture Models can provide better results. Moraes and Machado [34] proposed the composition GMM-RL as a kernel of a training evaluation system in virtual reality based environments. .1 Gaussian Mixture Models and Relaxation Labeling The GMMs and FGMMs methods were presented in previous sections as a good option for training evaluation based on virtual reality. Those methods are explained as follows. λl ) = 0. . λl ) < 0. denote a set of objects A = {a1 . RL will be used after a GMM classification. The two methods are based on those developed by Peleg and Rosenfeld [45]. λl are incompatible for at and at . λl ). . according to Tran et al. The correlation-based estimation of the compatibility coefficients is defined . The meaning of these compatibility coefficients can be interpreted as [59]: a) If rtt (λk . An initial probability is given to each object at having each label λk . then λk .3 Sequential Methods There are ways of improve the results of a classifier using another classifier after that. λl are compatible for at and at . These probabilities satisfy the following condition: k=1 ∑ N pt (λk ) = 1. .6. Then. λN }. λ2 . However. In training evaluation based on virtual reality some methods based on 1) Gaussian Mixture Models and Relaxation Labeling and 2) Fuzzy Gaussian Mixture Models and Fuzzy Relaxation Labeling can be found in literature. where rtt (λk .6. b) If rtt (λk . then λk . then λk . λl ) : Λ × Λ → [−1. For VR evaluation purposes.3. . [59]. λl are independent for at and at . aN } and a set of labels Λ = {λ1 . whose magnitude denotes the strength of compatibility. which is denoted by pt (λk ). . 1]. . 4. .Intelligent Decision Making in Training Based on Virtual Reality 105 It is important to note that the decision rule in (24) can also be applied to the first case when m is close to one. a2 . [51] introduced the Relaxation Labeling (RL): an interactive approach to update probabilities of a previous classification. c) If rtt (λk . Two possible methods employ the concepts of statistical correlation and mutual information for computing coefficients. It has been successfully employed in image classification [13]. Rosenfeld et al. 4. λl ) > 0. for all at ∈ A (26) The RL updates the probabilities pt (λk ) using a set of compatibility coefficients rtt (λk .

dos Santos Machado and R. update the probabilities for each object using the updating rule in (32). 5. In order to minimize the effect of dominance among labels. compute the compatibility coefficients using (28) or (29). . the modified coefficients are [59]: ∗ rtt (λk )(λl ) = [1 − p(λk )][1 − p(λl )]rtt (λk )(λl ) (28) and the mutual-information based estimate of compatibility coefficient is t ∑ pt (λk )pt (λl ) ∑ pt (λk ) pt (λl ) t=1 t t rtt (λk )(λl ) = log (29) t=1 The compatibility coefficients in (29) must be scaled in the range [−1.106 L. Marcos de Moraes as rtt (λk )(λl ) = t=1 ∑ [pt (λk ) − p(λk )][pt (λl ) − p(λl )] t σ (λk )σ (λl ) (27) where pt (λl ) is the probability of at having label λl and at are the neighbors of at . calculate the updating factor defined in (30). p(λl ) is the mean of pt (λl ) for all at . 3. 2. The updating factor for the estimation pt (λk ) at the mth interaction is: qt (λk ) = (m) t =1 ∑ dtt ∑ rtt (λk )(λl )pt l=1 t N (m) (λl ) (30) where dtt are the parameters that weight the contributions to at coming from its neighbors at and subject to t t =1 ∑ dtt = 1 (31) The updated probability pt (m+1) (λk ) for object at is given by: = pt (λk ) 1 + qt (λk ) k=1 (m) (m) (m+1) pt (λk ) ∑ pt (λk ) 1 + qt (λk ) N (m) (m) (32) The RL algorithm can be outlined as follows: 1. 4. estimate the initial probabilities for each object satisfying (26). 1]. repeat steps 3 and 4 until the change in the probability is less than a chosen threshold or equal to a chosen number of interactions. and σ (λl ) is standard deviation of pt (λl ).

the initial label probabilities and the compatibility coefficients need to be well determined. the updating rule in (32) should be now rewritten [59] as follows: pt (m+1) (m) (m) (λk ) = pt (λk ) 1 − qt (λk ) k=1 ∑ N (m) pt (λk ) (m) 1 − qt (λk ) (34) The GMM-RL algorithm for class of performance identification is stated as follows. there isn’t necessity to consider the compatibility between an input vector and its adjacent vectors [59]. Objects are now feature vectors considered in the GMM and labels are classes of performance identification. This leads to: pt (λl ) = pt (λl ) (33) which means that compatibility between different labels is only considered for a same object. Calculate the updating factor defined in (30). 1. the initial probabilities in the RL are defined as the a posteriori probabilities. In the GMM-based classification. where t = t (no neighbours considered). Compute the compatibility coefficients using (28) or (29). Thus. 2. . in performance identification. Update the probabilities for each class of performance using the updating rule in (32). for a successful performance in a relaxation method processing. Estimate the initial probabilities for each class of performance using the a posteriori probabilities in (3): pt (λk ) = p(λk | xt ) = [p(xt | λk )p(λk )] k=1 N (35) ∑ p(xt | λk ) p(λk ) where p(λk ) = 1/N and p(xt | λk ) is computed as in (1). xT } are known to belong to a certain class of performance λ . Wrong estimations of these parameters will lead to algorithmic instabilities. 3.Intelligent Decision Making in Training Based on Virtual Reality 107 Clearly. 6. 5. The probability of each class of performance p(λk ) after RL algorithm is computed by: p(λk ) = ∏ pt (λk ) t=1 T (36) . Therefore. . . 4. Repeat steps 3 and 4 until the change in the probability is less than chosen threshold or equal to a chosen number of interactions. where t = t and dtt = 1/T for simplicity. x2 . Unlike the relaxation labeling for image recognition where the m-connected neighbouring pixels may belong to different regions. all unknown feature vectors in the sequence X = {x1 . .

it was 1 N N chosen that proposed by Borotschnig et al.7. Among existing FRL methods in literature. the decision rule for class of performance identification is as follow [59]: Select class of performance k∗ if k∗ = argmax p(λk ) 1 k N (37) 4. 4.108 L. [51] has been widely used for several applications with good performance [45]. B. Machado and Moraes [22] proposed the use of a two-stage evaluator where: FGMM is the first stage and FRL is the second stage.3. The main advantage is that non-normalized models allow multiple labels for one object.7 Evaluation Methods Based on Hidden Markov Models 4. [7]: φt (m+1) (λk ) = j=1 ∑ max T rt j ∧ φt (λk ) (38) where ∧ means the logical operator “AND” which can be translated by a t-norm (as for example min). Therefore. In the classical HMM five elements are necessary to specify a discrete HMM [46]. the discrete output of system. i.6. dos Santos Machado and R. Denote a HMM by λ = (A. The FRL method stops when the changes in the pertinence degrees are smaller than a chosen threshold or when ends the number of interactions previously defined. when is needed: (i) the number N of states in the model.. (ii) the number M of distinct observation symbols per state. According to Borotschnig et al.2 Fuzzy Gaussian Mixture Models and Fuzzy Relaxation Labeling The RL model proposed by Rosenfeld et al.e. several non-normalized models were also proposed [17]. [7] non-normalized models are widely used as Fuzzy Relaxation Labeling (FRL). Marcos de Moraes where pt (λk ) is the a posteriori probability used in (10). It is possible to consider previous results obtained from the FGMM method as initial degrees of membership to each object at having each label λk which is denoted by φt (λk ) [5]. However. π).1 Discrete Hidden Markov Models (DHMM) Baum and his colleagues published in the late 1960s and early 1970s the basic theory of Hidden Markov Models (HMM) [2. The FRL method updates the membership degrees φt (λk ) and φt (n+1) (λk ) is obtained. 3]. .

how to efficiently compute P(O | λ). . oT and a model λ. are mixture coefficients and N(ot . Σ jk ) is a Gaussian with a mean vector μ jk and covariance matrix Σ jk for the kth mixture component in the state j.expectation-maximisation-method) [46]. . and (v) the initial state distribution π = {πi }..Intelligent Decision Making in Training Based on Virtual Reality 109 (iii) the state transition probability distribution matrix. k = 1. . Optimise the model parameters: How to adjust the model parameters λ = (A. the system must be trained from the beginning again. . . done by: N(ot .2 Continuous Hidden Markov Models For most applications. . is only needed to train the newly added class. N. oT and a model λ. M. the observations are continuous vectors. . we can use the Forward-Backward Procedure [46]. . Σ jk ) (39) where ot . the probability of the observation sequence. given a model? 2. Σ jk ) = (2π )N/2 1 1 × exp − (ot − μ jk ) (Σ jk )−1 (ot − μ jk ) 1/2 2 | Σ jk | (40) . . . denoted by B = {b j (k)}. denoted by A = {ai j }. . . . w jk . . . t = 1. there are three basic problems of interest that must be solved for the model to be useful in real applications. These problems are: 1. The last problem can be solved by an iterative procedure such as the Baum-Welch Method (or equivalently the EM . (iv) the observation symbol probability distribution in state j. In CHMM. Uncover the hidden part of the model: Given the observation sequence O = o1 . . j N and 1 k M. o2 . . T . to do this it is necessary training data for all classes. sT which is optimal in some meaningful sense (i.e. μ jk . μ jk . j = 1. s2 . It can be used Vector Quantization (VQ) to generate codebooks and use DHMM. b j (ot ) = P(ot | λ ) = M k=1 ∑ w jk N(ot . .7. . best “explains” the observations)? 3. However. In next section. how to choose a corresponding state sequence S = s1 . the notation used by Rabiner [46] will be followed. Evaluation problem: Given the observation sequence in time O = o1 . Given the HMM form. 4. . When a new class is added. o2 . B. . . . The second problem is an optimisation problem and it can be solved by using the Viterbi Algorithm [52]. π) to maximise P(O | λ)? To solve the first problem. μ jk . are observation vectors being modelled. . The densities are considered as a mixture of Gaussians. where 1 i.

λ1 ) 2 (43) 4. B2 . . Let ui jt = ui jt (O) be the membership function.4 Fuzzy Hidden Markov Models There are several fuzzy approaches for HMM and the continuous versions of Fuzzy HMMs are particularly interesting for training evaluation [35]. the first one is followed and denoted by FHMM. Let be λ1 = (A1 . . which is done by a distance measure between two HMMs. Unfortunately. k=1 ∑ w jk = 1. Hence. satisfying: 0 ui jt 1. it can be shown that the reestimation equations for . to which the observation sequence O belongs to state i at time t and to state j at time t + 1. Among those approaches.3 Comparison of Hidden Markov Models In a training evaluation is important to compare two HMMs. dos Santos Machado and R.7. In this section. i=1 j=1 ∑∑ N N ui jt = 1. 0< t=1 ∑ ui jt < T T (44) Using the fuzzy EM Algorithm [57]. λ2 ) = 1 log P(O(2) | λ1 ) − logP(O(2) | λ2 ) T (42) where O(2) = O1 . It is wanted to measure the similarity of two models. λ2 ): D(λ1 . this measure is nonsymmetrical. M +∝ −∝ b j (ot )dot = 1 (41) An iterative reestimation process is used to found the coefficients. two different HMMs. π1 ) and λ2 = (A2 . denoting the degree.110 L. 4. a natural expression for this measure is the symmetrised version: DS (λ1 .7. it can be cited those proposed by Tran and his colleagues [57. It is a variant of the classical K-means algorithm [19] for clustering data. π2 ). Thus. B1 . 60] and by Mohamed and Gader [30]. Marcos de Moraes and the following constrains are satisfied: w jk > 0. denoted by D(λ1 . (42) is a measure of how well model λ1 matches observations generated by λ2 . . OT is a sequence of observations generated by model λ2 . relative to how well model λ2 matches observations generated by itself [46]. λ2 ) + D(λ2 . . λ2 ) = D(λ1 . Rabiner [46] described a procedure for providing good initial estimates of the model parameters called the segmental K-means [20]. O2 .

∑ βt+1 ( j) jk (47) ηt ( j. .8. or 2) To use a discriminant criteria. ∑ jk where αt ( j) and βt ( j) are the forward and backward respectively [47]. ∑ ∑ M w jk N ot . m>1 (46) u jkt = ηt ( j. kt = k | O. . z = 1. kt = k | λ ) jkt jkt = − log and i=1 ∑ αt (i)ai j w jk N N ot . denote models of Z FHMMs modeled. μ jk . k). k) = P(st = j.Intelligent Decision Making in Training Based on Virtual Reality 111 coefficients of the mixture of densities done by (39) are: w jk = μ jk = Σ jk = t=1 T t=1 T t=1 ∑ um jkt ∑ um ot jkt T t=1 k=1 T t=1 ∑∑ ∑ T M um jkt (45) um jkt ∑ um (ot − μ jk )(ot − μ jk ) ikt t=1 ∑ um jkt T where the prime denotes vector transposition. m N M 1 is a weighting exponent on each fuzzy membership uy (x) (and is called a degree of fuzziness) and u jkt = i=1 l=1 ∑∑ d jkt dilt m=1 (2/(m−1)) −1 . μ jk . Z. it is possible to classify O in two different ways: 1) To verify the distances between each pair λz and λO and find the smaller [46].8 Evaluation Methods Based on Bayesian Models 4. . μ jk . where d 2 = d 2 (O) = − log P(O. 4. Let λz . to classify O into one of Z models [60]. Its application for training evaluation based on VR was presented by Moraes . Given a new feature vector sequence O . λ ) = {αt ( j)βt ( j)} j=1 ∑ αt ( j)βt ( j) jk k=1 N (48) × w jk N ot . . st = j.1 Maximum Likelihood Maximum Likelihood decision rule is the most common statistical technique used in data classification.

(54) can provide a significant reduction of computational cost of implementation. . (52) P(X | wi )P(wi ) . M. Formally. Marcos de Moraes and Machado [36]. M. the user’s interactions with the system are monitored and the information is sent to the evaluation system. P(wi ) is called a priori probability for wi and P(wi | X) is a posteriori probability for wi where X is known. In Bayesian theory. where Maximim Likelihood is in its kernel. the classes of performance for an user are done by: wi . the classification rule done by (50) is modified: X ∈ wi if P(wi | X)P(wi ) > P(w j | X)P(wi ) for all i = j. where i = 1. (50) However. the use of (54) has interesting computational properties [19].112 L. i = 1. dos Santos Machado and R. Equation (53) can be used to modify the formulation done by (52): X ∈ wi if gi (X) > g j (X) for all i = j. i=1 As P(X) is the same for all classes wi . (49) The probability done by (49) gives the likelihood that for a data vector X. (54) It is important to note that if statistical distribution of training data can assume multivariate Gaussian distribution. . then it is not relevant for data classification. . Thus. it can be convenient to use [19]: g(X) = ln[P(X | wi )P(wi )] = ln[P(X | wi )] + ln[P(wi )] (53) where g(X) is known as the log-likelihood function and it is known as discriminant function. Then. . P(X) (51) Equation (52) is known as the maximum likelihood decision rule or maximum likelihood classification. where M is the total number of classes of performance. . It is possible to determine the most probable class of a vector of training data X. the correct class is wi . . However. all the probabilities done by (59) are unknown. Classification rule is performed according to X ∈ wi if P(wi | X) > P(w j | X) for all i = j. probabilities denoted by P(X | wi ) can be estimated if there is sufficient information available for each class of performance. . . by conditional probabilities [19]: P(wi | X). So. . If training data cannot assume that distribution. Using the Bayes Theorem: P(wi | X) = where P(X) = ∑M P(X | wi ) P(wi )].

. with a membership function μA (x). now is assumed that wi are fuzzy sets over space of decision Ω.Intelligent Decision Making in Training Based on Virtual Reality 113 4. we assume that the fuzzy information font is a histogram of the sample data. Then a fuzzy event in Rn is a set A in Rn . . according a vector of data X. In this case. i = 1. let performance classes wi . where μA : Rn → {0. then the Fuzzy Bayes classification rule is to assign the vector of training data X from the user to wi class of performance if: X ∈ wi se μwi (X)P(wi | X)P(wi ) > μw j (X)P(w j | X)P(wi ) for all i = j (57) μwi (X)P(X | wi )P(wi ) P(X) (56) As in evaluation system based on Maximim Likelihood.8. be in space of decision Ω = {1.2 Fuzzy Bayes Rule Moraes and Machado [38] proposed an evaluation system based on Fuzzy Bayes Rule [54] that has similarities to the method presented in previous section and based on conditional probabilities. Again. φ. However. as the denominator is independent. where M is the total number of classes of performance. let the classes of performance for an user be done by wi . The probability of a fuzzy event A is defined by Lebesgue-Stieltjes integral: P(A) = Rn μA (x)dP = E(μA ) (55) In others words. a rule composition system of the expert system. . or a histogram of the sample data). the probability of a fuzzy event A with the membership function μA is the expected value of the membership function μA . M}. 4. M. . However. Formally. the evaluation system based on Fuzzy Bayes Rule analyses data from user’s interaction and emits a report about the user’s performance at the end of the training. P) a space of probability where φ is an σ-algebra in Rn and P is a probability measure over Rn .8. Based .3 Naive Bayes The Naive Bayes (NB) method. . 1} is Borelmensurable. the fuzzy probability of the wi class is given by the vector of data X: P(wi | X) = where P(X) = ∑i μwi (X) = 1. The probability of a fuzzy event is defined by [64]: let (Rn . . is a robust method for data classification. By the use of fuzzy probabilities and fuzzy Bayes rule [54] in the classical Bayes rule. where M is the total number of performance classes available. . also called discrete or multinomial NB. Let μwi (X) be the fuzzy membership function for each class wi given by a fuzzy information font (for example. .

wi ) (60) (59) From successive applications of the conditional probability definition over (60). X2 . X2 . Xn \wi )P(wi ) ⇔ P(wi | X1 . X2 . . . Xn . . X2 . . . due to: P(Xk \wi . . . . X2 . . . . dos Santos Machado and R. X1 . wi ) = P(wi )P(X1 . X1 ) . . . . It means that knowing the class is enough to determine the probability of a value Xk . X2 . for all k = l simplifies the equation above. Xn \wi ) = P(wi )P(X1 \wi )P(X2 . From (51): P(X | wi )P(wi ) ⇔ P(wi | X) = P(X) P(X1 . . Xn . . X2 . X2 ) . . it is not relevant for data classification and can be rewritten as: P(X | wi )P(wi ) = P(X1 . Xn \wi )P(wi ) Equation (59) is equivalent to the joint probability model: P(X1 . Xn . . . . (65) . Marcos de Moraes on that space of decision. . Xn ) > P(w j | X1 . . the classification rule for Naive Bayes is similar those done by (52): X ∈ wi if P(wi | X1 . Xn ) = P(wi ) ∏ P(Xk \wi ) S k=1 Then. . Xn \wi )P(wi ) = P(X1 . a Naive Bayes classifier computes conditional class probabilities to predict the most probable class of a vector of training data X. . Xn ) for all i = j and i. Finally. . . P(Xn \wi . X = {X1 . . . X1 ) = P(wi )P(X1 \wi )P(X2 \wi . . . . . Xn ) = (58) P(X) However. . . X2 . . . Xn \wi .e. . X2 . Xn−1 ) (61) The Naive Bayes classifier receives this name because its naive assumption of each feature Xk is conditionally independent of every other feature Xl . as P(X) is the same for all classes wi . = P(wi )P(X1 \wi )P(X2 \wi . where X is a vector with n features obtained when a training is performed. . X1 )P(X3 . can be obtained: P(X1 . (58) can be expressed by: n 1 (64) P(wi | X1 . . i. . . X2 . . wi ) = P(wi )P(X1 \wi )P(X2 \wi ) · · · P(Xn \wi ) (63) (62) n. X2 . . X2 . . . . . . X1 . . Xn . Xl ) = P(Xk \wi ) for each Xk and (60) can be rewritten as: P(X1 . This assumption unless a scale factor S. X2 . Xn }. . . . . . . . . . j ∈ Ω. Xn \wi . which depends on X1 . . X2 .. . . .114 L. . . .. .

4. Another problem is the proximity between a first penetration and the next (generally about millimetres). it was used a maximum likelihood estimator. Xn ) with ∗ = {i. .4 Quantitative and Qualitative Evaluation Based on Naive Bayes The evaluation system based on Naive Bayes can be adapted to analyze simultaneously qualitative and quantitative information and improve decision quality for special cases. X2 . multiple penetrations are required (up to 20 per patient) in order to obtain enough bone marrow for that transplant. in the bone marrow harvest procedure. Machado and Moraes [27] proposed a modified Naive Bayes to lead simultaneously with qualitative and quantitative variables and assist other requirements of an evaluation system for training based on VR [41].8. the repetitive procedures can cause loss of control or fatigue of physician. For this reason. named Pe : Pe (Xk \wi ) = #(Xk . wi )/#(wi ) Xk . the Assessment Tool based on Naive Bayes collects the data from their manipulation. . (66) where #(Xk . . Other approach is to use quantitative and qualitative variables simultaneously [11] and to compute its parameters from D. To estimate parameters for P(Xk \wi ) for each class i.Intelligent Decision Making in Training Based on Virtual Reality 115 and P(w∗ | X1 . as a fracture of the pelvic bone. However. In the traditional transplant technique. it is possible to apply Naive Bayes classifier in data with quantitative and qualitative variables using a discretisation method in the first stage to allow the use of the Naive Bayes method after. wi ) is the number of sample cases belonging to class wi and having the value . All probabilities of that data for each class of performance are calculated by (66) and. For example. #(wi ) as the number of sample cases that belong to the class wi . this approach can affect classification bias and variance of the NB method [62]. it is easy to find children with soft bone structures or bones with deterioration caused by disease. The information of variability about these procedures is acquired using probability models. mistaken procedures can cause serious injuries to the patient. some special cases can occur and it can require specific skills of physician. In the paediatric oncology. from the vector of the training data. Based on equation (64). . but to do not bypass it. j | i. j ∈ Ω }. which require sufficient dexterity of the operator. This procedure requires specific applied force to trespass the bone and to reach the marrow. In these cases. an expert performs several times the procedure labeling them into M classes of performance. Previously. at the end. is done by (64). When users perform the training in virtual reality simulator. the user is assigned to a wi class of performance by (65) [39].

n-c continuous variables. . Formally. c (0 c n) be categorical or discrete variables ob- tained from training data. According [6]. according to sample data D. . . Xc+2 . The continuous variables can be modeled by probability density functions. Augmented Naive-Bayes. where nodes represent variables in process and oriented arcs encode conditional dependencies between variables [44].116 L. X2 . X2 . dos Santos Machado and R. Marcos de Moraes Formally. which allows nodes to form an arbitrary graph. Cheng and Greiner [9] proposed a classification for Bayesian networks. it can be written as: P(wi | X1 . with X = Xcat ∪ Xcont . . . j | i. Thus. the modified Naive Bayes method computes conditional class probabilities and then predicts the most probable class of a vector of training data X. . The choice of a specific structure to knowledge representation depends on dependencies relationship between variables which describe that process. . . The parameters of the method are learning from data. . j ∈ Ω }. . a Bayesian network is defined as directed acyclic graphs.5 Bayesian Networks A Bayesian network is a probabilistic model that can represent a set of probabilities distributions from all variables in a complex process and also establish their relationships [44]. . The GBN is a generalized form of Bayesian networks. The graph G is a set of nodes and oriented arcs. Based on the same space of decision with M classes. Bayesian Multi-net nets e General Bayesian Networks (GBNs). X2 . . . X is a vector with n features. 4. Tree Augmented Naive-Bayes. obtained from training data too. Xc }.8. . rather than just a tree. . Xn }. All these distributions models can be adjusted and verified by statistical tests over the data [31]. . denoted by G and a probabilistic distribution denoted by P. where P(w∗ | X1 . Categorical variables can be modeled by multinomial distributions. is done by (67). let Xcat = {X1 . Another important characteristic is that each child node cannot be connected to the final classes of evaluation. Xn ) with ∗ = {i. The final decision about vector of training data X is done by (65). The dependencies are modeled by specific conditional probabilistic distributions [21]. in five classes: Naive-Bayes. That choice is critical. Xn ) = c n 1 P(wi ) ∏ P(Xk \wi ) ∏ P(Xk \wi ) S k=1 k=c+1 (67) and the classification rule is done by (65). because it changes the final results. according to their graph structure. Discrete variables can be modeled by count of events in sample data D or by discrete statistical probability distributions. Equation (67) defines a modification of Naive Bayes method. as in classical Naive Bayes and Xcont = {Xc+1 . .

If those dependencies are founding and. can have a bivariate Gaussian distribution. which did not occur in other structures.Intelligent Decision Making in Training Based on Virtual Reality 117 The General Bayesian Network is convenient to serve as base for a training evaluation due to its generality. . P(X1 . . For example. By the use of probabilities calculus is possible to find dependencies among nodes in a Bayesian network. Xn ) = P(X1 | w j ) P(X2 | w j ) P(X3 | w j ) (69) The node probabilities are associated to probability distribution. . an expert calibrates the system. a node A can have a Gaussian distribution and a node B. X2 . dependencies can be estimated using multivariate linear regression [44]. . Xn ) is the joint probability distribution and P(Xn | Xn−1 . The structure of GBN is learned from data. according M classes of performance. all combinations are searched and an arc is designed between two nodes to obtain an increment of initial score. This cycle is repeated until the total network score could be less than a predetermined value or a fixed number of cycles. . . X2 . . Let be w j . . Xn−2 . Xn−2 . Following. X1 . the class of performance for a user and Xi . . . this permits simplifications in (68). . . . At the end of the training. The users perform their training in the VR simulator and the Assessment Tool based on GBN collects the data from that manipulation. as well as the parameters of conditional probabilities. according (68). . If the conditional independence between variables is verified. . it is possible to verify dependencies between variables during network modeling and put them in structure nodes of GBN. Then. . Thus. . Xn−2 . . which depends on A. the assessment system reports the classification to the trainee. . Xn ) = ∏ P(Xn | Xn−1 . . with a mean vector and a covariance matrix [19]. a network is created with all independent nodes and an initial score is calculated. In a first moment. Previously. All probabilities of data for each class of performance are calculated by GBN and a wj class of performance is assigned to the user at the end of simulation. represents a node in GBN with n as the number of nodes in a graph. . X2 . X1 ) is the conditional probability of Xn conditioned by its predecessor nodes Xn−1 . The joint probability distribution (68) where P(X1 . In that network. Formally. Then. 1 in GBN for an event is done by: P(X1 . let the classes of performance be in space of decision Ω with M classes of performance. . . . The information of variability about these procedures is acquired using GBN based method [43]. the parameters for that nodes set are re-estimated using linear regression. if is possible to assume Gaussian distribution for nodes. the dependencies between nodes can adjust itself to real dependencies. Scores are used to help the estimation of the final structure of GBN for each class of assessment. X1 ) i=1 n i n. j ∈ Ω. .

each time the training is performed. Additionally. It means that the evaluation method to be used must evaluate the geometric design and also the variation related to the bone marrow harvest process. Then. users perform training and receive evaluation reports about the training. the path of needle is the main attribute to be measured. Marcos de Moraes 4. But. In some cases. A third perspective for the characterization of methods could take into account the training of each user or all performed trainings for a continuous evaluation. as applied forces and angles of needle path. there are other measures important to define the success of the procedure. complex simulations could demand a detailed report about errors and successes in the procedure execution. However. The effectiveness of each methodology depends on the context of the training simulated by VR and its adaptation to such context. there are two types of evaluation: that analyse one user and that analyse more than one user and also their interactions during the training simulation. In the second mode. machine learning methods and mixed methods (combining two or more of them for intelligent decision making). it is possible to affirm that the evaluation methods adequate to the first type of training could . the evaluation reports about the present training and also about the previous training. In the first case. Thus. Other example can be the simulator for training gynaecological exam. dos Santos Machado and R. the target is the internal region of the iliac bone for extraction of bone marrow. As example can be mentioned the training of bone marrow harvest in which a specific region of the body must be selected to insert the needle.118 L. In this case. In this case. These particularities define a set of evaluation methods that can be used. the evaluation is performed for each user and also for all users in order to generate a report about individual performance and its relation to the other users performance. could be enough to inform users about their condition. in the second case. it should identify if user touched all region necessary to complete the patient examination. In this training the user must palpate softly the vagina and cervix region with an angle that don’t hurt patient. logical methods. evaluation method must verify variations in forces and angles of touch. Other way to characterize evaluation methods for training could be related to the number of users evaluated simultaneously. Therefore. However. statistical measures).9 Considerations About Evaluation Methods Evaluation methods present in literature could be characterized in four main groups based on: descriptive measures (in general. The most important point related to evaluation methods for training based on virtual reality is the observation of the particularities of each training. This report includes a profile to allow users to identify recurrence of mistakes and strong points in the performance.

computational power and new algorithms are necessary to process multiple points of contact for haptic devices in order to provide realistic haptic reactions. haptic loops need to work at 1 000 Hz to allow user to feel tactile sensations and realistic 3D environments demand the processing and exhibition of two images with 30 frames per second rates. allows the development of safe. As the research area of evaluation methods for training based on virtual reality is recent. Future trends point to systems in which multiple users can be immersed in a same virtual environment. Additionally. whose architecture is shown in Figure 4. In this case. For those reasons. most haptic devices are little intuitive and need improvements to allow comfortable use for training purposes. a training simulator and its correspondent evaluation system are limited by the computational power available. Besides. and because VR simulators are real time systems. locally or through networks (collaborative activities) [40]. consequently. in better quality services. it results in better-qualified professionals and. . technologies for high immersion still are very expensive for social applications and this is a huge limitation for its use in large scale. 4. It explains why is difficult to cut functionalities of VR simulators: the sense of presence in the virtual environment cannot be compromised.5. Those technical limitations must be taken into account to construct a realistic simulator based on VR. it is the kernel of a larger system. It is important to observe that evaluation methods are more robust than a simple application of a pattern recognition method. Researches about autostereoscopic displays are in progress. The technologies for immersion in virtual environments still require the use of special glasses and other wearable devices. the evaluation method should fit to the architecture to take advantage of its functionalities. affordable and reliable tools for the training of new professionals. Socially.10 Future Trends and Conclusions Intelligent decision making for VR systems. and collects data from several devices and user’s interactions to perform evaluation.Intelligent Decision Making in Training Based on Virtual Reality 119 not be able to deal with the features of the second type of training. but their results will be improved in next years. For example. realistic. Nowadays. particularly for training evaluation. the several methodologies proposed cannot lead with all training situations because each training can use specific devices and/or specific interaction modes. computers limitation can demand specific architectures as flight simulators used for training pilots. Mostly. In fact.

120

L. dos Santos Machado and R. Marcos de Moraes

Computational performance is the most significant limitation for the use of evaluation methods in training based on VR. As VR systems must be real time systems by definition, the use of more complex and sophisticated evaluation methods is only limited by computational power. Some solutions were recently proposed and use embedded systems coupled to computer systems to execute evaluation methods [42]. However, it is a temporary solution while computers with larger processing power are not available at lower costs. Recently, another trend arose: continuous evaluation methods [37]. Continuous evaluation is a good tool used in present and distance learning to help the construction of the knowledge and the cognitive training. In training based on VR environments, the goal is to construct a diagnostic to help trainees to understand their difficulties along their several trainings.

Bibliography
[1] Alca˜ iz, M. et al. (2003), GeRTiSS: Generic Real Time Surgery Simulation, Studies in Health n Technology and Informatics, 94, pp. 16–18, IOSPress. [2] Baum, L.E. (1972), An inequality and associated maximization technique in statistical estimation for probabilistic functions of Markov processes, Inequalities, (3), pp. 1–8. [3] Baum, L.E. and Petrie, T. (1996), Statistical inference for probabilistic functions of finite state Markov chains, Ann. Math. Stat., (37), pp. 1554–1563. [4] Bezdek, J.C. (1993), A review of probabilistic, fuzzy and neural models for pattern recognition, Journal of Intelligent and Fuzzy Systems, 1 (1), pp. 1–26. [5] Bonnet, N. and Cutrona, J. (2001), Improvement of unsupervised multi-component image segmentation through fuzzy relaxation, Proc of IASTED Int. Conf. on Visualization, Imaging and Image Processing, pp. 477–482, Spain. [6] Borgelt, C. and Kruse, R. (2002), Graphical Models: Methods for Data Analysis and Mining, Wiley. [7] Borotschnig, H, et al. (1998), Fuzzy relaxation labeling reconsidered, Proc. IEEE World Congress On Computational Intelligence, FUZZ-IEEE, pp. 1417–1423. [8] Buchanan, B.G. and Shortlife, E.H. (1985), Rule-Based Expert Systems: The MYCIN experiments of the Stanford Heuristic Programming Project, Addison-Wesley. [9] Cheng, J. and Greiner, R. (2001), Learning Bayesian Belief Network Classifiers: Algorithms and System, Proc. Fourteenth Canadian Conference on Artificial Intelligence. [10] Dempster, A.P., Laird, N.M. and Rubin, D.B. (1977), Maximum Likelihood from Incomplete Data via EM Algorithm, Journal of Royal Statistical Society, Ser. B. 39, pp. 1–38. [11] Doring, C., Borgelt, C. and Kruse, R. (2004), Fuzzy clustering of quantitative and qualitative data, Proc. of the 2004 NAFIPS, pp. 84–89. [12] Dubois, D. and Prade, H. (1996), What are fuzzy rules and how to use them. Fuzzy Sets and Systems, 84 (1), pp. 169–185. [13] Fu, K.S. and Yu, T.S. (1980), Statistical Pattern Classification Using Contextual Information, Research Studies Press. [14] Gande, A. and Devarajan, V. (2003), Instructor station for virtual laparoscopic surgery: requirements and design, Proc. of Computer Graphics and Imaging, USA, pp. 85–90. [15] Gao, Z. and L´ cuyer, A. (2008), A VR Simulator for Training and Prototyping of Telemanipe

Computational Intelligence in Complex Decision Systems

121

[16] [17] [18] [19] [20]

[21] [22]

[23] [24]

[25]

[26]

[27]

[28] [29] [30] [31] [32]

[33]

[34]

[35]

ulation of Nanotubes, Proc. ACM Simp. on Virtual Reality Software and Technology, Bordeaux/France, pp. 101–104. Harstela, P. (1999), The Future of Timber Harvesting in Finland, Int. Journal of Forest Engineering, 10 (2), pp. 33–36. Hummel, R.A. and Zucker, S.W. (1983), On the foundations of relaxation labeling processes, IEEE Transactions on Pattern Analysis and Machine Intelligence, 5 (3), pp. 267–287. John, N.W. et al. (2001), Web-based surgical educational tools, Studies in Health Technology and Informatics, 81, pp. 212–217. IOSPress. Johnson, R.A. and Wichern, D.W. (2001), Applied Multivariate Statistical Analysis, 5th edition, Prentice Hall. Juang, B.-H. and Rabiner, L.R. (1990), The segmental K-means algorithm for estimating parameters of hidden Markov models, IEEE Trans. Acoustics, Speech and Signal Processing, 38 (9), pp. 1639–1641. Krause, P.J. (1998), Learning Probabilistic Networks, Knowledge Engineering Review, 13, pp. 321–351. Machado, L.S. and Moraes, R.M. (2003), An Online Evaluation Of Training in Virtual Reality Simulators Using Fuzzy Gaussian Mixture Models and Fuzzy Relaxation Labeling, Proc. IASTED International Conference on Computers and Advanced Technology in Education (CATE’2003), pp. 192–196, Greece. Machado, L.S. and Moraes, R.M. (2006), VR-Based Simulation for the Learning of Gynaecological Examination. Lecture Notes in Computer Science, 4282, pp. 97–104. Machado, L.S., Moraes, R.M. and Zuffo, M.K. (2000), A Fuzzy Rule-Based Evaluation for a Haptic and Stereo Simulator for Bone Marrow Harvest for Transplant, Proc. Phantom Users Group Workshop, USA. Machado, L.S., Mello, A.N., Lopes, R.D., Odone Fo., V. and Zuffo, M.K. (2001), A Virtual Reality Simulator for Bone Marrow Harvest for Transplant, Studies in Health Technology and Informatics, Amsterdam/The Netherlands, 81, pp. 293–297. Machado, L.S. et al. (2006), Assessement of Gynecological Procedures in a Simulator Based on Virtual Reality, Proc. 7th International FLINS Conference on Applied Artificial Intelligence, pp. 799–804, Italy. Machado, L.S. and Moraes, R.M. (2009), Qualitative and Quantitative Assessment for a VRBased Simulator, Studies in Health Technology and Informatics, 142, pp. 168–173, IOS Press. Mamdani, E.H. and Assilian, S. (1975), An Experiment in Linguistic Synthesis with a Fuzzy Logic Controller, International Journal of Man-Machine Studies, 7, pp. 1–13. McLachlan, G. and Peel, D. (2000), Finite Mixture Models, New York: Wiley-Interscience. Mohamed, M.A. and Gader, P. (2000), Generalized hidden Markov models - Part I: theoretical frameworks, IEEE Trans. On Fuzzy Systems, 8 (1), pp. 67–81. Mood, A.M., Graybill, F. and Boes, D.C. (1974), Introduction to the Theory of Statistics, 3rd ed., McGraw-Hill. Moraes, R.M. and Machado, L.S. (2003), Online Training Evaluation in Virtual Reality Simulators Using Gaussian Mixture Models, Studies in Health Technology and Informatics, 94, pp. 42–44, IOSPress. Moraes, R.M. and Machado, L.S. (2003), Fuzzy Gaussian Mixture Models for On-line Training Evaluation in Virtual Reality Simulators. Annals of the International Conference on Fuzzy Information Processing (FIP’2003), 2, pp. 733–740, China. Moraes, R.M. and Machado, L.S. (2003), Gaussian Mixture Models and Relaxation Labeling for Online Evaluation of Training in Virtual Reality Simulators, Proc. Int. Conf. on Engineering and Computer Education (ICECE’2003), Brasil. Moraes, R.M. and Machado, L.S. (2004), Using Fuzzy Hidden Markov Models for Online

122

L. dos Santos Machado and R. Marcos de Moraes

[36]

[37]

[38]

[39]

[40]

[41]

[42]

[43]

[44] [45]

[46] [47] [48] [49] [50] [51] [52] [53] [54] [55]

[56]

Training Evaluation and Classification in Virtual Reality Simulators, International Journal of General Systems, 33 (2-3), pp. 281–288. Moraes, R.M. and Machado, L.S. (2005), Maximum Likelihood for On-line Evaluation of Training Based on Virtual Reality, Proc. Global Congress on Engineering and Technology Education (GCETE’2005), pp. 299–302, Brazil. Moraes, R.M. and Machado, L.S. (2005), Continuous Evaluation in Training Systems Based on Virtual Reality, Proc. Global Congress on Engineering and Technology Education (GCETE’2005), pp. 1048–1051, Brazil. Moraes, R.M. and Machado, L.S. (2006), On-line Training Evaluation in Virtual Reality Simulators using Fuzzy Bayes Rule, Proc. 7th International FLINS Conference on Applied Artificial Intelligence (FLINS’2006), pp. 791–798, Italy. Moraes, R.M. and Machado, L.S. (2007), Assessment Based on Naive Bayes for Training Based on Virtual Reality, Proc. Int. Conference on Engineering and Computer Education (ICECE’2007), pp. 269–273, Brazil. Moraes, R.M. and Machado, L.S. (2007), Multiple Assessment for Multiple Users in Virtual Reality Training Environments, Lecture Notes in Computer Science, 4756, pp. 950–956, Berlin. Moraes, R.M. and Machado, L.S. (2008), A Modified Naive Bayes to Online Training Assessment in Virtual Reality Simulators, Proc. 3th Int. Conf. on Intelligent System and Knowledge Engineering (ISKE’2008), China. Moraes, R.M. and Machado, L S. (2008), Using Embedded Systems to Improve Performance of Assessment in Virtual Reality Training Environments, Proc. Int. Conference on Engineering and Technology Education (Intertech’2008), pp. 140–144, Brazil. Moraes, R.M., Machado, L.S. and Souza, L.C. (2009), Online Assessment of Training in Virtual Reality Simulators Based on General Bayesian Networks, Proc. VI International Conference on Engineering and Computer Education (ICECE’2009), Argentina, In Press. Neapolitan, R.E. (2003), Learning Bayesian Networks, Prentice Hall Series in Artificial Intelligence, Prentice Hall. Peleg, S. and Rosenfeld, A. (1978), Determining Compatibility Coefficients for Curve Enhancement Relaxation-Process, IEEE Trans. on Systems, Man and Cybernetics, 8 (7), pp. 548– 555. Rabiner, L.R. (1989), A Tutorial on Hidden Markov Models and Selected Application in Speech Recognition, Proc. of the IEEE, 77 (2). Rabiner, L.R and Juang, B.-H. (1993), Fundamentals of Speech Recognition, Prentice Hall PTR, New Jersey. Rheingold, H. (1991), Virtual Reality, Touchstone, New York. Rich, E. and Knight, K. (1993), Artificial Intelligence. New York: McGrawHill. Rose, F., Brooks, B. and Rizzo, A. (2005), Virtual Reality in Brain Damage Rehabilitation: Review, CyberPsychology and Behavior, 8 (3), pp. 241–262. Rosenfeld, A., Hummel, R.A. and Zucker, S.W. (1976), Scene labelling by relaxation operations, IEEE Transactions on Systems, Man and Cybernetics, 6 (6), pp. 420–433. Ryan, M.S. and Nudd, G.R. (1993), The Viterbi algorithm, Research Report CS-RR-238, Department of Computer Science, University of Warwick, UK. Shafer, G. (1976), A Mathematical Theory of Evidence, Princeton University Press. Terano, T., Asai, K. and Sugeno, M. (1987), Fuzzy systems theory and it’s applications, Academic Press Inc., San Diego. Tran, D. and Wagner, M. (1998), Fuzzy Gaussian Mixture Models for Speaker Recognition, Proc. Int. Conf. Spoken Language Processing (ICSLP98) special issue of the Australian Journal of Intelligent Information Processing Systems (AJIIPS), 5 (4), pp. 293–300. Tran, D. and Wagner, M. (1999), Fuzzy approach to Gaussian mixture models and generalised

Computational Intelligence in Complex Decision Systems

123

[57]

[58]

[59]

[60] [61]

[62] [63] [64] [65]

Gaussian mixture models, Proc. Computation Intelligence Methods and Applications, (CIMA’99), 154–158, USA. Tran, D. and Wagner, M. (1999), Fuzzy hidden Markov models for speech and speaker recognition, Proc. 18th Int. Conf. North American Fuzzy Information Society (NAFIPS’99), pp. 426–430, USA. Tran, D., VanLe, T. and Wagner, M. (1998), Fuzzy Gaussian Mixture Models for Speaker Recognition, Proc. Int. Conf. on Spoken Language Processing (ICSLP98), 2, pp. 759– 762, Australia. Tran, D., Pham, T. and Wagner, M. (1999), Speaker recognition using Gaussian mixture models and relaxation labeling, Proc. 3rd World Multiconf. on Systemetics, Cybernetics and Informatics/ 5th Int. Conf. Information Systems Analysis and Synthesis (SCI/ISAS99), 6, pp. 383–389. Tran, D., Wagner, M. and Zheng, T. (1999), A fuzzy approach to statistical models in speech and speaker recognition, Proc. FUZZ-IEEE’99 Conference 3, pp. 1275–1280, Korea. Weiss, P., Rand, D., Katz, N. and Kizony, R. (2004), Video capture virtual reality as a flexible and effective rehabilitation tool. Journal of NeuroEngineering and Rehabilitation, 1 (12), online (http://www.jneuroengrehab.com/content/1/1/12), BioMed Central. Yang, Y and Webb, F.I. (2003), On Why Discretization Works for Naive-Bayes Classifiers, Lecture Notes on Artificial Intelligence, 2903, pp. 440–452. Zadeh, L.A. (1965), Fuzzy Sets, Information and Control, 8, pp. 338–353. Zadeh, L.A. (1968), Probability Measures of Fuzzy Events, Journal of Mathematical Analisys and Applications, 10, pp. 421–427. Zadeh, L.A. (1988), Fuzzy Logic, Computer, 1, pp. 83–93.

Chapter 5

A Many-Valued Temporal Logic and Reasoning Framework for Decision Making

Zhirui Lu, Jun Liu, Juan C. Augusto, and Hui Wang School of Mathematic and Computing, University of Ulster at Jordanstown, Newtownabbey, BT37 0QB, Northern Ireland, UK E-mail: Lu-Z1@email.ulster.ac.uk, {j.liu, jc.augusto, h.wang}@ulster.ac.uk

Temporality and uncertainty are important features of real world systems where the state of a system evolves over time and the transition through states depends on uncertain conditions. Examples of such application areas where these concepts matter are smart home systems, disaster management, and robot control etc. Solving problems in such areas usually requires the use of formal mechanisms such as logic systems, statistical methods and other reasoning and decision-making methods. In this chapter, we extend a previously proposed temporal reasoning framework to enable the management of uncertainty based on a many-valued logic. We prove that this new many-valued temporal propositional logic system is sound and complete. We also provide extended reasoning algorithms that can now handle both temporality and uncertainty in an integrated way. We illustrate the framework through a simple but realistic scenario in a smart home application.

5.1 Introduction Decision making is a process of leading to a selection of a course of action among many alternatives and happening all over the world all the time. People try to collect as much information as they can to help them to perform the most appropriate decision for further steps. In nowadays society, large amount of information can be provided by media or other sources and loads of it may affect to people’s problems, however, it is rather difficult and even impossible for human being to manually deal with such amount of information. In such case, scientists have developed computer systems which help people to analyze that information according to some specific methodologies and return feedbacks which may guide people what and how to make the most suitable decision under certain situation.
D. Ruan, Computational Intelligence in Complex Decision Systems, Atlantis Computational Intelligence Systems 2, DOI 10.1007/978-94-91216-29-9_5, © 2010 Atlantis Press/World Scientific 125

126

Z. Lu, J. Liu, J.C. Augusto, and H. Wang

These systems are usually called Decision Support Systems (DSS), which couple the intellectual resources of individuals with the capabilities of computers to improve the quality of decisions [29]. Decision support system is an essential component in many real world systems, e.g., smart homes, disaster management, and robot control. These real world systems are characterized by uncertainty, where the states and transition between states are not deterministic due to various reasons, and temporality, where the transition between states is dependent not just on the current state but also on the time when the transition happens. Solving problems in such systems depends on the formal mechanisms used, e.g., logic systems, statistical methods, reasoning, and decision making methods. Logic is the study of principle of inference and demonstration [30]. In a formal system, logic should be able to investigate and classify the statements and arguments, through the study of inference and argument in natural languages. Logic-based formalisms play an important role in artificial intelligence realm and have made significant contributions both to the theory and practice of artificial intelligence. A completed logic should have its own axioms and inference rules to help it analyze the arguments or information from certain sources, and release the most reliable and reasonable result, which is what we require in the new DSS. Hence, a logic-based Decision Support System should be more rational, especially DSS based on the formal logic tools being able to represent and handle dynamic, inconsistent and uncertain information. There are logic reasoning systems that can be used to solve problems under temporal conditions, or under uncertain conditions. It is not hard to see that such a logic reasoning system would be quite complicated. The problem of decision making under uncertainty in a dynamic domain is computationally demanding and has recently been investigated by many researchers. Figure 5.1 shows the objective for our research framework to build up a many-valued temporal logic system for decision support, which is used to solve the decision-making problem containing uncertain information within dynamic environment, and expected to target many applications such as medical diagnosis problem, risk analysis, or some other similar problems which contain complex preference, uncertainty and dynamic environment. The present work only covers the preliminary results regarding the combination of two typical propositional logic systems with its theoretical properties, and the corresponding reasoning algorithms, illustrated by a simple but realistic scenario in smart homes.

A Many-Valued Temporal Logic and Reasoning Framework for Decision Making

127

Imprecision & Incomparable information

Representing and Reasoning for Building a Logic-based Dynamic Decision Support System

Logic-based dynamic decision support system

Logic system Preference Modelling Programming Language

Dynamic Environment

Figure 5.1 Our Research Framework Illustration.

5.1.1 Problem Description: a Realistic Scenario The ability to reason about both time and uncertainty is very important and desirable in a knowledge-based system. There are numerous applications with the need of reasoning about both time and uncertainty. For example, Smart Home systems (A Smart Home can be described as a house that is supplemented with technology, for example sensors and devices, in order to increase the range of services provided to its occupants by reacting in an intelligent way [5]), which rely on data gathered by sensors, have to deal with the storage, retrieval, and processing of uncertain and time constrained data. Consider a scenario where the task is to model a kitchen monitored by sensors in a Smart Home system. Let us assume the cooker is on (cookerOn, a sensor detecting cooker being activated), but the motion sensor is not activated (-atKitchen, atKitchen is a sensor detecting location of the patient in the kitchen). If no motion is detected after more than three units of time (umt3u), then we consider the cooker is unattended (cu). In this case, at the next unit of time, the alarm will be on (alarmOn) to notify the occupant. In this scenario, the cooker and the occupant are both monitored by sensors, but there may be a problem with these sensors which can not return accurate information about the status of cooker or position of the occupant, and some causal relationships may not be always certain. For example, due to the likely malfunction of sensor “atKitchen”, we can only assume that the patient is in the kitchen with e.g., 80 % certainty, or with ‘high’ confidence. What we want to know is if the alarm would be on or off under such uncertain and dynamic situation can be automatically inferred in order to

among others. provided 13 relations between different time intervals [1.128 Z. Cardenas Viedma et al. 12. the proposed system is similar to the one in [9]. among others.2 Relevant Works From the scenario giving above. 20]. 24. within the implementation of fuzzy set into time representation. in this scenario.C. 6. such that. many works have been done by researchers. J. however. some of them just focused on the uncertainty of time issue. 7. 32. The proposed reasoning framework aims to help deal with such uncertain and time constrained situations to make rational decisions. which.1. 19. it still . the uncertainty and temporality may exist at the same time. Mucientes et al. and Schockaert and Cock [28] extended classic time interval as used in temporal reasoning into fuzzy-time interval to make the reasoning system more flexible. lemmas and theorems which built up a theoretical foundation. One of the common approaches used to solve such decision-making problems with these characteristics is by combining many-valued and temporal logics. but it provided more mathematical definitions. how to solve this case is the main issue for developing the combined logic system. it shows that the smart home systems provided two kinds of information at least. Liu. uncertain and temporal information. such that. 8. Wang decide for further steps. However. Lu. 25. 18. [9] extended temporal constraint logic (TCL) into fuzzy temporal constraint logic (FTCL) which allows FTCL to handle uncertainty of time representation such as ‘approximately 90 minutes’. In this system. [23] suggested a fuzzy temporal reasoning method for robot control. the uncertainty may not only exist in states. which is one of interval temporal logic applications. 14. and H. However. EscaladaImaz [13] provided a many-valued temporal reasoning method which extends the truth value of a state to indicate the uncertainty of an element for real-time control systems. 10. such that. 31]. 2]. 5. we know that the rule may not always be true in some special case. from the real world experience. Allen suggested to separate the time issue into interval. it improves the ability of system for handling the decision-making problem which contains uncertain time issue. it implemented a valid area (only the interaction area of condition states) of conclusion part of rule. J. For the temporal property in logic can be referred to [16. From the above brief literature review. Some of them only considered the uncertainty of state and extend the truth value of state and gave the valid time interval for conclusion part. it shows that some of decision-making systems combining with fuzzy logic and temporal logic still have some weakness. can be referred to [26. To handle uncertainty through logic approaches. For example. they did not consider the uncertain information of state. Augusto.

5.e. our many-valued approach allows the consideration of uncertainty not only on states but also on rules. With it. 1] by using the Łukasiewicz implication operator. we follow the work from [15]. the temporality of this reasoning can be extended into time interval way for further research. We also present algorithms for forward and backward reasoning based on this logic framework. 1 − x + y} (x.2 Many-valued Temporal Propositional Logic Systems In this section. Uncertainty is introduced in the form of a many-valued logic and reasoning framework. by means of the Łukasiewicz logic [26. we present a novel reasoning framework for decision making under uncertainty and temporality. 1] with truth-values in [0. 31. This chapter aims to build up a framework of many-valued temporal reasoning in theory and provide an applicable algorithm for decision-making system to solve the real-world decision making problem. it should be able to solve the dynamic decision making problem which contains uncertainty of state and rule. In the rest of this chapter. We adopted a many-valued logic with truth values in the interval [0. 1] [26. . which integrates many-valued logic (mainly focus on Łukasiewicz logic) and temporal logic [15]. which allows users insert their owned problem and situation and return a reliable result which may lead the most acceptable choice. Łukasiewicz logic L[0. and prove the soundness and completeness theorems over the defined logical system. i. where 0 means ‘false’ and 1 means ‘true’. Łukasiewicz logic L[0. The temporal logic element is simpler than in the approaches cited above. semantic and inference rules. 1]) and −x = 1 − x is the ordinary negation operation. which provided a simple stratified causal rule for the time issue (only same-time rule and next-time rule) to allow a way to reason about the dynamic aspects of a system with less computational cost. As a difference with other proposals. How to handle such a problem is the main objective in this chapter. After this step. 21]. we provide the basic definitions for syntax. 31. which extends a temporal reasoning framework presented in [15] so that the representation and reasoning with uncertainty is allowed within the system.. We would like to build up a generic algorithm. y ∈ [0. We illustrate this framework through a simple but realistic scenario extracted from current research on the use of artificial intelligence applied to the design and implementation of Smart Home [3]. we can provide a simple program to identify if the system is suitable or not. 1] is a well known many-valued logic. where the implication operator “→” is the Łukasiewicz implication given by x → y = min{1. we outline a formal framework. In this chapter. 21].A Many-Valued Temporal Logic and Reasoning Framework for Decision Making 129 possibly exists in rule. Within this system.

x ∨ y = Max(x. and t1 . denoted as T . . If the operations are defined as follows: x ∧ y = Min(x. One can easily check that the truth-value degrees of the formulae derivable in the Kleene’s system and in the Łukasiewicz system do not. It implies that the formulae of their logics are syntactically equivalent to the formulae in classical logic. t2 . −. Then L is called the Łukasiewicz implication algebra on [0. Liu.1 Syntax and Semantics Following the notations given in [15]. Each atomic state comes in pairs.1. ∧. SD denotes the set of the dependent atomic states. in general. An independent state can only be initiated by the occurrence of initiating events. L will represent Łukasiewicz implication algebra. We also assume a set of rules. Lu. 1]. y). x ⊗ y = Max(0. . . . s1 ∧ s2 denotes the (non-atomic) state that holds when s1 and s2 both hold. 1] is by using the Łukasiewicz implication operator or some isomorphic forms of it.tn ∈ T . 5. . . coincide. ⊗ >. From now on. →. most important proposals and results used the Kleene implication (x → y = max(−x. 1]. sn ∈ S. x + y − 1) = −(x → (−y)). More importantly. characterizing relationships amongst states of the system. Formally. Wang Łukasiewicz logic has been studied in numerous papers on fuzzy and many-valued logic. and s1 . Augusto. So the development of a relatively efficient calculus based on the Łukasiewicz system seems desirable. we have: Definition 5. Let L =< [0. In extending the classical logical approach. SI denotes the set of the independent atomic states which does not depend on other states holding at the same time. denoted as R. −x = 1 − x. denoted as S. . Q = S ∪ R refers to as the set of all atomic states and rules.C. s2 . x ⊗ y = Max(0. J. Pavelka showed in [26] that the only natural way of formalizing fuzzy logic for truth values in the unit interval [0. 1]. x. y)). and H. . J.130 Z. y ∈ [0. whereas a dependent state can do so.2. x + y − 1). We also propose a set of time slot. −x = 1 − x = x → 0. we assume a set of atomic states. where ⊗ is called the Łukasiewicz product and → is called Łukasiewicz implication. and every . ∨. each positive atomic state s being paired with its negation −s. Note that. y).

TY = L ∪ {−. The semantic definition follows a universal algebraic point of view. and denoted as HoldsAt(p. 21]. If p is a state. Let X be a set of propositional variables. where each si is an atomic state and s ∈ SD . 1 − θ ). where s ∈ S. independent. θ ). Definition 5. represented as Ingr(s). ⊗ can all be expressed by – and →. i. θ ∈ [0. Ingr(s) is the event of s being initiated. or one of each. it may be that both p and −p are dependent.2. Every p ∈ Q is associated with a value A(p) ∈ [0. q ∈ LP(X). In this system. which may occur when an expert is unable to establish/predict a precise correlation between premise and conclusion but only with degrees of certainty. s ∈ S. θ ) is equivalent to HoldsAt(p. The certainty degree dictates to which extent a state is true. →}. it only considers two kinds of rules: Same-time rules: s1 ∧ s2 ∧ · · · ∧ sn → s.. 31. or credibility degree) of a proposition p (based on the application context). called a fuzzy premise set. 1]. Let A ∈ FL (Q). 1]. notice that Q = S ∪ R ⊆ LP(X). t. We also provide events to model impingements from outside of the system.A Many-Valued Temporal Logic and Reasoning Framework for Decision Making 131 time slot is independent to each other. A is called a non-logical fuzzy axiom set or fuzzy premise set. . In addition. Moreover. These events represent ingressions to a state. Time is represented as a discrete series of atomic instances. p ⊗ q ∈ LP(X) if p. note that ∨. labelled by natural numbers. p ∧ q. Note that L and LP(X) are the algebras with the same type TY. the knowledge in the proposed formal system can be represented as a triplet: (p. s1 ∧ s2 ∧ · · · ∧ sn → s represents the influence of s1 ∧ s2 ∧ · · · ∧ sn over s. ar(→) = 2 and ar(a) = 0 for every a ∈ L.t. but certain to some degree. →} be a type with ar(−) = 1. We assume that HoldsAt(−p. The semantic extension is also applied to rules. t. 1]. The propositional algebra of the many-valued propositional calculus on the set of propositional variables is the free T algebra on X and is denoted by LP(X).e.t. In order to manage uncertainty. Next-time rules: s1 ∧ s2 ∧ · · · ∧ sn → Os. see more details on this in [15]. so p ∨ q. θ ) which means p holds with the truth-value level θ at time t. Hence. ∧. Same-time rules are required to be stratified. where TY = L ∪ {−. FL (Q) represents the set of all the fuzzy sets on Q. as in [26. which is taken from [0. and s1 ∧ s2 ∧ · · · ∧ sn → Os represents delayed (next time) influence of s1 ∧ s2 ∧ · · · ∧ sn over s. we express that a state is not just true or false. If p ∈ Q. A(p) = α ∈ [0. they are ordered in such a way that the states in a rule are either independent or dependent on states which are heads of rules in previous levels of the stratification. 1] expresses that a state holds or a correlation holds with a certainty degree α . In a real-world application one may suppose that A(p) is the minimal truth-value degree (or possibility degree.

1]. . and sn holds. . 1]. (2) γ is a propositional algebra homomorphism with respect to the first argument. 1]. t2 ∈ T . (s. t ∈ T . β ). (sn .t) = α . . . β ) . same . Then γ is called a temporal valuation of LP(X). If α = 1 at time t. (3) For any t1 . α ⊗ β ) where α = Min(α1 . If there exists a temporal valuation γ such that γ (p) α at time t. J. γ (α .7. αn ). . Definition 5. and H.t. The definitions below cover all the possible ways the system can use to compute the value of a state based on other states and incoming events (which can also alter the truth value of states). . . J. . . αi ∈ [0. α1 ). β ∈ [0. then the truth-value level stating that s holds at the same time slot immediately is β .t2 ).2. Same-time β -Rule: (s1 ∧ s2 ∧ · · · ∧ sn → s.6 (Same Time Many-valued Temporal Modus Ponens Rule). then p is said to be valid with the truth-value level α at time t.5. where ∗ means any state belong to S. n).132 Z. . and sn ∈ S.t1 ) ≡γ (∗ . Definition 5. .3. 1] (i = 1. If a mapping γ : LP(X)× T → L satisfies the following properties: (1) For any α ∈ [0. The same-time rule is used to solve the truth value of the dependent state which is affected by independent state and/or another dependent state instantly.2 Inference Rules This section explains how reasoning with uncertainty in a dynamic system is performed in our system. β ∈ [0. which means that if s1 and s2 and . then γ (∗ .t. if t1 = t2 . According to the above definition.t. where s1 and s2 and . Lu. α ∈ [0.t. .C. Liu. Wang Definition 5. Next-time β -Rule: (s1 ∧ s2 ∧ · · · ∧ sn → Os. α2 . Definition 5. then p is said to be α -satisfiable at time t. . then p is valid in time t. αn )}. β ). . Definition 5.4. . Let T be the time set. . The time MTMP rule (s-MTMP) is defined as: {(s1 . 1]. Augusto. the same-time rule calculus is built up for application. Let p ∈ LP(X). If γ (p) α for every temporal valuation γ of LP(X) at time t. (s1 ∧ · · · ∧ sn → s. 5.

v) is used to express the event occurs and the rule for the occurrence of an event is defined as: Occurs(Ingr(s). while an event occurs. 1]. .t + 1. β ) . at t.t. σ ) to denote an ingression to s.t + 1. with the truth-value σ ∈ [0. v) → (s.t. i.t ∗. the next-time rule calculus is built up for application. αn ). then the consequence would be (−s. (s. .10.t − 1. v). α2 . (t − 1)∗. αi ∈ [0. αn )}.e. Occurs(Ingr(s). and sn holds. In addition. (s. . where s ∈ SI and v∗ . Definition 5. According to the above definitions. .t.t + 1. The symbol ‘O’ indicates that change on the truth value of state s will be delayed and take effect in the next time unit. this rule does not affect to the truth value of the related states instantly. . .t.10 provide a calculus for event. then the truth-value level stating that s holds in the next time slot is β . . The time MTMP rule (n-MTMP) is defined as: {(s1 . α1 ). α ⊗ β ). Occurs(s.A Many-Valued Temporal Logic and Reasoning Framework for Decision Making 133 which means that: if s1 and s2 and . the independent states can only be initiated by the occurrence of the initiating events.t. α ). 1]. Event Modus Ponens Rule (E-MP) The Event Modus Ponens Rule is defined as: (s. n). Definition 5. v) ↔ HoldsAt(s. where s ∈ SI and v ∈ [0. v).t + 1.8 (Next Time Many-valued Temporal Modus Ponens Rule). Different from the dependent state.. However. v∗ ) ∧ HoldsAt(s. i. then the truth value of state should be changed by E-MP. . Assume events occur instantly and an instant between t and t + 1 is denoted as t ∗ .t + 1. (s.9 and 5. 1]. Definition 5. An atomic state is said to follow the persistence rule if there is no event or rule that can be applied to affect the truth value of state s at time t. (s. The next-time rule is used to solve the truth value of the dependent state which is affected by independent states and/or another dependent state in next time slot. t ∗ . α ⊗ β ) where α = Min(α1 . β ∈ [0. β ) . 1 − (α ⊗ β )).t. 1]. . .e. it inherits the same truth value as that one at time t.t.11 (Persistence rule). next . We write (Ingr(s). . 1] (i = 1. such that. if s1 ∧ · · · ∧ sn → O(-s). Definition 5. Then at time t + 1. . .. (s1 ∧ s2 ∧ · · · ∧ sn → Os. β ) Definitions 5. v ∈ [0.9. (sn . .

t − 1. It follows from Definition 5.1.t − 1. J.3 Logical Calculus Definition 5. α ). we can see that it just inherits the truth value from the previous time slot. p1 and p2 ∈ LP(X). so p = s-MTMP(p1 . for any a. Then p = MTMP(p1 . From the above definitions. Hence [γ (p1 ) ⊗ γ (p2 )] → γ (p) = [γ (p2 → p) ⊗ γ (p2 )] → γ (p) = [[γ (p2 ) → γ (p)] ⊗ γ (p2)] → γ (p) = −γ (p) → [[γ (p2 ) → γ (p)] → −γ (p2 )] = [γ (p2 ) → γ (p)] → [−γ (p) → −γ (p2 )] = 1. which means they are considered from left to right until one can be applied. there are four different Modus Ponens types of inference rules in our logical system. and γ is any temporal valuation in LP(X) (so γ is a propositional algebra homomorphism with respect to any formulae in LP(X)). At a meta-logical level the following priority applies in their applications: E-MP>>s-MTMP>>n-MTMP>>Persistence.C. p1 and p2 ∈ LP(X). Wang Definition 5. such that.t. if: γ (p) = γ (M(p1 . The Persistence Modus Ponens Rule is defined as: (s. from the inference rule. 5. b. it follows from the property of Lukasiewicz implication that: b iff a → b = 1. γ (p1 ) ⊗ γ (p2 ). That is. p2 )) holds for any temporal valuation γ in LP(X). α ) (s. Liu. Persistence(s. The result for n-MTMP rule can be . Let MTMP denote both same-time and next time MTMP rules and p. p2 )) proved in the similar way.134 Z.13 (Logical Consequence). p2 ) is called a logical consequence of p1 and p2 . Lu. c ∈ L. J. and H. Then p = M(p1 . p2 ) is a logical consequence of p1 and p2 .2.13 that p = MTMP(p1 . γ (p) = γ (M(p1 . suppose p1 = s1 ∧ s2 ∧ · · · ∧ sn → s and p2 = s1 ∧ s2 ∧ · · · ∧ sn . Proof. Augusto. p2 ) is a logical consequence of p1 and p2 .12 (Persistence Modus Ponens Rule (P-MP)). γ (p1 ) ⊗ γ (p2 ) Lemma 5. p2 ) = s. Let M be an inference rule and p. and a Consider the s-MTMP rule. a → b = (−b) → (−a). a → (b → c) = b → (a → c). α ) The P-MP is used to handle the state which is not affected by any event or rule in specific time slot.

α1 ).α ) C} = ∨{B(C).α ) C. For each i.t).t. such that Cn = n-MTMP(C j . and (1) If Ci ∈ LP(X) and αi = A(Ci ) at time t.A Many-Valued Temporal Logic and Reasoning Framework for Decision Making 135 Definition 5. (Cn . C is called an α -logical consequence of A at time t. (C2 . 1 i n. II) If there exists a finite sequence in the following form: (ω. Ci ∈ LP(X). and C ∈ LP(X). . (3) If there exists an α -deduction of C from A at t. αn ). α2 ). B can be deduced from A at time t}. . If i = n. (2) If C ∈ LP(X) and C is not effected by the same-time rule. For each i. Ck ) and αn = α j ⊗ αk at time t + 1. α1 ). (C2 . αn = α . III) In case there is no event or rule changing the truth value of a state at time t. Definition 5.14.α ) C.t.16 (Many-valued Temporal Deduction). A (t. and for i = n. Definition 5. (Cn . .t. αi ) is the same as the above same time process. An α -temporal deduction of C from A is defined as follows: I) At a specific time t. (1) If there is an event Occurs(C. if for any temporal valuation γ in LP(X). αi ∈ L. 1]. . where Cn = C. denoted as (ω . (Ci .α ) OC. α ). Ck ) and αi = α j ⊗ αk at time t. such that. . Let A ∈ FL (LP(X)) be a fuzzy premise set.t. (t − 1)∗ . 1 i n. k < n. (2) if there exist j. α2 ). αn ). (1) if C ∈ LP(X) and α = A(C) at time t + 1. an α -deduction of C from A at time t. denoted it as A (t. γ is a temporal valuation γ in LP(X) satisfying A}. Ci ∈ LP(X). then C ∈ LP(X) and α = A(C) at time t. Let A ∈ FL (LP(X)) be a fuzzy premise set. and αi ∈ [0. or event occurring. . γ satisfies A at time t implies that γ (C) α at time t.t) : (C1 . occurs between time t − 1 and time t.t. Let A ∈ FL (LP(X)) and p ∈ LP(X). then denoted it as A +(t. or (2) If there exist j. or. next-time rule. and α = A(C) at time t − 1.t. k < i. such that Ci = s-MTMP(C j . where Cn = OC. Then it is called an α -deduction of C from A at time t + 1.t. Set Con(A)(C) = ∧{γ (C).t) : (C1 . A temporal valuation γ in LP(X) is called satisfying A at time t if γ (p) A(p) at time t for each p ∈ LP(X). then it should follow the Persistence rule. Here we set Ded(A)(C) = ∨{α . where C ∈ LP(X) and α = A(C). .15 (α -Logical Consequence). we will have C ∈ LP(X) and α = A(C) at time t. . αn = α . is a finite sequence in the following form: (ω . . denoted by A| =(t.

136 Z.1 (Soundness). Without lose of generality.t2 . If 1 < l(ω) < n. then γ (C) α at time t + 1. j < n. assume that Ci = si1 ∧ si2 ∧ · · · ∧ sin and C j = si1 ∧ si2 ∧ · · · ∧ sin → Cn . = tn = t and Cn = C. .4 Soundness and Completeness Theorems Theorem 5.2. the conclusion holds for each deduction. Without lose of generality. If t1 = t2 = · · · = tn−1 = t.t1 . αn = α . If t1 = t2 = . α ∈ [0. γ (Ci ) αi and γ (C j ) α j at time t + 1. then. assume that Ci = si1 ∧ si2 ∧ · · · ∧ sin and C j = si1 ∧ si2 ∧ · · · ∧ sin → Cn . Case 2. If 1 < l(ω) < n. C j ) and αn = αi ⊗ α j at time t. (2) If 1 < l(ω) < n. αn ). γ satisfies A at time tn implies that γ (C) Proof. Lu. J. and Cn = C. then we have to prove that it holds for l(ω) = n. If t1 = t2 = · · · = tn−1 = t − 1. j < n. γ (C) = γ (Cn ) A(Cn ) = αn = α at time t + 1. If there exists an α -temporal deduction of C from A in the following form: (C1 . then γ (C) = γ (Cn ) αn = α at time. Then it follows from Lemma 5. α1 ). at time γ (Cn ) γ (Ci ) ⊗ γ (C j ) αn = αi ⊗ α j . αn = α . where Cn = C or OC. . Let A ∈ FL (LP(X)). since γ satisfies A at time t. αn = α . Augusto. . . C j ) and αn = αi ⊗ α j at time t + 1. A(Cn ) = If l(ω) = 1. . α2 ). we shall prove by deduction on the length of ω (denoted as l(ω)) that if γ satisfies A at time t. since γ satisfies A at time t + 1.1 that αi and γ (C j ) α j at time t. then A(Cn ) = αn . γ (Ci ) t. at time t. Wang 5. Suppose we have i. For each i. tn = t + 1 and Cn = OC. (C2 . γ (Ci ) ⊗ γ (C j ) αn = αi ⊗ α j Case 3.1 that γ (Cn ) holds at time t + 1. then we have to prove that it holds for l(ω) = n. The proof is given as follows considering the different cases: α at time tn . Suppose we have i.tn . Case 1. and H. (Cn . we shall prove by deduction on l(ω) that γ A-satisfies S at time t + 1. Then it follows from Lemma 5. αi ∈ L. the conclusion holds for each deduction. J. and C ∈ LP(X). and Cn = s-MTMP(Ci .C. (1) If l(ω) = 1. Ci ∈ LP(X). Liu. then by induction hypothesis. then by induction hypothesis. . the conclusion holds for each deduction. then for any temporal valuation γ in LP(X). . and Cn = n-MTMP(Ci . tn = t. 1]. then γ (C) α at time t. then we have to prove that it holds for l(ω) = n. then A(Cn ) = αn . 1 i n.

This is just the classical Boolean logic case. Ci ∈ LP(X).t1 ). . there exists a B ∈ FL (LP(X)) (also deduced from A) such that B (s1 ) ⊗ B(s1 → s2 ) > γDed (s2 )..l} (LP(X)). Theorem 5. α2 ). Without lose of generality. and C ∈ LP(X).1) C implies that A there exists a deduction of C from A in the following form: (C1 . α ∈ L. In addition. . α1 ). .t1 . . Then it follows from the definition of Ded(A) that γDed (s) A(s). Case 4. So we need to prove the consistency of Ded(A) and Con(A). . where B ∈ FL (LP(X)) is the fuzzy set which is deduced from A. next-time rule or event occurring rule at time t. Persistence(C. we have: P-MP((C. α )) = (C.t2 .1) C. For each i. (Cn . it is easy to see that A| =(t.tn ). the completeness theorem can be equivalent to the statement that Con(A) Ded(A). .e. if γ satisfies A(p) ∈ {0. α )). For any temporal valuation γ of LP(X). γ is now a binary (t. it follows from the properties of → and ⊗ that γDed (s1 ) ⊗ B(s1 → s2 ) > γDed (s2 ). From the above soundness theorem. If t1 = t2 = . where Cn = C or OC. then according to P-MP rule. Suppose that C is not taken effect by any same-time rule.t1 ). Occurs(C. 1 Proof. αn = α . α ). . It is easy to see that the conclusion holds as well. α ) It is easy to see the conclusion holds. Let A ∈ FL (LP(X)). Because the system we focused on only considers a set of atomic states (denoted as S). for any p ∈ LP(X). αi ∈ L. so we assume that Ci = s ∈ S which is an atomic state. (Cn .2 (Completeness). Suppose that γDed (s1 → s2 ) < A(s1 → s2 ). i n. it only needs to prove that the temporal valuation γDed of LP(X) induced by the Ded(A) satisfies A. Then it follows from the definition of Ded(A) that γDed (s1 → s2 ) < A(s1 → s2 ) < B(s1 → s2 ) < 1.A Many-Valued Temporal Logic and Reasoning Framework for Decision Making 137 Suppose we have i < n.β ).1 can be stated as Ded(A) Con(A).tn . then γ (p) = 1.t − 1. t − 1. 1}. Hence. Obviously. Cn = C. A. tn = t. . Equivalently. Theorem 5. αn ). Based on the definition of Ded (A). = tn = t. According to the definition of Con(A). . so according to the definitions of syntax and semantic in LP(X). If α = 1. then there exists an α -temporal deduction of C from A in the following form: (C1 . and a set of rules (denoted as R). (C1 . (t − 1). assume that C j = s1 → s2 ∈ R. If C is an α logical consequence of A at time t. (t − 1)∗. (C2 . and Cn =E-MP((C. A ∈ F{0. and αn = α . . .t. γ (p) valuation. i.

very much as a query is answered in Prolog.138 Z. such that. Prolog) to identify how this system works and whether the result is reliable.g. They complement each other and are useful at different stages of the design of a system. it shows that although the system is simple. we can provide a program (e. from 0 until t. .3. The classical forward reasoning algorithm for stratified causal theory is introduced [15]. t. The forward reasoning algorithm is mainly used for simulations that can show how the system changes (all state truth values) as time evolves. This leads to a contradiction. t. which means that there are only two statuses in state. it is sound and complete. So we have B∗ (s1 ) ⊗ B∗ (s1 → s2 ) > γDed (s2 ). According to Theorems 5. 5.3 Practical Many-valued Temporal Reasoning There are two algorithms to implement reasoning within our system. J. The backward reasoning based algorithm instead is more efficient. 5.2.e. completes the proof. and it also assumes that all the rules in knowledge base are always true. In this section. Liu. So. It is a goal based mechanism and to find out if HoldsAt(p. One follows a forward reasoning style and the other works by backward reasoning.16) such that B(s2 ) = B (s1 ) ⊗ B(s1 → s2 ) > γDed (s2 ). A(s1 → s2 ). in the future. Therefore.. it is obvious that B∗ is also deduced from A.. Prolog). The main drawback of this algorithm is that to find out if HoldsAt(p. then it will only considers those states and times the rule base indicate are strictly needed to answer the query. it extends the classical forward reasoning algorithm into many-valued one which allows users to do the simulation with uncertain information. α ). there exists a B ∈ FL (LP(X)) which is deduced from B∗ (based on MTMP rule in Definition 5.1 Forward Reasoning Algorithm The forward reasoning algorithm is used to simulate the given decision-making problems containing their owned initial assumptions and list all the running results from t = 1 to n. and H. Augusto.C. Wang Set B∗ = B ∨ B . included both states and rules under uncertainty. α ) the algorithm will compute the value of all states time after time. Lu. true or false. γDed (s2 ) = Ded(A)(s2 ) This means that γDed (s1 → s2 ) B(s2 ) > γDed (s2 ). both of them provide a theoretical foundation for us to implement such a system into a program (e. i.1 and 5. J.g.

1]) or not (when its degree of truth is in [0. Output: a history of the values of all states up to a time t. α ).t − 1. if (s. which are specified by determining α ∈ [0. α ) holds and (s. then apply any next-time rule. α ). (a) for each Occurs(Ingr(s). notice that unless otherwise known (s. α ) that arrives. λ )). then assert (s. 1 − α ) holds then assert (s. and β ∈ [0. it applied such an algorithm to solve a healthcare scenario which is used to check if some specific case is true or false in a specific time slot under the initial assumption.t − 1. We compute the resulting history as follows: (1) At t = 0. an event occurrence list (E). . .2) of same-time rules. (ii) Apply persistence: for any co-k-dependent state s. Rn ∈ R.3. if (s. (b) for each co-independent state s. β ) for r ∈ R. 1 − α ) was not asserted. 2. a set of next-time rules. if (s. (i) For k = 1. initial conditions. α ) for s ∈ SI and β ∈ [0. We say a rule is live if it can be applied. . A threshold λ is assumed and used to determine if a state has supportive evidence of holding (when its degree of truth is in [λ .2 Backward Reasoning Algorithm The classical backward reasoning algorithm for this system has already been provided in [15. α ). we extend such a backward reasoning algorithm under classical level into many-valued level. 3. Input: a stratified set of same-time rules (Rs ). t ∗ . then assert (s. α ) is assumed for all s ∈ SI . This extension allows users to enquiry some uncertain temporal decision-making problem in specific way under their assumption. 1 − α ) has not been asserted.A Many-Valued Temporal Logic and Reasoning Framework for Decision Making 139 Input: a stratified set (see Section 2. This is called ‘applying persistence’ to the state s. a set of next-time rules (Rn ). . (c) Apply any live next-time rule. and (s. .t. and a query of type . 1] in (s. apply any live same-time rules under the order of increasing stratification level.t ∗ . 0. α ). 3.. 0.t. β ) for r ∈ R.t. apply any live same-time rule of Stage k. (t − 1)∗. 1] in (r. α ). which is a set of formula in the form Occurs(Ingr(s). In the paper [4].t. . 2. α ).t. 4]. (2) For t = 1. where Rs . 0.t − 1. an initial condition list (Ic ). Once all possible same-time rules were applied. 1] in (r. which is a set of formula in the form Occurs(Ingr(s). and an event occurrence list.. 5. 0. In this section.

5. n. We then apply the following steps (where αi represents the truth-value of pi at time t): (1) IF t > 0 THEN REPEAT a) IF HoldsAt(q. (t − 1)∗. . for i = 1.e. for i = 1. at time t − 1 (i. J. Wang holds(q. . applies this algorithm recursively for each member of the antecedent of the rule) THEN the query is true. applies this algorithm recursively for each member of the antecedent of the rule) THEN the query is true. . J. e) ELSE t = t-1 (i. Augusto. 2.4 Scenarios Consider a simplified scenario where the task is to model a kitchen monitored by sensors in a smart home system as mentioned in Section 5. but the motion sensor is not . i. θ ) ∈ E and θ the query is true. . 0. d) ELSE IF there is a next-time rule (p1 ∧ p2 ∧ · · · ∧ pn → Oq. θ ) ∈ Ic OR Occurs(Ingr(q). Lu.e. θ ) ∈ E and θ < α . β ) THEN IF αi 1 + α − β .1. c) ELSE IF there is a same-time rule (p1 ∧ p2 ∧ · · · ∧ pn → q. β ) THEN IF αi 1 + α − β .140 Z.t. (t − 1)∗.1. t. THEN the query is false. . 2. Liu. θ ) ∈ Ic OR Occurs(Ingr(q).. and H. a sensor detecting cooker being activated).t.e. α . . THEN α.C. Let us assume the cooker is on (cookerOn.. Output: the answer to the query holds(q. try to find if there is a proof that is true at an earlier time and the value persists) UNTIL (the proof of q or −q is found) OR (t = 0) (2) IF t = 0 THEN IF HoldsAt(q.t. n.. . b) ELSE IF HoldsAt(q.t. .e. α ) (this could easily be extended to also provide a justification). α ) by which we ask the system to prove that the truth value of state q is equal or greater than α at time t. at time t (i. θ ) ∈ Ic and θ THEN answer the query is true ELSE the query is false..t.

95). 80% certainty. cu → alarmOn.5 and provide the events as follows: Occurs(ingr(atKitchen). for various reasons.95). alarmOn → O(− cookerOn) Lets assume the rules above have different degrees of confidence attached: (R1 being the less reliable given the sensors involved: R1(0.. If no motion is detected after more than three units of time (umt3u). the cooker and the occupant are both monitored by sensors. due to the likely malfunction of sensor “atKitchen”. We need to decide if the alarm should be activated under such uncertain situation. ± hazzard. Occurs(ingr(cookerOn). –atKitchen∧cookerOn∧umt3u→ cu R2. e. 1∗ . R5(0. − umt3u. −cookerOn→ −umt3u R5.9). R8(1). −cookerOn→ −alarmOn R3. may not be always certain. ± atKitchen.9). umt3u. 0. we define the trigger level of all rules to be 0. Occurs(ingr(−atKitchen). status of cooker or position of the occupant. 0∗ . For example. R3(0. 0∗ . Let us assume the following representation of the scenario discussed above: Independent atomic states: cookerOn. e. Occurs(ingr(umt3u). 0. Sensors are. then we consider the cooker is unattended (cu).95).9).. the alarm will be on (alarmOn) to notify the occupant that the cooker has been left unattended. 0. −cookerOn→ −hazzard R4. In this case.g. −cookerOn→ −cu Stage 2: R6. The proposed reasoning framework aims to help to represent and reasoning with such situations and to help in the design and development of systems with such characteristics. Besides. we can only assume that the house occupant is in the kitchen with. atKitchen is a sensor detecting location of the house occupant is in the kitchen).R7. 0. Same-time rules: Stage 1: R1. unreliable and some relationships.A Many-Valued Temporal Logic and Reasoning Framework for Decision Making 141 activated (-atKitchen.9) Questions: 1) The occupant wants to know what will happen from time= 1 to time= 6.g. In this scenario. − cookerOn. R6(1). at the next unit of time.cu → hazzard Next-time rule: R8. R4(0. 4∗ . R7(1).95). R2(0. ± alarmOn.9). Dependent atomic states: ± cu. .

J. the value of alarmOn becomes 0. Liu. Wang 2) The occupant wants to check if the alarm will turn on automatically by the system at time= 5.9)) changes the value of atKitchen to be 0.25 by s-MTMP.2 at next time unit by R8 and by n-MTMP.9. This makes R1 live.9 by E-MP. Answer for question 2): Assume we want to know with a confidence level of at least 75 % whether the alarm will be triggered t = 5. According to R6. the change of alarmOn causes the value of cookerOn changes to be 0. The result of applying the forward reasoning algorithm up to time unit 6 to the scenario described above is summarized in Table 5.C. 0. 0. It also triggers R3.8 not 100 percent true. and what the result is at each time slot. From the forward reasoning algorithm. other related states are also changed their truth values into a low level. such that.5. the alarm is activated because of the unattended cooker is on for more than 3 time units. Through it. 0∗ . t = 3 and 4.8 by s-MTMP.25 by s-MTMP. the value of atKitchen is set to be 0. t = 5: Occurs(ingr(umt3u.9 and cookerOn to be 0. By R4. Such that. the value of cu is changed into 0. the system believes that it may lead to dangerous situation and turn on the alarm to notify the occupant that the cooker is on and unattended at the moment.142 Z. the values of the states we follow do not vary. when t = 5. Lu. Augusto. but it is uncertain and only contain 0. then we can use the following query (via Backward Reasoning algorithm): holds(alarmOn. the value of umt3u changes to be 0. 0.8 by s-MTMP.9)) and Occurs(ingr(cookerOn. the value of alarmOn becomes 0. the system turns off the cooker automatically at time 6 and provides a low truth value of cookerOn.1∗.8 is in a high level of trust.0. we can see what happened during the time passed.1 above. the truth value of hazzard becomes 0. 0. Answer for question 1) Let us follow the evolution of some key states in this scenario (Forward Reasoning algorithm): t = 1: Because of Occurs(ingr(atKitchen. By this change.1. Besides.1 by E-MP. From Table 5. and H. However. According to R5.25 by s-MTMP. thus. 0. t = 2: Occurs(ingr(−atKitchen. .25 by s-MTMP. then the value of hazzard becomes 0. t = 6: The triggering of R8 makes R2 live.9)) then the value of umt3u changes to be 0. 4∗ .9)).8 by s-MTMP.0∗. we can see that it provides a simulation of the scenario according the initial assumptions. By R7.75) and apply backward reasoning. J. the value of cu becomes 0.

95 0. i.9 0.9). the system will calculate that cu must be greater than 0. Because cu is a dependent state the system will identify the states it depends on: –atKitchen.95 0.95 0.8 0.1 0.9 0. so the system has a way to prove that each state involved in the query is supported at the right level (under or above the threshold 0.9). the intuitive outcome is obtained.95 0.5 threshold they can trigger rule 6.95 0. we can see that the backward reasoning algorithm always traces the related state of the query.95 0.8 5 0..9 0.e.1 0 0 0 0 2 0.95 0. so it decreases the time of . Occurs(ingr(−atKitchen).75 to change the truth value of alarmOn greater than 0.5 as requested by the rules).95 1 1 1 5 0.95 0.9 0.95 1 1 1 Because alarmOn is a dependent state. such that.95 0.1 0 0 0 0 3 0.9 0.75. cookerOn and umt3u.9 0 0 0 0 1 0.9).9 0.A Many-Valued Temporal Logic and Reasoning Framework for Decision Making 143 Table 5.95 1 1 1 6 0. From R6.25 0. The results show that according to the initial setting of the scenario and the inference performed by our system.95 0.5 threshold and cookerOn and umt3u are above the 0.25 6 0. The system can trace back in time the knowledge to the points where specific events triggered those values: Occurs(ingr(cookerOn).2 0. 4∗ .95 1 1 1 4 0.0∗.9 0. 0.95 0.95 0.9 0. 1∗ .9 0. the backward reasoning algorithm provides a more sufficient way to identify the specific query than only forward reasoning algorithm in the decision-making system. in this case R6.95 0.95 0.9 0. then the system will start by finding a related rule. Compared to the forward reasoning algorithm.95 1 1 1 3 0. From the above example. the alarm will be triggered to notify the occupant when the cooker has been left on without attention for a specified length of time.9 0. it would ignore some unrelated components and avoid the unnecessary paths.25 0.95 0.95 0.95 0. and Occurs(ingr(umt3u).95 0.1 Exhaustive simulation by forward reasoning Time State cookerOn atKitchen umt3u cu hazzard alarmOn Time implication R1 R2 R3 R4 R5 R6 R7 R8 1 0. Because at t = 5 atKitchen is under the 0.8 0.95 1 1 1 2 0.1 0. 0.9 0.1 0 0 0 0 4 0.25 0. 0.

However. such as sensors can be reflected into our reasoning system. from the uncertainty issue. which means that R1 is available for most of cases. but if there are some effects which may last several time slots (time interval case). the uncertain information collection from Smart Home systems. and we have also presented practical algorithms for forward and backward reasoning. such that. but also applied into rules. so the system may need to have an event Occurs(-cookerOn. we have provided basic definitions of syntax. and if the system wants to trigger R1. but if we can extend the same-time rule and next-time rule into time interval form based on [17]. and inference rules for a many-valued temporal propositional logic system that allows us reason under both temporality and uncertainty. it needs to provide a high truth degree of each state in condition part. We have proved the soundness and completeness of this logic system. We will also consider extending the reasoning framework to a . it may need to create more states or events to handle it. So. the further research we may do is how to extend the time issue into a time interval situation. 1] interval. if the occupant sets up the cooker that should be open for 30 minutes (time controller). 5. For example. Lu. and H. the uncertainty is not only implemented into states. The results from the reasoning system may be more reliable by handling uncertainty both on states and rules. all the truth values are in the [0. Augusto. 1). 5. In future work we will implement our reasoning framework in Prolog and explore problems in other domains. J. semantic. this system can not be applicable. the further research should be implementing the lattice values which can handle the non-numerical uncertain information based on [31]. the truth values of rules show that rules may not be always true in smart home systems and they may contain uncertain information in some specific cases. it requires more accurate information for trigger. the system can only predict or handle some situations in the same time slot (same-time rule effect) or next time slot (next-time rule effect).30.6 Conclusions In this chapter. J. If the information can not be represented in a numerical form. So far. but in the real world application.5 Discussions According to the simple scenario. it is not sufficient to represent all the uncertain information. Wang searching and calculating process to improve the sufficient of the query process. by the definitions. For example. but not all of them. such that. Moreover.C.144 Z. Liu. this added event may be avoided. So. there is only a 80 % truth degree of R1.

: Methodology. December.M. Int. Co-located event of IJCAI 2007. Elsevier. Dubois. [10] R. [14] D. 21 (3) 209–247. 1994. Algebraic foundations of many-valued reasoning. India.A. HadjAli. Fuzzy temporal constraint logic: a valid resolution principle. [15] A. Berlin. of 9th Int. Augusto. 2000. Springer. 6th-7th of January 2007. pp. Liu. Yang. Navarrete Sanchez. [7] J. 1992. Ambient intelligence: applications in society and opportunities for AI. Bulgaria. . [13] G.J. Croft and J. and M. 1. Hodkinson. International Journal of Computational Intelligence System. D’Ottaviano. on Artificial Intelligence (IJCAI-07). Symposium on Temporal Representation and Reasoning. Emerson. [5] J. Temporal Logic (Volume 1): Mathematical Foundations and Computational Aspects. Hyderabad. Special Issue on Actions and Processes. the MIT Press. U. Vol. Bibliography [1] J. A General framework for Reasoning about Change. Borowik. Temporal and modal logic. on Art. Wang. 91–100.A. 1(4): 361–378. Mundici. J. [6] J. Prade. 33/1 pp. and H. A temporal many-valued logic for real time control systems. Proceedings of 2nd Workshop on Artificial Intelligence Techniques for Ambient Intelligence (AITAmI07). JUCS. Joint Conf. [2] J. and J.Computational Intelligence in Complex Decision Systems 145 more general lattice valued logic framework and to consider different types of time durations whilst retaining simplicity. Hyderabad.C. Cook. 52–54. Temporal reasoning for decision support in medicine. Handbook of Theoretical Computer Science.C. 2003. Actions and events in interval temporal logic. New Generation Computing. Enhanced healthcare provision through assisted decision-making in a smart home environment. Augusto and D.C. and Applications. Augusto. Management of uncertainty and spatiotemporal aspects for monitoring and diagnosis in a Smart Home. Journal of Logic and Computation. C. [11] D. Chapter 16. Augusto. Theoretical Foundations. Systems. 1994. Ltd and Springer-Verlag. P. Marin Morales. In Proc. 2008. McCullagh. Cardenas Viedma. Tutorial Lecture Notes. Augusto. [9] M. I. pp. 2003. Ferguson. [12] E. H. of Intelligent Systems. [4] J. Gabbay. 9 (9): 1168–1195. Fuzzy Sets and Systems. 2000. Artificial Intelligence in Medicine. Fuzziness and uncertainty in temporal reasoning. pp. 2005. 117: 231–250. 1990. 1991. Reynolds. Oxford University Press. Walkden. [8] L.K. [3] J. Allen and G. Dordrecht: Kluwer. R. 1–24. Conf. Many-Valued Logics. McCullagh. Cignoli. Time and time again: the many ways to represent time. India. Ohmsha. 6 (4): 1–14.C. In Proceedings of the 9th Int. A. Bolc and P. 27–32.-A. and D. and I. Stratified causal theories for reasoning about deterministic devices and protocols. Acknowledgment This work is partially supported by the research project TIN2009-08286 and P08-TIC3548. P. 2001. Intl J. Galton and J. I. 4: 531–579. 2007. Augusto. Allen. Escalada-Imaz. V. 20th Int.

Lattice-valued Logic: An Alternative Approach to Treat Fuzziness and Incomparability. Song. Temporal Logic and State Systems. A resolution-like strategy based on a lattice-valued logic. Logik und Grundlagend. Mockor. 2008. and J. Causal Reasoning for Alert Generation in Smart Homes. [28] S. Augusto. New York: Litton Educational Publishing. De Cock. [32] Y. Iglesias. Artificial Intelligence.C.).. J. [26] J. London: Academic Press. [22] Z..V. C. 2006.P. C. 9. 2003. of the 8th Int. and S. in Goble. 447–464. 2003. vol. Baldock. 1987. [27] N. F. Wang [16] A. K. I. Spain. Ruan. IEEE Trans. Turban. Fuzzy Sets and System. Galton. 1989. Liu. FLINS Conference on Computational Intelligence in Decision and Control. Hajek. Blackwell. Mucientes. Metamathematics of Fuzzy Logic. 57–70. 301–306. McGraw Hill. pp. J. Schockaert and M. 11. and H. Zeitschr. Qin. Springer-Verlag. Multi-valued temporal reasoning framework for decision-making.. The Blackwell Guide to Philosophical Logic. 2008. 2008. 2001. September 21–24. Mathematical Principles of Fuzzy Logic. and Z. D. Lu. 1969. Fifth Edition. Lu. Introduction to Logic. Suppes. 172: 1158–1193. R. [25] V. Augusto and H. Decision Support Systems and Intelligent Systems. Novak. J. [30] P. Inc. 25: 45– 52. Dordrecht: Kluwer. Merz. 1998. [23] M. Fuzzy Sets and Their Applications. [24] V. Liu. and J. 134:83–99. Novak. 2003. [20] F. J. ed.E. Kr¨ ger and S. Regueiro. Springer. Studies in Logic and Computation. Temporal Logic. In Proc.146 Z. A fuzzy temporal rulebased velocity controller for mobile robotics. Wang. 1957. A. Dec. 2000. Designing Smart Homes. [17] A. o [21] J. Barro. Math. Math. [29] E. A Treatise on Many-Valued Logics. 1997. J. 11(4): 560–567. Liu. Xu. [19] P.. Liu. 119–134. Temporal reasoning about fuzzy intervals. [31] Y. II: enriched residuated lattices and semantics of propositional calculi. Perfilieva. [18] S. D. Prentice Hall. C. Y. Bugarin. The Springer. Gottwald. on Fuzzy Systems. Augusto and C. D. Xu. and J. Aronson. Many-Valued Logic. Nugent (eds. Bristol: Adam Hilger. Dordrecht: Kluwer. On fuzzy logic I: multiple-valued rules of inference. pp. Venema. Rescher. Temporal Logics and Their Applications. Galton. Research Studies Press Ltd. 1979. . Lou. Pavelka. Ruan. 2001. New York. III: semantical completeness of some multiplevalued propositional calculi.

Pleinlaan 2. Ruan. it provides a statistical tool for interpreting decision-making processes for large samples of customers. Belgium E-mail: pkunsch@vub. correlations between criteria.be Complex decision problems have most of the time multi-criteria dimensions. Belton and Stewart [4]. which is valid for any dimension. This methodology serves two objectives: firstly. like rank reversal. It has been moving in recent decades towards Multi-criteria-decision aid (MCDA) (see Roy [1.ac. it provides a support for multiple-criteria ranking of alternatives in the presence of uncertainties. Malhotra [6]. This chapter investigates an original approach using the correlations between criteria. Figueira et al.g.Chapter 6 A Statistical Approach to Complex Multi-Criteria Decisions Pierre L. rather than to provide a final decision. addresses uncertainties. The proposed methodologies are therefore basically normative: the intention is to assist the decision-making process.1 Background Multi-criteria decision-making (MCDM) has become a very important instrument of Operations Research. Many existing multi-criteria decision tools are however plagued with difficulties due to uncertainties on data and preference weights. Geometrical interpretations are given for two and three-dimensional problems. secondly. multiple decision-makers. As a consequence. Such normative approaches necessitate a rather tedious process of preference elicitation. 2]. as a measure of distance between ranking solutions.1007/978-94-91216-29-9_6. system-dynamics simD. BE-1050 Brussels.. Kunsch MOSI department. and how it enhances the rank-reversal immunity.. such methods are not directly adapted to descriptive models. like in many marketing-research applications. or clients on markets. Atlantis Computational Intelligence Systems 2. © 2010 Atlantis Press/World Scientific 147 . see e. and there are therefore practical limitations on the number of participating decisionmakers. DOI 10. Vrije Universiteit Brussel. Vincke [3]. and last but not least. 6. and several future research directions are also indicated. undesirable properties. [5] for overviews of existing techniques). etc. The on-going development of the approach. It is shown how the proposed framework. Computational Intelligence in Complex Decision Systems.

Finally. agent-based modelling (ABM). they leave no space to the way real people are coming to a final ranking. In many situations it is impossible to find a justification for this disturbing phenomenon. to gather all relevant technical parameters. and therefore statistical treatments are required.. The latter is not an outranking method. Ref. [9.8]. because MCDA approaches are normative. 20]. have shown that rank reversals appear quite frequently with the PROMETHEE method. The main objective is here to apprehend in a credible way the behaviours of usually many agents with different preferences. e. see also Maystre et al. see Awang and Triantaphyllou [25]. it is difficult to include dependencies between criteria. Outranking methods are based on pair-wise comparisons.g. De Smet et al. [7. For the coming discussion I draw from my own contribution [24] to this problem. 10]. This constitutes an important obstacle for meaningful modelling. it is therefore difficult. and/or mimicking. irrelevant one. see e. [14]. in outranking methodologies. It is only recently that rank reversal has been shown to appear many times in the ELECTRE methodology. Another very important drawback for the correct and robust use of MCDA methodologies in a statistical framework is the important anomaly of rank reversal: there is no independence of the ranking of two alternatives with respect to a third. but it is also based on pair-wise comparisons of alternatives. 2. Kunsch ulation models like the choice of a departure time of car commuters by Springael et al. see Brans and Vincke [15]. consumers in some market. Macharis et al.g. etc. 13]. A simple example is extrapolating the market-share evolution of different product brands. Belton and Stewart [4] first identified the mechanism of rank reversal in the Analytical Hierarchy Process of Saaty [21]. for which . Furthermore. Brans and Mareschal [17. Brans et al.148 P. PROMETHEE. 19. [11]. and to establish how they affect the comparisons. For large sets of decision-makers. e. These MCDM problems are characterized by a large number of decision-makers. like ELECTRE developed by Roy [1.g. and to validate them against the observations. or impossible. 18. Ferber [12].L.. and an answer note by Figuera and Roy [26]. and giving some improvement capabilities of the brands.. These difficulties are an obstacle for using these wellknown techniques in many practical problems. in Marketing Research [6]. moving decisions in urban areas by Kunsch et al. weights. [23]. for analysis. Kunsch [24]. [16]. All require assumptions on scales. [25] makes reference to important test criteria concerning the invariance properties regarding rank reversals that must be claimed for robust MCDM methodologies. more specifically when using the PROMETHEE methodology. indifference and preference thresholds. based on initial consumers’ enquiries. De Keyser and Peeters [22]. etc. by definition of decision aiding.

. C2 . Note that these assumptions are made without loss of generality. . as serving as a normative MCDA instrument. . n-vector: any real vector in Rn .2. . n. 1]K . H¨ rdle and Hl´ vka [27]. Unit n-vector: vector obtained by centring the n-vector to its mean value and normalising it to one. 6. K criteria.4 presents the more general methodology. . It is also assumed that all decision-makers share the same knowledge about the criteria.Ck . It is easily shown that the technique is immune against the denounced non-rational rank reversals [24]. .2 Basic Elements of the MCDM Methodology 6.1 An ordinal procedure for finding the Rank Matrix In the following it is assumed that there are many decision-makers representing multiple possible preferences between n alternatives. containing real ordinal or cardinal data. . . . which is provided in a n × K evaluation table. . can provide a a an extremely simple. . .. . Section 6. . In Section 6. . Let me first introduce some useful definitions. given K criteria. and robust MCDM technique with a dual use: this technique can be as well used for interpreting statistical data from observations. Alternatives: A1 . 2. Criteria: C1 . independent of preferences. In Section 6. . n alternatives measured by their evaluation vectors.. unambiguously defined by their rank vectors. and how it can be extended to the case of mixed ordinal and cardinal data.. and they may be easily removed when needed. and an outlook is given on future research on this approach. V : the set of all unit n-vectors. illustrative examples are presented with 2-D and 3-D geometrical representations. it is shown that basic concepts of Multivariate statistics. Ai . The latter is first developed in the case when only ordinal rankings of alternatives are available from the decision-makers for multiple criteria. the author presents the basics of the proposed MCDM procedure.3. it will be assumed that weights representing the priorities between criteria are fully stochastic in the ranges [0.1. .g. Section 6. see e. . bringing important properties described with more details in Appendix A. A2 . . 2. . i = 1. n .5 presents conclusions. . In this chapter. k = 1. To make statistical evaluations.2.A Statistical Approach to Complex Multi-Criteria Decisions 149 there are strong needs to adjust parameters to observed data. .

1. 682 + 1. for criteria k = 1. (2) For example. .5. K. a preference ranking A1 > A5 = A4 > A3 > A2 gives in this notation the rank n 2. i = 1. Rank vector: n-vector of ranks rik of n alternatives Ai . which is obtained by ranking the alternatives on the basis of an evaluation vector. 2. 2. P(7) = . Preference ranking: Ai > A j > Ak = Al > . indicates the preferences between alternatives: Ai > A j indicates that Ai is better ranked than A j . P(n): Number of ordinal rankings with ties for n alternatives. 4. . . P(6) = 4. The traditional way of indicating a rank. . obtained by aggregating all criteria rank vectors by means of MCDM procedures for vector optimisation to be defined later in this chapter. is to use the average tie rank: 1 trad (1) rik = l + (p − 1) 2 In this chapter. P(5) = 540 + 1. 1) in which all alternatives belong to the same indifference tie. . Unit criteria rank vector: unit n-vector of the criteria rank vector. 2. . 2. . Overall rank vector: n-vector indicating the ranks of the n alternatives in the overall ranking. . P(n) is given by the recursive formula: P(n) = 47. Kunsch Evaluation vector: n-vector providing ordinal or cardinal evaluations of the n alternatives for some criterion.L. say p alternatives at position l in the rank vector. 2). Ak = Al indicates that Ak and Al have the same rank. .. 4. . . 2. rather than (1. P(n) is always an odd number because it contains the trivial ranking (1. thus simply: rik = l vector (1. Overall ranking: Ranking provided by the overall rank vector. the integer value of the rank is used. and using an adequately defined indifference threshold for identifying ties of alternatives. Criteria rank vector: Rank vector containing the ranking obtained for some criterion. given fully stochastic weights with rectangular probability distributions in ranges [0. 5. as traditionally used. . n−1 i=1 i ∑ P(i)Cn + 1.150 P. P(4) = 74 + 1. 5. for reasons that will become clear later on. with the possibility of ties. . n. Associated rank vector (to an evaluation vector): Rank vector. Rank Matrix: the n × n matrix indicating the probability of any of the n alternatives to occupy any rank. P(1) = 1 (3) P(2) = 2 + 1. etc. . 1]K . P(3) = 12 + 1. 292 + 1. in case of a tie.5).

1.. otherwise some negative preferences would be introduced. Theorem 6. Consider now all possible unit n-vectors obtained by combining those K basis vectors with random non-negative coefficients. 1]K generate a set of n-vectors.A Statistical Approach to Complex Multi-Criteria Decisions 151 On the basis of these definitions. has a rank at most equal to (n − 1). which is generated by all possible linear combinations of these vectors n serving as basis vectors. Using K independent unit criteria rank vectors. they constitute a basis in V{K} . 1]K . But all Pareto-optimal linear combinations must necessarily lie within the solid angle spanned by the unit criteria rank vectors. with column vectors from V n .2. The set V{K} ⊂ V n has dimension K.e. Every vector in V n may be obtained as a linear combination of a basis made of (n − 1) independent vectors in V n . Any matrix. ignoring for the time how to identify ties by means of indifference thresholds..1 does not imply that the positive coefficients in the linear combination of basis vectors are identical to the actual preference weights used by decision-makers to mark . Call V{K} the subset of V n .e. The following Theorem 6. distributed according to rectangular distributions in [0. i. obtained from these unit n-vectors. Let us further consider the associated rank vectors to the combination unit n-vector (see definitions. These results have very important consequences for ordinal MCDM ranking problems. which is the dimension of the space of all unit n-vectors.1. Corollary 6. and its Corollary A. Proofs are provided in Appendix A. Consider a set {K} of K n n − 1 independent unit vectors in V n . i.1: Basic Property A.1. Proof. I now introduce a seemingly harmless. the set of associated rank vectors to these n-vectors certainly contains all ordinal overall rank vectors which are Pareto-optimal.1 applies: Theorem 6. and combining them linearly with K non-negative weights from [0. n Assuming K independent unit criteria rank vectors. above in this section). but very powerful Basic Property A. solutions of the vector-optimisation MCDM problem with stochastic weights.1. in short basis vectors.1.1. that any matrix consisting n of column vectors from V{K} has at most rank K. it could not be associated with any linear combination of the unit criteria rank vectors showing non-negative coefficients. and normalising them to unit vectors – note that the n-vectors are automatically centred through the property of unit basis vectors. Herewith a combination unit n-vector is obtained for each instance of the weight set.1. If a possible Pareto-optimal overall ranking would not appear in the set of associ- ated rank vectors. and all have non-negative coefficients.

n).5 OR (overall ranking) : A1 = A2 > A3 w1 . R2 are different. . Kunsch their priorities between the criteria. . This is just a mathematical trick: it is not meaningful from the point of view of preferences represented by the weights. so that there are in a tie regarding the overall ranking (OR). 2. . 2. so that it will not affect the overall ranking associated to the linear combination. this is not always the case. . the positive coefficient is thus equal ck = wk j to the preference weight. the positive coefc1 = w1 . In order to preserve the ranks with respect to the normalisation constants. . K = 2) and three alternatives (A1 . w1 = 0. respectively. w2 = 0. A1 and A2 have symmetrical positions. k = 1. . Because the weights are equal. . 5/3. 2). . σ2 = √ 3 UR1 = (4) (5) where σ1 . i = 1. w2 are the preference weights associated to C1 . c2 . . K (7) σ0 In the case that there is no tie in the criteria rank vector. the standard deviations used in the normalising process are not the same. and UR2 gives: R1 − (2. The overall ranking will however be affected by the fact that σ2 < σ1 . because of the tie presence.152 P. the following transformation must be applied: σ2 w2 (6) σ1 In general. 3). A3 is dominated by both. σ2 are the standard deviations of the rank vectors. the normalisation process is a mathematical operation. 2) R2 − (5/3. Call R1 and R2 the criteria rank vectors: C1 : A1 > A2 > A3 . . 2. . Because there is a tie in C2 . The mean values of R1 . C2 . which has nothing to do with actual preferences. 5/3) √ √ . A2 . C2 . rather than the traditional notation of (1).5 C2 : A2 > A1 = A3 . Indeed. σ0 being the standard deviation of the natural rank (1. given less dispersion in the case of C2 . To see that.. UR2 = σ1 2 σ2 2 1 σ1 = 1. c2 = w2 ficients to be used in the linear combination of the basis vectors are linked to the preference weights by the following relationship: σk . 1. n. R2 = (2. A3 ). R1 = (1. let us consider the criteria rank vectors and the normalisation process in the following simple example with two criteria (C1 . The centring and normalising process to unit rank vectors UR1 . . 2. because of the use of (2).L. Due to the use of unit n-vectors. but this difference is the same for all three alternatives. Combine now linearly these basis vectors with positive coefficients c1 . 2.

whatever the used approach. because there is no tie any more in the second basis vector: OR : (A2 = A1 > A3 ) (9) The only way for avoiding such influence of a third irrelevant alternative A3 on a pair (A1 . i.A Statistical Approach to Complex Multi-Criteria Decisions 153 The rank-reversal phenomenon could also occur when the linear combination depends on the normalising process. because A3 is dominated in all cases.14) changes to C2 : (2.A2 ). and K n − 1 independent unit rank vectors representing the criteria. because after the transformation in (7) the denominators of the linear combination terms of the unit criteria rank vectors are now the same. and finding the associated rank vector. The proof is trivial. using the unit criteria rank vectors with equal weights produces the following overall ranking.. the same associated rank vector is obtained than simply adding the weighed un-normalised rank vectors. Using the weighed sum of criteria unit vectors without the adaptation of (7) could produce a rank-reversal situation. This looks like a complication. the latter result is the same as obtained by simply adding the weighed un-normalised rank vectors. which is obtained by a simple equivalence consideration. 1]K . 3). as easily verified: OR : (A2 > A1 > A3 ) (8) This ranking is different from A1 = A2 > A3. Consider now in addition that C2 : (2. 2) in (2. It is expected that the overall ranking would not be affected. is indeed using (7). 1. All Pareto-optimal overall rankings are obtained by considering non-negative fully random weights drawn from [0. But the corresponding overall ranking is now certainly as indicated in (4). any Pareto-optimal overall rank vector is the associated rank vector to the weighted sum of the K rank vectors. a third alternative A3 gets worse in the second criteria. as follows: Combination rule of rank vectors: The linear combination of the unit criteria rank vectors with the positive coefficients c j ’s defined in (7) gives the same final ranking as the direct combination of the un-normalised criteria rank vectors with the preference weights w j ’s. when ties are present in some criteria. given some set of K preference non-negative weights. given the individual rank vectors per criterion. 1. . This immediately defines the procedure for aggregating rankings vectors corresponding to different independent criteria: Ordinal MCDM procedure: In a MCDM problem with only ordinal data for n alternatives. In the example of (4). indicated in (4).e. but in fact it is not: a strikingly simple procedure indeed results for obtaining the overall ranking.

This is due to the missing cardinal data. or new dominated alternatives are introduced. see Kunsch [24]. and the ranks of the alternatives in the pair are not affected. as by the given procedure the overall rank vectors is the weighed sum of the criteria rank vectors. Q = 0. providing information about the relative positions. This rank-reversal phenomenon would also occur here if the linear combination would be dependent on the normalising process. Rank reversals appear because these approaches. and it has suf- . is not respected by outranking methods. and while ranking give the same rank to two numbers differing by less than Q.L. which is dominated by the pair in all criteria. Consider an indifference thresholds Q. Kunsch The Pareto-optimal overall rank vectors obtained with this procedure satisfy the following invariance property: Invariance property of the ordinal MCDM procedure: When the ordinal MCDM procedure is applied for obtaining the overall rank vectors.154 P. In order to identify ties in the overall ranking obtained in the weighed sum of criteria rank vectors. Note that this simple invariance property. introduce interdependence between alternatives. rather than to (1). The need for avoiding interdependency between alternatives also justifies clearly why the ranks must be defined according to (2). n].g. when a third alternative with no better ranks. gets worse. and the linear combination must be made directly on the ranks. as shown in Awang and Triantaphyllou [25]. e. The components of the resulting combined n-vector are numbers belonging to the interval [1. so that the ranking within a pair is not immune against a third irrelevant alternative getting worse. or the addition of additional dominated alternatives. in particular ELECTRE [25] and PROMETHEE [24]. to avoid the dependency of the rank on the number of alternatives p involved in a tie. it is why (7) has to be applied to remove this dependency. Kunsch [24]. in case of purely ordinal rank vectors. which excludes not-explainable rank reversals. the following approach is best used: before calculating the weighed sum of ranks re-normalise the sum of weights to one. The proof of this invariance property is trivial.1. and changes of positions of alternatives. though it is convincing. the ranks of alternatives in a pair are not affected when an alternative. becomes worse. This phenomenon may be interpreted as resulting from Arrow’s impossibility theorem [26]. The proposed procedure is disappointingly simple. The same property applies when alternatives dominated by the pair in all criteria are added to the alternative set.. with any procedure. Note finally that some explainable rank reversals are unavoidable.

the trivial 1-D ranking (1. 1. 2. but it is certainly well justified. so that now up to K = n independent criteria. n + 1) in the space of higher dimension (n + 1). . An interesting . But it may not be possible to demonstrate by mathematical means only that this approach is the only ‘correct one’ to the ill-defined vector-optimisation problem. . . If for example n = 2. 1. .1. It is certainly not to be expected that people have a computer in their head. as further developed in appendix A. The (n + 1) dummy alternative is chosen such that it is dominated by all n ‘real’ alternatives.2. because they are quite bad in all respects: they are placed in a dominated indifference tie at the very end of the ranking.1 imposes a restriction on the maximum number of criteria. may be considered. K = 2. why are not considered in detail by the decision-makers. To solve the MCDM problem a third dummy dominated alternative is introduced. The projection of these basis vectors on the 1-D axis perpendicular to the vector corresponding to (1. 6.1 and 6. 3) on the 1-D axis. because they are anti-correlated.2 for additional explanations. the two unit criteria rank vectors are obviously linearly dependent.3.2. like in ELECTRE. The corresponding (n + 1)-vector is shown to be orthogonal to the set of n-vectors.1. 1. (2. 3) gives the 1-D situation. so that now the unit vectors corresponding to (1. 1) is the point projection at the axis origin of (1. 1). pointing in opposite directions along the 1-D axis. when considering (n − 1) independent unit criteria vectors. see sub-section 6. or threshold-rich preference functions. or that they use sophisticated graph-theoretical approaches. . 1. Nevertheless this procedure is expected to be close to the way actual people reason while aggregating partial rankings of individual criteria. The point is here that all criteria are made commensurable by working with unit vectors. as confirmed by K > n − 1. A dummy (n + 1) alternative stays for all alternatives. 2) and C2 : (2. and its projection is at the origin of the coordinate system using (n − 1) orthogonal basis vectors in V n . 1) into the game.2 The maximum number of criteria The ordinal MCDM procedure in sub-section 6. like in PROMETHEE. see below the 2-D visualisation in Fig. see sub-sections 6. and basis vectors. . Finally we are brought back to rather unsophisticated weighed sums of ranks. Note first that. This restriction has never been pointed out before. C1 : (1.3. 6.1. 1. it is always possible to add a dummy (n + 1) alternative. . . in all n criteria. 1. 3). 3) are used as a basis. which now becomes (1.3.A Statistical Approach to Complex Multi-Criteria Decisions 155 ficient mathematical underpinning. for all criteria. This addition brings the trivial n-vector ranking (1. to my knowledge.

what is the largest possible positive integer x. with two or more dummy variables in the dominated tie. in position (n + 1). given n alternatives? When the criteria vector does not exhibit ties between the significant alternatives. question is to know how many such dummy alternatives may be added to the dominated unique indifference tie? In other words. like: (1) Two or more criteria rank vectors are identical. smaller than.156 P.L.e.. given n alternatives: The actual maximum number of acceptable criteria in any MCD problem is given by the maximum rank of the basis matrix built with the criteria unit rank vectors as columns. This leads us to the following upper bound on the number of criteria: Upper bound on the number of criteria. Other dependency situations between criteria may happen in practice. because for x > 1. plus the number of criteria. Kunsch Figure 6. or equal to n. This number is. Some more theoretical investigation may precise the value of this upper bound on the maximum number of criteria in the general case. in any case. the number of independent criteria. placing dummy alternatives in a tie. i. and thus the largest acceptable K. which show ties in the n alternatives. the x last rows would be identical in the matrix built by the basis vectors as columns. x = 1 at most.1 Shows the 12 non-trivial rank vectors for n = 3 alternatives. in each criteria rank vector. with (n − 1) + x = K. In this case they may be combined into . the number of alternatives.

55]N . etc.1 is thus only valid in practice for a number of independent criteria. the dependent basis vectors have to be eliminated from the set.1. such that the maximum rank of the basis matrix is r < K. (3) It may happen that K does not exceed the limit.1. [23] for similar suggestions.45. and some unit criteria rank vectors are linear combinations with non-negative coefficients of a basis of L independent unit criteria rank vectors.. the latter may grow forever. generally not much larger than n. In this case. as it suits the decision-makers: it is clear from intuition. According to our knowledge this restriction has never been reported before in MCDM problems. (4) In case the number of criteria exceeds the upper acceptable limit. etc. their weights are shared according to this combination between these basis vectors. Simulation tests [25] also indirectly confirm these conclusions for ELECTRE . and (2) is not applicable.A Statistical Approach to Complex Multi-Criteria Decisions 157 one criterion. so that more criteria than the limit. In such cases it is better to adopt a hierarchical structure. In this way separate MCDM analyses are performed at different levels. and its full implications. see section 6. for example. say N. the more correlations between criteria appear. before starting the analysis. the more important criteria with respect to the weights have to be selected. The proposed procedure presented in sub-section 6. it may be appropriate to decompose the problem by selecting random weights in [0. but some basis vectors are linearly dependent. say L. like in AHP of Saaty [21].1. 0. have to be considered. and to consider for each instance of the weight set. there may be a point when they start outweighing by their very number more essential criteria like the technical performances. used for calculating the Rank Matrix. the L most important criteria. and their weights may be added. etc.2. (2) In case K is above the limit. that the more criteria beyond n. and Corollary A. and confirmed by the Basic Property A. but such limitation looks quite reasonable: when no limit is placed on the number of criteria. the price. such as to maximise the sum of criteria weights. that may artificially ‘push’ some alternatives in an unmerited way towards the top of the ranking.2. say L.1. the purchase of a car: when more and more small design features are added as criteria in the MCDM analysis. pursuing the analysis at higher level on the obtained more global criteria. up to this acceptable limit. see Macharis et al. Space is lacking for discussing in detail this important aspect. the number of alternatives. by first combining commensurate criteria of the same nature.2. Consider. (5) When several important criteria have equal weights.

. 2. [28]: n = 3 car types. This is in my opinion coherent with the above remark that the increased correlations between criteria may consolidate ranks in an artificial way. The curves are bell-shaped. For validation purposes a preference ranking of the criteria is asked as well. 1) to be ignored. 2. According to (3) there are for each criterion P(3) = 12 + 1 possibilities for the alternative rankings. corresponding to the permutations with ties of (1. 2) (10) An enquiry is performed on a sufficiently large sample of individuals (‘interviewees’) to collect their opinions on car rank vectors. with their percentages with respect to the full sample (100 %).2. 1. and their maximum values are observed to be around K = n. These results are used in a second step.1. when the car characteristics change. K = 2 = n − 1 evaluation criteria.g. C1 . In these tests. are offered on the market. The objective is here to deduce the preferences between criteria for each group. The rank vectors provided by several groups of interviewees are given in the upper part of Table 6. 3). the pair-wise outranking relations and the subsequent ranking of the alternatives is more likely to become stable than before’.1. 3). e. and in later steps. how to reconstruct the preference rankings of criteria. 1. C2 are considered. or new brands enter the car market. The inverse question is also useful. and to validate these results against the collected interviewees’ data. The analysis in this first step is expected to provide enough information on preferences. when analysing data from multiple decision-makers. independently of individual preferences: C1 : (1. 6. without performing anew an expensive market enquiry. . A pay-off table is available from which ordinal rankings are deduced for each criterion. Kunsch II and ELECTRE III.3 The inverse procedure for finding preference weights In sub-section 6.L. and how to perform a statistical analysis of the different preference groups in the decision-makers sample? Assume for applying the reverse procedure the example presented in Kunsch [29] from Ruan et al. C2 : (3. this includes the trivial ranking (1. The authors [25] observe that the decrease with larger numbers of criteria is due to the fact that ‘if the number of criteria is beyond some value. the rates of rank reversals in some sample examples from the literature are plotted against the number of criteria K for different numbers of alternatives n.158 P.2. a normative procedure has been discussed for aggregating the criteria rank vectors to an overall ordinal ranking. on the basis of an evaluation table with two criteria. for market surveys: given statistical results on overall rankings. In this way the changing market shares may be re-computed.

Label R % sample Ranking Weight C1 Weight C2 Preference R1 = C1 (15 %) (1. 1. 1) (12) Assuming that preferences of the interviewees’ groups have remained the same. The preferences are indicated in the lowest row of Table 6. the positive coefficients in the linear combination of the basis vectors for obtaining the rank vectors are equal to the preference weights.1. as all five groups have chosen Pareto-optimal rankings. 1. 1. C2 has a tie.1 correspond to the direction of the unit criteria rank vectors representing C1 : (1. compatible with the preferences indicated in Table 6. 3) 1 0 C1 alone R2 = C2 (8 %) (3. 1. see sub-section 6. it is easy to recalculate the new markets shares. as follows: C1 : (2. 2. and C2 : (3.1: C1 > C2 indicates that C1 is more important that C2 . C1 (C2 ) alone indicates that C2 (C1 ) has a vanishing weight in the linear combination.1 The rankings with their percentages for five group of interviewees from the observed sample in the Market survey enquiry from Kunsch [29] in Ruan et al. 67 %. due to some evolution of the brands. Because now. 3) 1. according to the Basic Property A. 2) (see Fig. C2 : (2. In a second step assume that the criteria rank vectors change.2.1).1. Because the two basis vectors contain no tie.1. 1. there is no need for such corrections in this example. C1 = C2 indicates that both criteria have the same importance. or zero. 6.A Statistical Approach to Complex Multi-Criteria Decisions 159 Table 6. The market-shares vector (MS1 ) for the three cars is computed as being: MS1 : (33 %. 1. These results are then compared to the interviewees’ declarations about their criteria preferences. 0 %) (11) In this evaluation. Any other ranking is obtained by combining linearly the criteria rank vectors used as basis vectors.2. 2. appropriate corrections to non-rational rankings are made. 3). equal shares of ranks are assumed for the cars involved in a tie. 3) 1 1 C1 = C2 R5 (16 %) (2.1. 3). by inverting the correlation matrix of the two basis vectors. 2) 0. This is done for each group by using the sets of two weights.1 for each preference group.577 C1 > C2 R4 (25 %) (2. according to (7).1. [28].155 C1 < C2 The two criteria in Table 6.5) in appendix A. 2.577 1. They are easily calculated from (A. all weights being positive. the c2 coefficient positive coefficient is related to the weight w2 .1 and its corollary A.155 0. indicated in the lower part of Table 6. 2) 0 1 C2 alone R3 (36 %) (1.

the matrix. 3) counter clockwise. 6. each overall rank vector is determined in a unique way by a fixed set of combination coefficients of the basis vectors. Kunsch as indicated in (7). 1).1 is easy to perform.2. and (2. respectively. 6.2. and Y -axis. and it becomes: c2 = σ2 1 w2 = √ w2 σ0 3 (13) For the three groups R1 . Note also that because the basis criteria are independent.1 is a logical one: starting from any ranking in any direction: e. the two weights are immediately obtained.1. starting from (1. as either one is vanishing. R2 . K = (n − 1) = 2. R4 . there are P(3) = 12 rank vectors according to (3). and n = 4. Assuming two maximum criteria. K = 2. 1) is perpendicular at the origin to the plane of this figure. The 12 non-trivial rank vectors are shown in Fig.3. which we first ignore. assuming that the aggregation procedure is indeed a weighed sum of ranks. The Pearson’s correlation between two unit rank vectors is equal to cos(θ ). 6. and so is the associated rank vector.1 The 2-D case For n = 3.3 2-D and 3-D Visualisation in the Ordinal Case In order to illustrate the ordinal procedure. 84 %. . making a 2-D representation possible. the sequence of rankings in Fig. 3). 1. 6. 1. has rank= n − 1 = 2. 2.g.L. according to the MCDM ordinal procedure in sub-section 6. A Rank Matrix is obtained for each interviewees group. Monte Carlo calculations are performed using the weight ranking of the group. Multiplying the probabilities of each car brand of being ranked first by the group percentages in the whole interviewees sample gives the modified market-shares vector (MS2): MS2 : (0 %. The trivial rank vector (1. 16 %) (14) The inverse procedure to the MCDM ordinal procedure in sub-section 6. 2. 1.1.1. or both are equal. The orthogonal unit rank vectors corresponding to the rankings (1. K = 3 are discussed in detail in this section.1. and to visualise the results in the 2-D and 3-D space the two cases of n = 3. which has all unit rank vectors as columns.160 P. According to the Basic Property A. R5 : C2 > C1 . not including the trivial rank vector (1. For group R3 : C1 > C2 . 0 vectors [27]. 2) have zero correlations: they may be chosen as orthogonal positive X-axis.2.1 and Corollary A.. This angle is equal to 30◦ θ π is the angle between these between any two adjacent rank vectors.

are obtained by performing Monte Carlo calculations on the weights. The MCDM compromise obtained with equal preference weights being given to both criteria is thus (1. and the (C1 .g. leading further to (2. and increasing A3 values. 3). 3). 3). 1]2 . 2. 2. 3) are possible Pareto-optimal overall rank vectors. c2 = w2 = 0. 1. c2 = w1 = 1. These frequencies are visualized. It is clear that all Pareto-optimal solutions of the (C1 . The sum of the linear-combination coefficients is not necessarily equal to one. or (2. the corresponding unit rank vectors are: 1 1 C1 : − √ .. or clockwise from (2. 2. and a quadratic normalisation rule applies.C2 ) MCDM problem must necessarily be one of the overall rankings associated with such combination vectors. √ . then a further increase of (2) beyond (1). Consider the rank vectors correspond to K = 2 criteria given as C1 : (1. 1. Thus the possible compromise rank vectors. are the following ones. see (15) for the unit vector (1. and C2 . 1. then an increase in (3) leading to (2. 1. 1. 1. 3). For arbitrary weights from [0. combining them linearly by means of two non-negative random coefficients in [0.A Statistical Approach to Complex Multi-Criteria Decisions 161 one finds successively: an increase of (2) towards (1) first leading to (1. 3) : c1 = w1 = 1. 1. as corresponding to the relative surfaces of the sectors lying between the basis vectors. because of the logical evolution sequence starting counter-clockwise from (1. and their positive coefficients. 2). only (1. they must be overall rank vectors lying in the circle sector between C1 . identical to the preference weights in this basis without ties. Note again that the positive coefficients are equal to the preference weights. 3). 3).C2 ) basis vectors. + √ 2 2 (15) 1 1 C2 : 0. (1.1. one obtains . 2) would not be logical: starting from C1 : (1. 1. see (7). 0. 3) it would move clockwise towards non-increasing A2 values. 3). 3) : c1 = w1 = 0.e. see (15) for the unit vector The positive combination coefficients are calculated by means of (A. (1. 3). √ 6 6 6 (16) C2 :(2. and re-normalising. c1 . 1. − √ . 3) : c1 = w1 = c2 = w2 = 1. etc.. as given by the Rank Matrix. 3). in the absence of ties in the basis vectors. i.5) in appendix A. 1] [2]. + √ 2 2 Using these vectors as basis vectors. The ranks frequencies of alternatives. 2. In this example. where both the ordinary and unit rank vector are indicated: C1 :(1. 2. C2 : (2. in this simple case. Choosing e. 3). as the combination result must be a unit vector. 1. 1 −1 +2 −√ . 1. c2 in the combinations. away from C2 : (2. 2. produces a unit vectors lying in the circle sector between C1 and C2 .

162 P. In these cases the Pareto-optimal overall rank vectors will either lie within. the matrix. 0 %) A2 : (50 %. when discussing the 3-D case with n = 4. 1). 6.1 and Corollary A. A. It is interesting to note that the traditional representation in (1) would produce the higher C4 symmetry.. but an additional 30◦ sector is now added. see Fig. i. 50 %. This visualisation shows that the rank vectors are organised in a symmetrical way around the unit sphere: they build a symmetry group. Using C2 : (2. 0 %) A3 : (0 %. 0 %.2. the 2-D overall rank vectors must always lie within the solid angle defined by the basis vector. K = (n − 1) = 3. which is certainly interesting to analyse. so that there is no C4 symmetry around the orthogonal XY Z axes. see sub-section 6.. 100 %) It is also clear from the 2-D representation why the preference weights are not the same as the positive coefficients in the linear combinations of basis vectors. They appear on the unit sphere in Fig.e. this representation would not be suitable. Note that for n = 3. Two types of octants are observed for each Z 0.2.3.1. 3. 1.2. Z 0 with respectively 16 and 13 unit rank vectors. 3) provide here the six directions of orthogonal XY Z axes.1. there are P(4) = 74 rank vectors according to (3). 1. and higher dimensional spaces.1. 6.Y ) plane are shown in two Figs. while A3 is dominated: A1 : (50 %. This is elaborated in the following sub-section. and any rank reversal is avoided. but this is beyond the scope of the present chapter. Z 0: but. A2 . which may be first ignored. A. But by application of (7) the frequency invariance is re-established. when the latter show ties. 3) must not change the ranks frequencies.1. which has all unit rank vectors as columns. This is not necessarily true for cases with n > 3. Projections in the plane from above and below perpendicular to the (X.1. as discussed earlier.2 The 3-D Case For n = 4.2 in Appendix A. or close-by the edges of the solid angle spanned by the basis vectors. because it introduces (17) . and four identical octants for each Z 0. 2) instead of C2 : (2. so that their representation in the 3-D space is easy. that seem to favour A2 for being ranked first.L. 1. 1. 50 %. the criteria rank vectors.2. 6. 1. According to the Basic Property A. Kunsch immediately equal ranks frequencies for both A1 . has rank= n − 1 = 3. This symmetry lowering is due to the use of integer ranks in (2). The six permutations in the rank vector (1. not including the trivial rank vector (1.

and using (7) for taking into account the tie effect on the weights in C1 and C3 .5 −1 0 −0.2 The unit sphere shows for n = 4 the p(4) = 74 non-trivial ranking vectors indicated by dark points. 17 %.5 Figure 6. 0 %. 100 %) A4 : (42 %. 3). 0 %) An equal-sharing rule of ranks probabilities is applied to alternatives involved in a tie. which may lead to rank reversal situations. Assume the following three independent basis vectors.2 0 −0. C2 : (1. 25 %. 57 %. 1. 1).8 −1 1 0. let us consider the 16-vector octant shown in Fig. 4. 3. 0 %. 41 %. 4. 0 %) A2 : (13 %. impossible to explain on a rational basis. 1) (18) A Monte-Carlo analysis is performed by using random weights in [0. 2.4 −0. The computed Rank Matrix is obtained: A1 : (45 %.3. dependencies between alternatives. 6. (19) . being the criteria rank vectors: C1 : (1.5 0 −0. This explains why A1 has a somewhat larger probability of occupying the first rank than A4 . C3 : (3.6 0.8 0.5 −1 0.2 1 −0.A Statistical Approach to Complex Multi-Criteria Decisions 163 1 0. 1] [3]. 0 %) A3 : (0 %. To further illustrate the proposed approach.4 0. 4. 30 %. 30 %.6 −0.

1. (1. so that this MCDM problem is equivalent to three 2-D problems. Figure 6. (1. (4). 1) diagonal these 12 + 1 solutions reconstitute all unit (n − 1)-vectors in the 2-D example (see Fig. The sketched solid angle spanned by the unit basis vectors described in the text is evidenced. the diagonal of the octant shown in Fig. 6. this explains the observed difference in the first-rank frequency. (4). (4).1. compare Fig. This is a case where a dummy dominated variable (here A3 ) is added in order to be able to address the larger K = 3 MCDM problem. including the trivial (n − 1 = 3) rank vector (1. considering that the dominated A3 has the fixed 4th rank. It can be remarked that alternative A3 is always dominated. 1). (4). 1. or close to its edges. 1) ranking for (n − 1) = 3 is located at the origin as discussed for the case n = 2 in section 6. All 13 permutations of (1.2. 2. Kunsch though both alternatives are equivalent regarding their criteria ranks. 6. 6. 6. 3) with the fixed third position (4).L.3. 2. (4). This may be immediately verified from Fig. . (4). (4). This has been discussed in sub-section 6. 3). 1. 3. obtained by combining pair-wise the three criteria.3.1).3. (4).1 and Fig. 3) are present. 2. 1) are in the sequence (1. 3. (1. (4). For example the Pareto-optimal rank vectors located about the edge (1. are the P(3) = 12 rank vectors of Fig.164 P. The rank vectors within this solid angle. 2). 1).3 An upper octant with 16 vectors. 6.1. 2. The trivial (1. by looking at all permutations rankings of (1. 2). 4. 6.3. 3. A3 being fixed in the 4th position. 3) to (1. 2. When projected on a plane perpendicular to the (1. Nevertheless A4 has to share one time the first rank with A2 .2.

in the same way as basis vectors with integer components are used in the case of purely ordinal scales. 6. being less important for the decision-makers. but a procedure satisfying the property of invariance to rank reversals must be worked out. though it may .1 for designing the ordinal MCDM procedure. It would mean. It also shows that Paretooptimal rank vectors are not all located within the solid angle spanned by the basis vectors. as additional difficulties arise here. but some are located outside.2. see Fig.1 A MCDM procedure The nice thing about the Basic Property A. that unit n-vectors obtained from the cardinal evaluations of alternatives may be used as basis vectors.4. the set of evaluation vectors is infinite. 1. 6. but close to its edges. For cardinal scales things are thus far less clear than for ordinal scales. but they must not change the final overall rankings: the required invariance properties of the ranks of dominating alternatives would not be granted if unit evaluation vectors are used without care as basis vectors. corresponding to the trivial (1.1. given a set of preference weights.1 and Corollary A. the correct use of (7) is not fully clear. is that it is not only applicable to unit basis vectors corresponding to ordinal rankings but to any unit n-vector in the Kn dimensional subspace V{K} ⊂ V n . 1) solution in the projected 2-D case for n − 1 = 3. K n − 1.1. The cases in which all criteria have comparable importance will bring the solution (1. used in subsection 6. at least in principle.A Statistical Approach to Complex Multi-Criteria Decisions 165 This graphical analysis confirms that the introduction of a dummy dominated alternative will not cause any rank reversal among the dominating solutions.2. which reference vector should be used to calculate σ0 in (7)? A second difficulty arises because some alternative evaluations are uncertain.4 An Extension of the Approach to Cardinal Scales 6.3. But this is not so easy in practice. and any modification in some values affects the re-normalising process of unit vectors serving as a basis. While the number of discrete ordinal rank vectors is finite. some could be given arbitrarily bad values. concerning the relationship between the positive combination coefficients to be affected to the basis vectors. located in the octant diagonal. on which some combination coefficients are vanishing. and K criteria. given n alternatives. in which some criteria rank vectors have vanishing preference weights. First. 1). 4. the same happens when dominated alternatives are added to the set. 1. Because they come somewhat outside the edges of the solid angle. these rank vectors have one of several negative linear-combination coefficients with respect to some basis vectors: it means that they are each solutions of a reduced MCDM problem.

or ‘Least Preferred’. . . ‘INT’ computes the closest integer value. this is important considering that evaluations are many times imprecise or uncertain. e. Perform an affine transformation of cardinal evaluations to an m-pointscale (1. 5.166 P. ‘worst’ corresponds to the threshold value beyond which the score takes the worst value= m. it should be acceptable to modify this value. In addition it must be inspired by the wish of keeping things simple.L. some pertinent cost: the universe of discourse would be in the open interval: ] Most preferred. better alternatives.g. m). or right of the boundaries keep the value ‘Most Preferred’. meaning that values left. m] are obtained.4 for cost data in the universe of discourse ]20. Without loss of generality. Consider as an example a minimum criterion. 6. i = 1. so that for each criterion — one obtains an integer n-vector with five possible echelons in the interval [0.. Select the ‘universe of discourse’ for each criterion.g. The following approach [24] is proposed: Assume without loss of generality a minimum criterion. . 2. and close to the way the rather limited human brain is able to combine criteria evaluations by means of preference weights. They thus produce a 5-point scale for the scoring. they are in first approximation independent of the precise value of the alternatives. An example is shown in Fig. respectively. Least preferred [. In this way integers scores in [0. with flat ramps to the left and right of the universe-of-discourse boundaries. independently of possible uncertainties on its evaluation. so that the ‘best’ value is chosen close-by and fixed. Consider five echelons covering the universes of discourse uniformly by means of five triangular membership functions MFi . In this case the fuzzy rules are different from the usual Mamdani rules used in control theory (20) . Of course should the best value change drastically – due for example to technology development.. using simple ‘mental rules’. Kunsch not be unique. . n 5. m]: score (best) = 1 score (worst) = m value-best · (m − 1) score (value) = INT 1 + worst-best where: ‘best’ corresponds to the best value expected for the criterion. To address the mentioned difficulties the evaluations of alternatives with any criterion should be transformed into well-defined and discrete scores. the first-ranked alternatives in the criterion evaluation vector may be given a score= 1. Klir and Yuan [30]. and the coming up of new. see e. Let us elaborate on this approach by using the language of fuzzy logic. To achieve that result first consider that all criteria scales are cardinal. 120[. the adjustment for maximum criteria being trivial.

in the example of a 5-point scale to the least ranked solution. 5.5) in case there is a hesitation between two echelons (i). 5 (21) This rule system attributes the integer score SC(VALUE) = i in case there is no hesitation between two echelons. This value should however be fixed.A Statistical Approach to Complex Multi-Criteria Decisions 167 [31]. if a new better alternative shows up later on. five 0. Consider first the case when there are no purely ordinal criteria: a comparison has to be made between any alternative and the best available one.5 . It is better to choose a threshold value indicating a least preferred value for some criterion: all alternatives beyond this threshold will be given the same worst score. and an alternative with a given criterion VALUE. the choice of a value for score 1 may be revisited. or half-integer score SC(VALUE). or up to the next integer. .e. Of course. still worse alternatives.. being equally unattractive regarding this criterion. the scores are robust against changes in the evaluations of the alternatives due to uncertainties. 5. corresponding to the most preferred alternatives in the provided set.55 AND MGi (VALUE) 0. In this procedure. the number of points in the scoring scale. (i + 1). Thanks to the rounding down. 2. . . for each criterion. three parameters must be fixed: (a) The ‘most-preferred’ left boundary. . i. indicating that they are indifferent to the decision-makers. i = 1. This value may be fixed such that the first ranked alternatives indeed gets score 1. an application is considered between the membership grade (MGi ) of the MFi . . that may cause rank reversals [24]. modifications of evaluations of least-preferred alternatives. because it influences all other scores. or the introduction of new. creating an unwanted dependency between alternatives. such as not to be influenced by uncertainties. 2.55 THEN SC(VALUE) = i IF MGi+1 (VALUE) i = 1. (a) The choice of the most-preferred left boundary in the universe of discourse is not an issue. (c) The last important parameter is the number of echelons ‘m’ in the scoring scale. and (c) ‘m’ the number of preference echelons. which receives score 1. .e.45 THEN SC(VALUE) = i + 0. i. and the half-score SC(VALUE = (i+0. A simple way is choosing this value such as to give the highest ‘m’ score. assuming a constant range of values: an equal shift in the scores of all alternatives does not affect the final ranking. (b) The choice of the least preferred value in the universe of discourse is more delicate. as follows: IF MGi (VALUE) > 0. .. (b) The ‘least-preferred’ right boundary. . except for some hesitation in the middle range. and an integer.

168 P. If n > 5 decision-makers may not be able. 3) cannot be distinguished from (1. 2. 5). to establish a fine distinction between least-preferred alternatives. I therefore recommend the following formula for choosing an ‘m’ value: m=n m=5 m = 5. and not including possible hesitations between half scores. Medium Preferred (MediumP).L. alternatives I recommend to use 5 or 7-point scales. This is because again the human brain is probably not in a position to make refined distinctions between more than seven echelons indicating relative preferences. Rather Most Preferred (RatherMP). the score 1 being always present at least once. Kunsch echelons seem to give a sufficient precision to the judgments. When there are ordinal scales. Figure 6. Least Preferred (LeastP). or more. and the same score= m is given. see (3).4 The five Memberships functions for the minimum criterion ‘COST’ using a 5-point scoring scale with Most Preferred (MostP). The MCDM procedure . Of course the number ‘m’ of echelons must be the same for all criteria. the number of echelons must not be larger than n: this is because a ranking like (1. to be compared with P(5) = 540(+1) rank vectors available with ordinal scales. Five echelons in a 5-point scale give (5n − 4n ) possibilities for scoring n alternatives. 7 n n 5 (22) 7 n=6 When there are seven. Rather Least Preferred (RatherLP). 55 − 45 = 2101 integer scores are possible to evaluate (n = 5) alternatives. which is meaningless in an ordinal setting. or willing. in order to indicate indifference between ‘least preferred alternatives’. For n = 5. 3.

• Select weights in [0. choose the ‘m’ number of preference echelons as indicated in (22). The following procedure would then be adapted to cardinal data. 1. • Find for each weight instance the associated overall rank vectors to the combined evaluation vector by using an indifference threshold Q. (1. • Evaluate all alternatives on the m-point scale using the rule system described in (21).2. m].2 described in sub-section 6. 1.1. 1]K . as before in sub-section 6.1. 5. 5. leads again to (7). choose the ‘most preferred’ value to be given the score ‘1’. • Calculate K basis vectors obtained by centring and normalising the K criterion evaluation vectors.2. 5) has the largest standard deviation of all evaluation vectors.1 for identifying ties..2.2. integer or half-integers values on the m-point scales: call them criterion evaluation vectors. • Calculate the Rank Matrix in case of stochastic weights.g.2. Q = 0. Regarding the presence of ties in some basis vectors the same line of reasoning.1.1 for ordinal rank vectors is applicable to the scoring vectors obtained by means of the m-point scale.1 are valid for any set of inden pendent basic vectors in V{K} . This is easy to verify. 5.A Statistical Approach to Complex Multi-Criteria Decisions 169 developed in subsection 6. choose the ‘least preferred’ value serving as a threshold for the least attractive alternatives receiving the same score m. as in sub-section 6.2. for an even n = 6. 5) may take this part. The score vector with the largest standard deviation σ0 may be chosen as a reference for (7). from which the associated rank vector may be obtained. that the un-normalised basis vectors must be directly combined in order to obtain an overall score n-vector for the alternatives. as explained in sub-section 6.1 applicable to ordinal rankings. and m = 5. use each instance of the normalised weights to combine the independent basis vectors. Otherwise reduce the dimension of the basis. containing the score evaluations. 1. and mixed ordinal/cardinal data: MCDM Procedure with cardinal or mixed data: • For each criterion with cardinal evaluation data. and it may be used for the σ0 reference. 3. Verify that the K basis vectors are independent. Invariance Property of the cardinal/mixed-data MCDM procedure: When the MCDM .1 and Corollary A. For example with an odd n = 5 and m = 5. e. for transforming the preference weights into positive combination coefficients of the basis vectors. The same conclusion is obtained. because the Basic Property A. (1. and to obtain a combined evaluation vector with components in the interval [1. K n-vectors are obtained.

that it should be dominated in all criteria rank vectors. .3. It is the localisation of a composting plant in Ticino (Switzerland) developed in Maystre et al. these rankings are also shown in the comparison in Table 6. [32] for illustrating the ELECTRE methodologies.2. The same property applies when dominated alternatives are added to the set of alternatives. 20]. 19. the nature and units of criteria.2 An application of the cardinal procedure For comparing existing methodologies with the present one. 6. which was dominated by the pair in the overall rank vector. EIII) there are ascending (asc) and descending (desc) distillation processes leading each to two complete rankings.2.1. Note that ELECTRE and PROMETHEE have two rankings.2.4. The indifference and preference thresholds appear in the lower part of this table. 20] there are two complete rankings corresponding to the positive (dominance) flows (PROM+). while this is not necessarily true in case of ordinal ranks.4. II. as in the ordinal case. 19. The proof of this property is the same as for the ordinal MCDM procedure presented in subsection 6. .In PROMETHEE I [18. .1. The cardinal evaluation data.1. and the 5-point scale (5-pts) approach introduced in section 6.In ELECTRE II and III [14] (EII. and ranges for the criteria weights are given in Table 6. They are used for the evaluation with ELECTRE III [32]. because of the possibility of incomparability between pairs of alternatives. Note however that it is here sufficient that the third alternative be dominated in the overall rank vector: it is no longer required. In the last column on the right the correlations are indicated between the obtained rankings from the different methodologies. for which linear criteria with two P (preference). and the negative (dominating) flows (PROM−) respectively.170 P.3 provides the rankings resulting from the application of the different methodologies on the MCDM problem described in Table 6.2. This is because the scores of the dominating alternatives in the pair are unaffected whatever happens to the third alternatives in any criterion. Table 6. see sub-section 6. Q (indifference) thresholds are assumed [18. or mixed data for obtaining overall rank vectors. the ranks of alternatives in a pair are not affected when a third alternative gets worse. and also for the comparison with PROMETHEE I. Kunsch procedure is applied to cardinal. a MCDM application is presented in this sub-section.L. Of course this invariance is only granted if the worsening of dominated alternatives does not modify the m-point-scale range of values. There are n = 7 alternatives and K = 5 linearly independent criteria.

2 The evaluation data for the MCDM localisation problem32 . Criteria Alternati-ves A1 A2 A3 A4 A5 A6 A7 Weights minaveragemax Indifference thresholds Preference thresholds Site Price SF/m2 -120 -150 -100 -60 -30 -80 -45 20-25-30 Transport 103 km/y -284 -269 -413 -596 -1321 -734 -982 35-45-55 Environmental #Residents status affected Qualitative Qualitative 5 2 4 6 8 5 7 8-10-12 3. The comparison in Table 6.5 4.3.5 8 7.5 10-12-14 Availability of site Qualitative 18 24 17 20 16 21 13 6-8-10 15 40 80 350 1 3 0.A Statistical Approach to Complex Multi-Criteria Decisions 171 Table 6.5 3. For the sake of keeping assumptions to a minimum. which are equal to the difference between the positive and the negative flows.e. which .19. though there are also not entirely immune against rank reversals with preference thresholds P = 0. In general it can be shown that the complete ranking induced by negative flows in PROMETHEE I provides more robust results than the net flows used for the complete ranking in PROMETHEE II. i.1.5 5. Of course this assumption deserves to be further analysed by means of sensitivity studies depending on actual preferences of the decision-maker groups. which are very different from those obtained by other methodologies.3 is very instructive as far as it confirms the following results: 2. ELECTRE II and III provide results.2. are also indicated for comparison in Table 6. assuming that the preference thresholds P = 0 for all criteria. the range of the m-point-scale for each criterion is chosen to be the interval between the best value (score ‘1’) and the worst value (score ‘m’). .5 1 5 .The PROMETHEE II [18. see section 6. Minimum criteria appear with a negative sign (SF=Swiss Francs). except for the ascending distillation in ELECTRE III [14].20] complete ranking (PROM II) is obtained from the net flows.5 4 8.The results of ordinal MCDM procedure. An interesting result [24] is that they are identical to the complete ranking provided by the negative flows in the ordinal version of PROMETHEE I..

A1 EII asc EII desc EIII asc EIII desc PROM+ PROMPROMII Ordinal 5-pts 7-pts 4 4 2.1).5 5 4.5 5.L. Nevertheless.e. the ascending distillation could be compared to the PROMETHEE I negative flow ranking [24].5 7 7 6 6. indicated as average weights in Table 6.5 3. A4 > A3 > A1 > A2 > A7 > A5 = A6 (23) .5 A3 4 4 4 2. in this example.5 5 4. the ordinal MCDM procedure in sec. the evaluations of the criteria are sufficiently well apart: the interdependency between alternatives introduced by preference thresholds remains rather small [24]. PROMETHEE I.973 0. This indicates that.873 0. discussed below.982 0.5 3.5 5. For the same reason the very high correlation of PROMETHEE II with the 5-point-scale results should be considered with caution. when the weights are taken to be stochastic in the ranges [min.2.000 0.580 0.580 0.5 3 2. and 7-point-scale results.5 2 3 A4 2 2 1 1 1 1 1 1 1 1 A5 7 7 6 5. i. the 5-point (5 pts) and 7-point (7 pts) scales in the MCDM procedure in sec.633 0. etc.5 5. max] indicated in Table 6.663 0.800 1. The abbreviations are further explained in the text. 2. (6. (6. The negative flows in PROMETHEE I provide a ranking. the addition of new dominated alternatives.5 2 2.5 7 5.172 P.5 5. as it gives no guarantee for robust results in case technical parameters or evaluation data would change.5 4. These results show that the crisp ranking is relatively stable in this range. which is closely related to both 5-point-scale. Kunsch Table 6.5 7 A6 4 4 7 5.2.5 6 A7 6 6 5 2.5 3 3 2 A2 1 1 2.3 A comparison between the rankings of ELECTRE II.5 Correlation 5-pt scale 0. may introduce unwanted rank reversals to the PROMETHEE I results.5 6 7 6.945 comes acceptably close to other results.2. Maystre et al.4 provides the Rank Matrix for the 5-point scale. III. [32] assume crisp weights.1). It is important in the present approach to make statistical analyses on these weights as well. PROMETHEE II. Table 6..5 2 4 4 2 4 4.4. changes in threshold values.5 5.

Kendall-τ [6]. not explainable rank-reversals.0 % 0.1 % 2.0 % 8.4 The Rank Matrix obtained from stochastic weights varying according to a rectangular distribution in the range indicated in Table 6.1. These procedures may be used either as interpretative instruments of statistical data in multi-criteria enquiries.6 % 0.8 % 69.1 % 0.7 % 0.4 % 16.3 % Rank 3 74. though they have a strong mathematical underpinning detailed in appendix A. agent-based modelling.e. Regarding the properties of different correlation measures. the Spearman’s rank correlation [6] gives the same correlation values as Pearson’s correlation in the absence of ties.8 % 0. not only quite technical ones in Marketing Research. [28] have shown that the cosine is the most adequate e a measure of distance between rank vectors when trying to achieve a collective consensus for group decision-making.0 % 0. as shown in [24].0 % 0.0 % 0.0 % 0. but it is not applicable to tie situations.0 % Rank 2 15. and close to ‘mind computing’ in many every day’s MCDM problems.0 % 0. has not the satisfactory property of Pearson’s correlation providing the cosine as a measure of distance between unit vectors.0 % 6.6 % Rank 5 0. Many applications are certainly thinkable. quite common in popular outranking methods.0 % 84. Rank 1 A1 A2 A3 A4 A5 A6 A7 0. Garcia and P´ rez-Rom´ n [33] in Ruan et al. though it has the capability of handling ties.0 % 0.0 % 0.0 % 0.0 % 42.5 Conclusions and Outlook This chapter presents simple procedures for solving MCDM problems. i.8 % 1. It is thus not useable in the MCDM context. Springael [34] has investigated .2 % Rank 6 0. but also in many cases of ordinary life.0 % 0.. The simplicity of the approach enhances its credibility.0 % 0.A Statistical Approach to Complex Multi-Criteria Decisions 173 Table 6. decision aiding tools for MCDA. In particular this approach eliminates dependencies between alternatives.2 % 68.0 % 21.2. however that explainable rank reversals are always possible with ordinal rankings.5 % 2.0 % 0.0 % 7.7 % 21.0 % 49.0 % 0. by making it more realistic.2 % Rank 7 0..1 % 0. Note.0 % 0. The key statistical concept is a measure of distance by means of the ordinary (Pearson) correlation between two rank vectors. because it does not make possible the development in K-criteria basis vectors of any unit rank vector. or non-transitivity situations.8 % 0.2 % 50. which are common in MCDM.8 % Rank 4 8.8 % 47.0 % 0.0 % 0. or directly as normative.0 % 0. due to the poor information available on the dynamic positioning of alternatives.0 % 0.0 % 0.0 % 7. etc.0 % 100. like proposed in this chapter.

1. its properties substantially differ from these of the Pearson’s correlation explicated in appendix A. which are beyond any acceptable range for some criteria. instead of one. In most applications the only information available on the many preferences is ordinal. One result obtained numerically for n 7 is that the Kendall’s correlation matrix of the P(n) rank vectors has rank 2(n − 1). 1]K intervals.1. it may not be elected for the first rank in the overall ranking. 14] like the incomparability between actions. showing the ranks probabilities of the different alternatives. but not always well-understood techniques.Eliminate beforehand alternatives. In each such group a Rank Matrix is obtained by Monte-Carlo simulation. Also an analysis from a group-theoretic point of view of the P(n) permutations with ties of n alternatives is of great interest for characterising the Pareto-optimal .Introduce IF rules for avoiding undue compensations between ranks. the second degenerate eigenvalue has a much smaller value. the set of Pareto-optimal rankings may be fully. like the following: IF alternative I has rank 5 for one criterion at least. A formal proof of this property for all n values is still outstanding. . are partially. The first non-vanishing eigenvalue is close to the expected value P(n)/(n − 1). generally in the form of some criteria ranking. Such treatment avoids the purely mechanical processing by elaborated. and make the results much more credible than crisp values of the weights. and in aggregating them into an overall ranking with ties.1. compensatory or non-compensatory effects. The present chapter insists on the need to consider stochastic analyses of the preferences as manifested by groups of decision-makers. veto thresholds. But perhaps in many cases prior pre-conditioning of data may bring an adequate substitute for these concepts. An aspect to be discussed is the choice of indifference thresholds in both situations of establishing criteria rankings. 3. Some possibilities to be further explored in practice may be the following ones in addressing several incomparability and compensation issues: . see Property A. Of course some important concepts introduced in outranking methods [1.L. 5. see Property A. shared by more or less homogeneous decision-maker groups with similar priorities. and do not make superfluous detailed sensitivity analyses.174 P. 2. Kunsch more in detail the Kendall-τ as a measure of distance between unit vectors. etc. or totally lost in the approach proposed in this chapter. and two sets of (n − 1) degenerate eigenvalues.4.4. which are in general impossible to justify. etc. More complicated constraints on the weight sets may be handled as well. K being the number of independent criteria. In the case of ordinal data. and univoquely determined by means of fully stochastic weights in [0.

How to take into account hesitations between two adjacent preference echelons? A half-score approach has been proposed in the procedure of sub-section 6. should more complicated cases be considered. In both cases of only cardinal data. choosing the least preferred threshold corresponding to the ‘m’ score depends on variable decision-maker preferences. . mainly regarding the MCDM procedure for cardinal and mixed data presented in sub-section 6. or of a mix of ordinal and cardinal data. are the following ones: . The approach proposed in this chapter is certainly not a final answer to all MCDM aspects. m] range of cardinal values corresponding to the scores? While the choice of an origin for the best score 1 is not really an issue. If the m-pointscale MCDM procedure is adopted.4.How to choose an indifference threshold for finding the associated overall rank vectors to combined evaluation vectors? The same issue appears in both ordinal and cardinal cases. as more information is needed from the side of the decisionmakers to achieve the statistical analysis of the Pareto-optimal rankings.1: is it useful to have a finer score graining? How does the graining influence the final aggregation of the score vectors for individual criteria? . and will always be ill-defined problems. This stresses the fact that MCDM problems are.How many preference echelons to choose? I have assumed that the human mind is not really in a position of evaluating more than five to seven echelons on a scale. the human freedom in defining preferences makes it impossible to design completely mechanical MCDM algorithms.1.1. Key problems on which further research could elaborate.How to choose the [1. or other non-linear functions.4. even if the assumption is made of fully stochastic weights for obtaining the Rank Matrix.May different scales be used when both ordinal and cardinal data are present? .How to convert cardinal data into scores? Is the proposed linear transformation sufficient. .A Statistical Approach to Complex Multi-Criteria Decisions 175 set given arbitrary preference weights for the criteria. but in the cardinal case it only arises at the level of aggregating criteria through .4. but this is certainly open for discussion. like logistic curves. it is why additional sensitivity analyses may have to be performed to capture this uncertain dimension of the scoring process of alternatives. at the risk of being less close to the ‘mind computing’ way of ordinary people? . there is more space for discussions regarding some important choices to be made than in the ordinal case. Moreover. as introduced in sub-section 6. the situation is not so straightforward.

and demonstrated.176 P.1) (A.1. identification of incomparability. 13] insist on several objectives of outranking methods. Kunsch the weighed sum of criteria scores. Another point for future research is elaborating the relationship between the number of alternatives. or not only interested in complete ranking of alternatives. Call V n ⊂ Rn the set of all real centred and normalised n-component vectors.2. see a discussion on this aspect for the ordinal case in sub-section 6. Therefore statistical aspects are in the foreground. 2.3) ∑ ∑ vi = 0 v2 = 1 i i=1. vn ) ∈ V n i=1.. etc. v2 . is superfluous in all circumstances. Definition A. in short unit n-vectors. 2. the latter are not primarily.. To complete this discussion. It may be expected that more freedom is given in case of cardinal data.. Other aspects of pair-wise comparisons in outranking methods are classification.n (A.. are detailed.2) (A. Appendix A A. Therefore the impression should not be awakened that all sophistication introduced in current MCDA methodologies.2. and particularly outranking methods.L.2. and Roy at many places [1. .. A further field of investigation is to assess the possible theoretical input of the proposed vector-optimisation methodology to these MCDA aspects. and the use of m-point scales for cardinal data. . but this needs to be verified and quantified.. .1. in short unit n-vectors v: v = (v1 . . 3. This theory serves as a basis for the MCDM procedures developed in the main text of this chapter.2. but not at the level of defining individual criteria rank vectors with possible ties in the ordinal case.n . description. Figueira and Roy [26].. MCDA is more concerned over decision-aid for individual decision-makers [1.. and the maximum number of admissible independent criteria in the basis set. note again that the present chapter primarily discusses the ranking of alternatives of MCDM problems in the presence of many decision-makers showing various preferences profiles. 5].1 Properties of the space of n-vectors In this theoretical appendix the mathematical properties of the space of normalised and centred real vectors with n components.

.1.1. Corollary A. Call ρv (b) the column vector containing the correlation coefficients of v with the K basis vectors. Its rows are linearly dependent. Thus the corresponding row and column in the correlation matrix are linear combinations of the K first rows or columns respectively. If there would be any linear relationship between the rows and columns of this correlation sub-matrix.1. Any matrix. The rank of CK is at most equal to K. Proof. which is the dimension of the space of all unit n-vectors. which constitute a basis by definition. This corollary is trivial.1.1. because it is not possible to find more than K independent n vectors in the sub-set V{K} . . Definition A. On the contrary any additional vector ∈ V{K} in the matrix is a linear combination of the K basis vectors. Consider that the top left K sub-matrix of CK (m) is the correlation matrix of the K basis vectors. Consider a set {K} of K n − 1 independent unit vectors in V n . i. which is generated by all possible linear combinations of these n vectors serving as basis vectors. Call Cb the correlation matrix of K n n − 1 basis vectors in V{K} .2.3. the same linear relationship would exist between the basis vectors. Thus v is linearly dependent on the (n − 1) independent vectors. that any n matrix consisting of column vectors from V{K} has at most rank K. K Definition A.1. this is excluded because the basis vectors are independent. The second part of the proposition is obvious because by definition any n × n matrix consisting of (n − 1) independent column vectors plus any other column vector ν ∈ V n has a rank equal to (n−1).2).2.A Statistical Approach to Complex Multi-Criteria Decisions 177 Basic Property A. the latter includes all possible linear combinations of the set {K} of independent basis vectors. according to (A.e. The rank of any such a matrix can obviously not be larger than n. Property A. Proof. The rank of this subn matrix is thus equal to K. Call n V{K} the subset of V n . as their sum is equal to zero. Therefore the rank of such matrix is at most equal to (n − 1). Every vector in V n may be obtained as a linear combination of a basis made of (n − 1) independent vectors in V n .2. Call CK K n − 1. has a rank at most equal to (n − 1). Proof. The sub-set V{K} ⊂ V n has dimension K. (m) (m) the Pearson’s correlation matrix of m n K vectors in V{K} . with column vectors from V n . in short basis vectors.

(4) All rows ri and columns c j of the correlation matrix C(n) are eigenvectors of the q non-vanishing eigenvalues λ = P(n)/q. (2) The m × m correlation matrix C(n) = XX and the n × n matrix D(n) = X X have the same rank q = (n − 1). see (3) in the main text.2. (3) All unit n-vectors in V n are eigenvectors of D(n) with the same eigenvalue λ = P(n)/q.2.4) immediately results when using the relation v = Bcv (b) ρv (b) for calculating the correlation coefficients in matrix of v with the K basis vectors. n Property A.4. Consider the matrix X (m = P(n). The following properties refers to the matrix built with the P(n) rank vectors of n alternatives. . The following relationships hold.3: ρv (b) K = Cb cv (b) (A. (A.6) .178 (b) P. 1). Its rank can be at most equal to n. according to Property A. . which is the q-denerate eigenvalue of C(n) .1.1.L. 1. (1) is obvious from the definition of the matrix X. .4) (A. Kunsch Further call cv the column vector containing the coefficients of the linear development of v in the basis vectors. . having the rank = K. (A.5) results from (A.5) (b) cv = K −1 (b) Cb ρv Proof.4) by considering that the correlation matrix of the K basis vectors is invertible. . . Property A. given Definitions A. and q non-vanishing eigenvalues. (2) It is obvious that the correlation matrix is C(n) = XX by definition of the data matrix X.1. not including the trivial rank vector (1. Consider any vector v ∈ V{K} and the basis matrix B = (b1 . the rows of which are the P(n) normalised and centred rank vectors (unit rank vectors) in V n : (1) The rank of the matrix X is q = n − 1. D(n) .3 in the main text. but each row sums up to zero. b2 . . . bK ) where bi are the basis column vectors. see sub-section 6. n).3.1. so that its rank is equal to (n − 1). Proof. Denote: D(n) uα = λα uα C(n) vα = μα vα (A. The next property gives us the means for calculating the coefficients of the linear combinan tion of basis vectors for any vector v ∈ V{K} .

this vector is an eigenvector of the correlation matrix. and it is at the same time one of its row. Consider any unit vector v ∈ V n . According to (A. and D(n) regarding eigenvalues.11) vα has for components the correlation coefficient of itself with the P(n) rank vectors. say o. D(n) . Note that this is a well-known result of factor analysis.6). There is thus only one nonP(n) = d −o q P(n) P(n) P(n) − =− o= n n−1 n(n − 1) λ = (A. Using the same approach the same result is obtained for other non-vanishing q eigenvalues.7): vα = Xuα (A. eigenvalue λ of C(n) . say d. as μ1 is the largest eigenvalue by definition.11) Take for the eigenvector uα any of the P(n) unit rank vector.9) (A. Because D(n) has the same eigenvalues as C(n) . it is also of rank q = n − 1. one obtains: XD(n) u1 = XX Xu1 = C(n) Xu1 = λ1 Xu1 (A. Multiplying now the second equation on the left by X . Because all alternatives are equivalent regarding all possible positions.A Statistical Approach to Complex Multi-Criteria Decisions 179 where λ1 ··· λ2 ··· λq are the q non-vanishing eigenvalues of D(n) . which it is thus equal to the trace of D(n) according to (2). therefore d = P(n)/n. or column. It is immediately seen that: D(n) v = (d − o)v vanishing q-denerate. and n(n − 1) equal non diagonal element.7) shows that Xu1 is an eigenvector. such that: (A.7) (A.e. equal to P(n). The sum of eigenvalues is the trace of C(n) . .8) indicates that any unit n-vector is an eigenvector. and μ1 μ2 μq are the non-vanishing eigenvalues of C(n) . see H¨ rdle and Hl´ vka ?. and λ1 is an eigenvalue of C(n) . a a (3) D(n) is by definition the variance-covariance matrix of the n alternative positions in the rankings with ties. q = (n − 1). and thus λ1 μ1 . this matrix has n equal diagonal elements. and multiplying the first equation on the left by X.. Considering the first eigenvalue α = 1 in (A. one also sees that X v1 is an eigenvector of D(n) with respect to the eigenvalue μ1 and thus λ1 μ1 . According to (A. i. Therefore λ1 = μ1 .8) (A.10) (4) results from the equivalence of C(n) .

L. as C(n) has q = (n − 1) degenerate non-vanishing eigenvalues. (A.1.2 Projection in the XY -plane of the 74 ranking vectors (n = 4) on the unit sphere in R3 Figure A. .4. 3.5. . A. The ending point of the unit rank vector comes approximately at the position of the first label component.12) Proof. and all its rows and columns are eigenvectors.1 The unit sphere with the p(4) = 74 non-trivial ranking vectors. Kunsch Property A. 2. . The formula results immediately for k = 2 from Property A. T he correlation matrix C(n) is such that C(n) k = P(n) n−1 k−1 C(n) = λ k−1C(n) ∀ k = 1. .1. The formula for any k power results by successive application of this formula. as seen from above.180 P.

49–73 (1991). De Smet. 31. Multicriteria Methodology for Decision Aiding. Despontin “A socio-economic model of moving decisions in cities”. Ph. J. Y. 189–202 (2001).. 81–90 (2000). 1992).L.2 The unit sphere with the p(4) = 74 non-trivial ranking vectors. (AddisonWesley. M. Bibliography [1] [2] [3] [4] [5] [6] [7] B. The ending point of the unit rank vector comes approximately at the position of the first label component. 40 (1-2). P. State of the Art Surveys. De Smet. (Economica. Inc. Brans (2000) Traffic crowding in cities . Martens and M.L. N. 1999). VUB/MOSI working document. 17 (VUB. [8] [9] [10] [11] [12] [13] . London.Control with urban tolls and flexible working hours A group multicriteria aid and system dynamics approach. (Kluwer. Springael . Vincke. Corrigendum in 41 (3-4). 2007). JORBEL. Figueira. MA. Multiple Criteria Decision Analysis: An Integrated Approach. Y. P. Towards statistical multicriteria decision modelling: a model and its application. S. V. (John Wiley&Sons. Roy.Computational Intelligence in Complex Decision Systems 181 Figure A. J. Springael. Belton and T.. Multiple Criteria Decision Analysis. 2002). M´ thodologie multicrit` re d’aide a la decision. Inc. A new framework for describing the decision behaviour of large groups. Ehrgott Eds. Ferber.L.P. 10. Kunsch. Kunsch. Marketing Research: an Applied Orientation. (Prentice Hall. J. 5th edition. Roy. 305–313 (2002). 1996). as seen from below. Springael and P. 2005). Kunsch and J. Dordrecht. J. Brans. J. A multicriteria-based system dynamics modelling of traffic congestion caused by urban commuters. Kunsch and J.L. Upper Saddle River. 2005). Theory and Decision. NY.-P. Greco. 81–97 (2002). Central European Journal of Operational Research. e e ` B. Kunsch. Research Paper. B. Boston. Multiagents Systems: An Introduction to Distributed Artificial Intelligence. Springael and P. Stewart. P. Roy. (Kluwer. The outranking approach and the foundations of the ELECTRE Methods. Antwerpen. E. Multi-criteria Decision Aid. Brussels. Journal of Multicriteria Decision Aid. Chichester. 2004). Paris 1985).L. J. 2004-011 (Universiteit Antwerpen. 11/6. NJ. Malhotra. (Springer Science+Business Media.

J. [21] T. Triantaphyllou. D’hondt. 1994). 31(6). in Ref. [32] L. [24] P. Awang and E. European Journal of Operational Research. 2008). 12. Figueira. “Ranking irregularities when evaluating alternatives by using some ELECTRE methods”. [26] J.A. Maystre.182 P. Berlin. D. 2007). M. in Ref. Pictet. H. Figueira and B. [30] G. 629. in J. Strengthening PROMETHEE with ideas of AHP. Kunsch. [23] C. Mareschal. submitted for publication (2010). 36. Springael. Description. J. S. Springael. 1996). P. H¨ rdle and Z. PROMETHEE and AHP: The design of operational synergies in multicriteria analysis. Brans and Ph. J. France. PROMETHEE methods. [16] J. Klir and B. Simos. De Keyser and P. E. De Brucker and A.J. 153. 163–195. Vincke and B. Yuan Fuzzy sets and Fuzzy logic – Theory and Applications (Prentice-Hall. [31] D.P. Kunsch [14] L. Ehrgott Eds. 307–317 (2004). [22] W. Brans and B. European Journal of Operational Research. 647–656 (1985). GAIA and BANKADVISER Software.. [34] J. The PROMCALC. Simos. Macharis. New York. a a [28] D. [18] J.L. [33] J. (World Scientific. 89. Reinfrank An Introduction to Fuzzy Control. conseils prae e tiques et cas d’application a la gestion environnementale. private communication (2008). 24. Bana Costa Ed. Hl´ vka. Maystre. 2005) p. Montero. Ranking irregularities when evaluating alternatives by using some ELECTRE methods. Omega. 37. [20] J. 216–242. C. e e [19] J.L.. in. Remedies to rank reversal in multi-criteria decision methodologies. New Jersey.P. Ruan. 1995). J. J. Mareschal PROMETHEE-GAIA.L. Some consensus measures and their applications in group e a decision-making. 28. Proceedings of the 8th International FLINS Conference. 611. 45–63 (2008).P. (Springer. by Wang and Triantaphyllou. Decision Support Systems. Management Science. Garcia. 28.P. Peeters.L. Chap. 2002). Exemple didactique commun: localisation d’une usine de compostage au Tessin. (Springer. K. M. Driankov.Y. Greco. A statistical Multicriteria Decision Aiding technique. J. Berlin. Brans and B. 2nd ed. Ph. in Ref. European Journal of Operational Research. Kerre Eds. M´ thodes multicrit` res ELECTRE.E. Computational Intelligence in Decision and Control.. Readings in Multiple Criteria Decision Aid. Brans and B. Mareschal The PROMCALC and GAIA decision support system for MCDA. Vincke A preference ranking organisation method: The PROMETHEE method for MCDM. P´ rez-Rom´ n. Pictet. Kunsch. 14.P. 457–461 (1996). Roy A note on the paper. 1990) p. The Analytic Hierarchy Process (McCraw-Hill Inc. [15] J. The PROMETHEE Methods for MCDM.P. Mareschal. Verbeke. J. (Ellipses. How to select and to rank projects: The PROMETHEE method.Y. Multiple Criteria Decision Analysis: state of the art surveys (Springer. (Presses Polytechniques et ` Universitaires Romandes. Saaty.. Multivariate Statistics. [17] J. Brans. Brans and B. Lu. 228–238 (1986). 731–733 (2009). Mareschal. [25] X. J. Comparison of the properties of different rank correlations formulas used as a distance measure. p. New York. Berlin. 297–310 (2004). . (Springer. [27] W. Une m´ thodologie d’aide a la d´ cision en e ` e pr´ sence de crit` res multiples. 3. Martinez. [29] P. London. 1980). p. L. Lausanne. Omega. A note on the use of PROMETHEE multicriteria methods.L. Hellendoorn.

Combined with thoughtfully designed interfaces. distributed assessments revealing variations in performances.Chapter 7 A Web Based Assessment Tool via the Evidential Reasoning Approach Dong-Ling Xu Manchester Business School. UK Ling. discussion forum. opinion survey and assessment have become increasingly common. By applying the ER algorithm for information aggregation. DOI 10. Manchester. This chapter describes the interface design of a web based assessment tool based on the Evidential Reasoning (ER) approach [29. The tool has been validated and used by a number of business users and examples of the application areas include business innovation assessment. sensitivity of the outcomes to uncertainties in different parameters. 31]. supplier evaluation. web based voting. Some web sites also allow users to construct their own decision models through the Internet [10]. the tool offers a flexible and friendly environment for web users to express their assessment judgments consistently.1 Introduction With the advance of computer and Internet technologies. The interfaces of the tool allow users to enter data using a built-in assessment criteria hierarchy and to display aggregated outcomes in both text and graphic formats. The ER approach is developed to handle multi-criteria decision analysis (MCDA) and assessment problems possibly with hybrid types of uncertainties. Atlantis Computational Intelligence Systems 2. objectively and accurately. 7. and performance assessment. and user specific reports highlighting key areas for attention.uk This chapter describes the interface design and the applications of a web based assessment tool on the basis of the Evidential Reasoning (ER) approach. The ER approach is a generic method that can be used for modelling and investigating MCDA problems under various types of unD.Xu@mbs. assessment outcomes generated from the tool can include average scores and ranking. Ruan. The University of Manchester.ac. © 2010 Atlantis Press/World Scientific 183 . M15 6PB.1007/978-94-91216-29-9_7. It uses belief decision matrices for problem modelling and data collection. providing new ways for supporting decision making processes. Computational Intelligence in Complex Decision Systems.

184

D.-L. Xu

certainties [31]. It offers unique features and flexibilities for supporting users to conduct comprehensive assessments and make informative decisions in a way that is consistent and reliable. Many assessment schemes, such as impact assessment of policies, customer satisfaction survey and business performance assessment, are structured as typical multi-criteria decision making problems [30, 1]. Typical assessment schemes include the EFQM (European Foundation for Quality Management) model [7] for business excellence assessment and many pre-qualification questionnaires (PQQ) used to select suppliers by different organisations including the public sector in the UK. These schemes normally consist of a criteria hierarchy, criteria weights, the detailed explanation of each criterion, and a scoring standard for assessing each criterion. The construction of an assessment scheme requires domain specific knowledge and is normally completed by experts in the assessed areas possibly with the help of decision analysts. The web based assessment tool discussed in this paper is developed by the requests of a number of industrial collaborators and business advisors. It has been applied to implement a number of assessment schemes used in supplier selection, business innovation assessment and policy impact assessment. The interfaces of the tool are supported by a relational database which is specifically designed to manage the assessment schemes and related knowledge. Assessment information entered by users through the interfaces is also stored in the database. To aggregate information on sub criteria in order to obtain overall performance assessments of an entity such as a business or a policy, the ER algorithm [31] is built into the tool. The tool has an automatic report generator and interfaces for displaying aggregated outcomes in both text and graphical formats. In the ER approach, belief decision matrices are used to model multiple criteria assessment problems and record assessment information. Using belief decision matrices, users are able to select multiple answers and associate a belief degree with each answer. This is useful in situations where an assessment is felt to be somewhere between two adjacent grades, or where an assessment is provided by a group of people with different opinions. Belief decision matrices can also accommodate missing or unknown information in an assessment. This flexibility is one of the unique features of the tool. The tool can be used in a wide range of assessment activities such as customer satisfaction surveys, product assessment and selection, and identification and ranking of consumer preferences. It can be used in both individual and group assessment situations. Most importantly, it can handle hybrid types of uncertainties including missing data, randomness and

A Web Based Assessment Tool via the Evidential Reasoning Approach

185

subjective judgments rationally and consistently. The chapter is organised as follows. First the ER approach and its features are briefly described in Section 7.2. An illustrative example is given in Section 7.3 to demonstrate how the ER approach can be applied to model assessment problems and aggregate information with uncertainties. In Section 7.4, the interfaces and the specific considerations about how to accommodate the special features of the ER approach in the interfaces are illustrated using examples and screen snapshots from applications. Using the examples, the aggregated results and features are explained. Concluding remarks are made in the Section 7.5.

7.2 The ER Approach The ER approach is developed to deal with MCDA problems having both quantitative and qualitative information with uncertainties [29, 30, 31]. Using the ER approach for MCDA includes two major steps. The first step is to use a distributed modelling framework in the format of a belief decision matrix to model a MCDA problem. The second step is to aggregate the assessment information of alternatives using the ER algorithm. The belief decision matrix is an extension to a traditional decision matrix [22, 9, 11] for modelling a MCDA problem. In a decision matrix, each element is a single value, representing the “accurate” or average assessment of an alternative or option on a criterion. In the belief decision matrix, an assessment is represented by a paired vector, rather than a single value. A paired vector, referred to as a belief structure or a distributed assessment in the ER approach, consists of the permissible referential values of a criterion and their associated degrees of belief. For example, if people are asked to rate the impact if UK adopts the Euro, the answers could be positive and negative to certain degrees. In a belief decision matrix, such distributed judgements are allowed and various types of uncertainties are captured and explicitly modelled. The criterion referential values can be quantitative scores, qualitative assessment grades in different terms and numbers, or a mixture of both. Assessments made on different scales and terms can be converted to a common scale in terms of utility equivalence using the rule and utility based information transformation techniques [29]. The ER algorithm is developed on the basis of Dempster-Shafer theory of evidence [17, 4, 31]. The theory is related to the Bayesian probability theory in the sense that they both deal with subjective beliefs. However, according to Shafer [17], the evidence theory includes the Bayesian probability theory as a special case. Its subjective beliefs are also required to obey the probability rules. The biggest difference being in that the former is able to deal with

186

D.-L. Xu

ignorance, while the latter is not. The evidence theory has found wide applications in many areas such as expert systems [4, 2, 27], pattern recognition [8, 12], information fusion [18], database and knowledge discovery [5], multiple attribute decision analysis [3, 31], and audit risk assessment [21, 20]. The ER algorithm is used for aggregating distributed assessment information recorded in a belief decision matrix. The outcome of the aggregation is also a distribution, rather than a single score. If necessary, a score can be calculated from a distribution by first quantifying the assessment grades and then adding the grade values weighted by their associated belief degrees in the distribution [29, 31, 32]. Note that a score generated using the ER algorithm may be different from that generated using the weighted sum method because a distribution is generated through a nonlinear aggregation process. There are two main reasons for not using the weighted sum method for information aggregation in the ER approach. Firstly, unlike the ER algorithm, the weighted sum method requires the stringent additive preferential independence condition to be met by assessment criteria [13], which is difficult to satisfy and hard to check if many criteria need to be taken into account. Secondly, the weighted sum method can only provide an average score for assessing an alternative with little capability to handle uncertainty, whilst the ER algorithm can generate a distribution and, when there is missing information, a lower and an upper bound of a belief degree to which an assessment statement is believed to be true. Such pragmatic yet rigorous outcomes provide scopes for global sensitivity analysis of alternative rankings under uncertainties [14, 15]. The above features of the ER approach bring the following three practical benefits. Firstly, assessment questions and choices of answers can be presented to users in a user friendly format. Secondly, the distributed assessment information generated by the ER algorithm reveals performance diversities and supports the identification of strengths and weaknesses. Thirdly, the number of criteria in the assessment model is much less of a concern to the ER approach than to other conventional approaches, because the former does not require the satisfaction of the additive preferential independence condition. The ER approach including the ER aggregation algorithm and some of its features is illustrated using a simple example in Section 7.3. 7.3 Illustration of the ER Approach The following example is based on an excerpt from a real world application of the ER approach.

A Web Based Assessment Tool via the Evidential Reasoning Approach

187

To assess and help to improve the innovation capability of small to medium sized enterprises in the UK and Europe, assessment criteria and grading standards were developed by Xu et al. [28]. One of the criteria is Company Culture which is assessed through several sub-criteria including the following two: (1) How well does the company motivate talented people? (2) Are employees empowered to test new ideas? The directors, managers and some representatives of the employees of a company were asked to self-assess its performance on the sub-criteria using the three grades {Not Very Well, Fairly Well, Well} for sub-criterion (1) and the other three grades {No, Some Times, Yes} for sub-criterion 2. Not every one agrees with each others assessments. On the first sub-criterion, 30 % of the answers are Not Very Well, 50 % Fairly Well, and 20 % Well. On the second one, the answers are 40 % No, 30 % Some Times, and 20 % Yes. It should be noted that the total percentage for the second one did not add up to 100 % because some of the employees did not provide an answer to the assessment question. Instead of approximating the company performances on each sub-criterion using a single grade, which is normally the case if a decision matrix is used, the ER approach representing the two assessments using the following two belief structures: P(S1 ) = {(Not Very Well, 0.3), (Fairly Well, 0.5), (Well, 0.2)} P(S2 ) = {(No, 0.4), (Some Times, 0.3), (No, 0.2)} (1) (2)

The numbers associated with the assessment grades are called belief degrees. Because belief degrees are in essence probabilities, the sum of them in each assessment, P(S1 ) or P(S2 ), should be 1 or less. If the sum is less than 1, this means that there is some missing information. When the sum of the belief degrees in a belief structure is 1, it is said that the assessment is complete, otherwise incomplete. Accordingly, the assessment P(S1 ) in (1) is complete while P(S2 ) is incomplete. Suppose the utilities [13] of the above two sets of assessment grades are evenly distributed in the utility space. Based on the assumption, if the three grades are denoted by {H1 , H2 , H3 } and their utilities by u(H1 ), u(H2 ) and u(H3 ) respectively, then we have {H1 , H2 , H3 } = { Not Very Well, Fairly Well, Well } = { No, Some Times, Yes } and u(H1 ) = 0, u(H2 ) = 0.5, and u(H3 ) = 1 (4) (3)

188

D.-L. Xu

If the utilities of the two sets of grades are distributed differently, then both sets of grades need to be converted to a common set using utility equivalent transformation techniques developed by Yang [29]. Using the notation in (3), (1) and (2) become P(S1 ) = {(H1 , 0.3), (H2 , 0.5), (H3 , 0.2)} P(S2 ) = {(H1 , 0.4), (H2 , 0.3), (H3 , 0.2)} (5) (6)

To know how well the company is doing on the Company Culture criterion, its performance on the sub-criteria needs to be aggregated. The ER aggregation algorithm is given as follows. Suppose sub-criterion (1) is more important with a weight of 0.3, and the weight for subcriterion (2) is 0.2. In the ER algorithm, the weights of sub-criteria need to be normalised to 1. Suppose there are L sub-criteria and ωi is the weight of the ith sub-criterion (i = 1, . . . , L), then the normalisation is defined so that 0 and

ωi

1

(7)

i=1

∑ ωi = 1.

L

(8)

In the example, the normalised weights for sub-criteria (1) and (2) are

ω1 = 0.3/(0.3 + 0.2) = 0.6 and ω2 = 0.2/(0.3 + 0.2) = 0.4
respectively. The ER aggregation algorithm takes the two assessments P(S1 ) and P(S2 ) as its input and generates a combined assessment, denoted by P(C) = {(H1 , β1 ), (H2 , β2 ), (H3 , β3 )}, as its output. The belief degrees in the assessment, β1 , β2 , and β3 , are obtained through the following steps: Step 1: Calculate basic probability masses pn and qn (n = 1, 2, 3): pn = ω1 βn,1 (n = 1, 2, 3) pH = 1 − ω1 = 0.4 pH = ω1 1 − ∑ βn,1
n=1 3 3

(9a) (9b)

=0 0.4

(9c) (9d)

pH = pH + pH = 1 − ω1 ∑ βn,1 = 1 − ω1 =
n=1

A Web Based Assessment Tool via the Evidential Reasoning Approach

189

P(S1 ) ⊕ P(S2 ) q1 = 0.16 {H 1 } q2 = 0.12 {H 2 } q3 = 0.08 {H 3 } qH = 0.6 {H} qH = 0.04 {H}

P(S2 )

Table 7.1 Probability Masses P(S1 ) p1 = 0.18 p2 = 0.3 p3 = 0.12 {H 1 } {H 2 } {H 3 } p1 q1 = 0.0288 p2 q1 = 0.048 p3 q1 = 0.0192 {H 1 } {Φ} {Φ} p1 q2 = 0.0216 p2 q2 = 0.036 p3 q2 = 0.0144 {Φ} {H 2 } {Φ} p1 q3 = 0.0144 p2 q3 = 0.024 p3 q3 = 0.0096 {Φ} {Φ} {H 3 } p1 qH = 0.108 p2 qH = 0.18 p3 qH = 0.072 {H 1 } {H 2 } {H 3 } p1 qH = p2 qH = 0.012 p3 qH = 0.0072 {H 2 } 0.0048 {H 1 } {H 3 }

pH = 0.4 {H} pH q1 = 0.064 {H 1 } pH q2 = 0.048 {H 2 } pH q3 = 0.032 {H 3 } pH qH = 0.24 {H} pH qH = 0.016 {H}

pH = 0 {H} pH q1 = 0 {H 1 } pH q2 = 0 {H 2 } pH q3 = 0 {H 3 } pH qH = 0 {H} pH qH = 0 {H}

qn = ω2 βn,2 (n = 1, 2, 3) qH = 1 − ω2 = 0.6 qH = ω2 1 − ∑ βn,2
n=1 3 3

(10a) (10b) (10c) (10d)

= 0.4 × 0.1 = 0.04

qH = qH + qH = 1 − ω2 ∑ βn,2 = 1 − ω2 × 0.9 = 1 − 0.36 = 0.64
n=1

where βn,1 and βn,2 are the belief degrees associated with Hn (n = 1, 2, 3) in (5) and (6) respectively. For example, β1,1 = 0.3 and p1 = ω1 β1,1 = 0.6 × 0.3 = 0.18. The calculated pn and qn are called probability masses assigned to grade Hn . The terms pH and qH in (9d) and (10d) are the remaining probability masses initially unassigned to any individual grades. The term pH consists of two parts, pH and pH , as shown in (9d). The first part pH represents the degree to which other criteria can play a role in the assessment. It should eventually be assigned to individual grades in a way that is dependent upon how all criteria are weighted and assessed. The second part pH = 0 because ∑3 βn,1 = 1 as n=1 calculated using (9c). Similarly, qH consists of two parts, qH and qH . Note that qH is not zero due to the incompleteness of the assessment P(S2 ), or ∑3 βn,2 = 0.9. In fact, pH and n=1 qH represent the remaining probability mass unassigned due to the incompleteness in their corresponding original assessments. They are proportional to their corresponding missing belief degrees and criterion weights, and will cause the subsequently aggregated assessments to be incomplete [31]. The values of those calculated probability masses are given in the 1st row and 1st column of Table 7.1. Step 2: Calculate combined probability masses: The ER algorithm aggregates the above probability masses to generate combined probabil-

190

D.-L. Xu

ity masses, denoted by rn (n = 1, 2, 3), rH and rH , using the following equations: rn = k(pn qn + pH qn + pn qH ) (n = 1, 2, 3) rH = k(pH qH ) rH = k( pH qH + pH qH + pH qH ) where ⎛ ⎜ k = ⎝1 − ∑ From Table 7.1, we have k = (1 − (0.0192 + 0.048 + 0.0144 + 0.0216 + 0.024 + 0.0144))−1 = 0.8584−1 = 1.1650 r1 = k(p1 q1 + pH q1 + p1 qH ) = 1.1650 × (0.0288 + 0.064 + 0.108 + 0.0072) = 0.2423 r2 = k × (p2 q2 + pH q2 + p2 qH ) = 1.1650 × (0.036 + 0.048 + 0.18 + 0.012) = 0.3215 r3 = k × (p3 q3 + pH q3 + p3 qH ) = 1.1650 × (0.0096 + 0.032 + 0.072 + 0.0048) = 0.1379 r H = k(pH qH ) = 1.1650 × 0.24 = 0.2796 rH = k( pH qH + pH qH + pH qH ) = 1.1650 × 0.016 = 0.0186
3 3 t=1 n=1 n=t

(11) (12) (13)

⎞−1

∑ pt qn⎠

(14)

If there are more than two sub-criteria, the combined probability masses can then be combined with the assessment on the third criterion in the same way. The process is repeated until the assessments on all sub-criteria are aggregated. The combined probability masses are independent of the order in which individual assessments are aggregated [17]. If there are several levels in a criteria hierarchy, the aggregation process is carried out from the bottom level until the top of the hierarchy is reached.

a utility score [13] can be calculated from the aggregated assessment.3623. 6]. denoted by U. (H2 . If necessary.0259 It could be partially or completely assigned to any combination of the three grades depending on the message in the missing information. indicating only 97.3215 = 0. 23] and supplier evaluation [19]. Applying the ER approach manually as discussed above is tedious but unnecessary.2796 Because of the incompleteness in one of the assessments. Suppose the final aggregated assessment is represented as follows: P(C1 ) = {(H1 .4463. If it were assigned to grade H1 .1915. U = ∑ u(Hi )βi = 0. new product design selection [16. they are given by 0. safety and risk assessment [24. β2 ). can be calculated as follows with the belief degrees as weights. .2423 0.2796 1 − 0. One of its latest applications is the prioritisation of customer voices for the General Motors Corporation [25]. More details on the properties of the ER algorithm can be found in Yang and Xu [32] and Yang [29]. Suppose the utilities for the 3 grades are given in (4). such as organisational self-assessment in quality management [30]. β3 )} Then βn (n = 1. product evaluation [25]. Therefore βH represents the combined effects on aggregated outcomes of missing information in the assessments on any sub-criteria. 3) are generated by: (15) βn = rn 1 − rH (16) For the example. The algorithm can be easily programmed and incorporated into different applications. The unassigned part is given by.9741 = 0. the aggregated assessment is also incomplete and it is shown by the sum of the three belief degrees which is 0.3364. the belief degree associated to the grade could be as high as Pl(H1 ) = β1 + βH = 0. (H3 . and β3 = 0. The expected utility score of the assessment given by (15).4147 i=1 3 (17) The score will normally be different from that calculated by using the weighted sum approach because the ER aggregation is a nonlinear process in which harmonic judgments will be reinforced more than proportionally and conflicting ones weakened accordingly. 2. βH = 1 − 0.9741.41 % of the belief degrees are assigned. The same applies to the belief degrees associated with grades H2 and H3 . β1 ). It has been applied to a wide range of real world problems.A Web Based Assessment Tool via the Evidential Reasoning Approach 191 The belief degrees in the aggregated performance distribution are calculated from the combined probability masses. β2 = = 0. β1 = 1 − 0.

The hierarchy provides users with a panoramic view of the assessment scheme – how many questions need to be answered and how they are related to each other in the hierarchy.192 D.Net and supported by a MySQL database.-L. Display the overview of the assessment scheme: The assessment questions are displayed in the left frame of the interface (Figure 7. Xu 7. The following are some of the features of the interfaces.4 Interfaces for Data Input To take full advantages of the ER approach.1 Criterion Hierarchy Displayed in Left Pane. in a hierarchical (tree) format. a web based assessment tool is developed.1). The interfaces of the web assessment tool are designed using ASP. There is a hyperlink to each question which directly leads users to the page (interface) where they can . Figure 7.

We believe that using belief structures to model it can fulfill such purposes. they still have the flexibility. as described later. then a belief structure can model almost all types of uncertainty. the automatically assigned belief degree will be 100 %. Therefore for users who are unwilling to deal with belief degrees and prefer conventional single-answer assessment. a common practice is to either guess the missing answers or reject the whole data set. It is different from and provides more flexibilities than most other assessment tools which normally allow only one assessment grade to be selected. Users may then edit the automatically assigned belief degrees to match a performance with grading standards which can be accessed from the interface if required. users can tick one or more check boxes beside the appropriate assessment grades or answers. Model uncertainty: If uncertainty means not 100 % sure. However for those who feel uncomfortable to categorise a mixed performance into a single grade. some obvious ones being subjective judgments. instead of just living with it.1). Considering that users may not be familiar with the flexibility provided. it will normally be required to eliminate the uncertainty first. clarify and reduce it as much as possible. as remarked by many industrial users and practitioners. For example. If there is uncertainty in collected data and an approach based on decision matrices is used for data analysis and information aggregation. This results in not only information loss or distortion. If only one answer is selected. a belief degree in percentage is assigned and displayed automatically beside the ticked box. probability distribution and ignorance. . and assign a belief degree to each grade.A Web Based Assessment Tool via the Evidential Reasoning Approach 193 enter or review their answers and comments to this question. when there are missing answers to some assessment questions. the belief degrees are assigned equally to the selected answers and add up to 100 %. It displays an assessment question and its multiple choice answers. The purposes of uncertainty modelling are to encourage users to understand. using belief structures to represent their assessments can boost their confidence in the aggregated assessment outcomes. Uncertainty is common in many decision making and assessment situations. To answer the question. the interface is designed in such a way that as soon as an answer is selected. The page is located in the lower right frame (Figure 7. Initially. but also a false sense of certainty in the outcomes. This is equivalent to conventional assessment schemes. Answer questions: Each question is given a dedicated page. Therefore belief structures allow different formats of uncertainty to be modelled and handled in a unified way.

Xu Figure 7. the information in all available data can be fully used in an assessment process. and meanwhile the effects of the uncertainties can be considered.194 D. By modelling uncertainties using belief structures. analysed and revealed.2 Side by Side Explanations of Question and Answers. the originality of data can be maintained. Those .-L.

the interface can also be used to collect quantitative data with or without probability uncertainty. For example. nor do they need to guess missing answers. The collected data are then recorded in the format of belief structures and stored in the supporting database.A Web Based Assessment Tool via the Evidential Reasoning Approach 195 are also some of the main purposes of modelling uncertainties in the ER approach. Figure 7. The process of assigning belief degrees is also a process of prompting users to exam and clarify any uncertainty. users neither have to approximately classify a mixed performance to a single grade.3 Aggregated Assessment Outcomes – Distributed Assessment Information. In addition to accepting qualitative assessments through grades and belief degrees as discussed above. Such a process can help to reduce uncertainty in a number of ways. Explain Question: It is often the case that a set of assessment criteria are developed by . Data with missing information is accommodated by assigning less than 100 % belief degrees.

196 D.1). When the calculation is complete. If a question needs to be explained in more detail.2). The list is getting shorter as more questions are answered.1). Save and retrieve data: The answers to each question are saved as soon as users click the “Save and Next” button (Figure 7. The explanation can be displayed in a separate browser at the top left corner of the main browser (Figure 7. users are directed to the pages where results are . Xu experts in the related area and words with special meanings may be used when assessment questions are asked. For some comprehensive assessment tasks. Saved answers to a question can also be retrieved for editing at any time by simply clicking the question displayed in the left frame of the browser (Figure 7. If users enter anything other than a number between 0 and 100 as a belief degree. To use this function. it is possible that they may take considerable amount of time to finish. which serves as a progress bar. the assessment question and the grading standards are displayed side by side so that users can map what they know about the assessed alternatives into the assessment grade or grades with appropriate belief degrees. Data validation: The sum of belief degrees assigned to the answers of any one question needs to be 100 % or less. The hyper link behind each question number can also be clicked to display the unanswered question if a user wishes to answer the questions in a different order. Therefore the interface is relatively error-proof. Explain Answers: Answers to assessment questions are often given in forms of qualitative grades. the aggregated overall performance and the performance in each sub-area are computed in the background by the ER algorithm. given that the answers are valid. Note that the Explain Question and Explain Answers browsers are in much smaller sizes and are placed in positions that are least likely to interfere with the focal points of the users. We call this process “Evidence Mapping”.1). In this way. The link “Explain answers” located at the top of the answers leads to a new web browser displaying the standards (Figure 7. The standards of those grades may need to be clearly defined in order to maintain consistency in assessments for different alternatives conducted by different web users in different locations at different times. a link “Explain this question” is placed beside the question (Figure 7.2). After all questions are answered. If it is the case.2). or the sum of the numbers entered is greater than 100 %.-L. a message will be displayed to remind them about the rule. Unless the correction is made the answers will not be saved. Monitor progress: A list of the un-answered question numbers is displayed at the bottom of the lower right pane (Figure 7. users have to give their email or user names as identities so that data associated to a specific user can be retrieved later. users can “Have a Break” and come back to finish it later.

4).6.5 Interfaces for Outcome Display The outcomes displayed in browsers are similar to those shown in Figures 7.A Web Based Assessment Tool via the Evidential Reasoning Approach 197 displayed (Figures 7.3 and 7. Figure 7.3-7. 7.4 Aggregated Assessment Outcomes – Performance Assessment Overview with Tailored Text Report. The tool provides four types of aggregated assessment information – text summary. distributed .

6.198 D. [15]. Note that there could be missing answers to one or more assessment questions.6. From this point of view. in contrast to local sensitivity analysis found in most literature. the scores are displayed in a chronological order to reflect trend of their opinions that may change over time (Figure 7.5). they may turn out to be favourable or unfavourable. it is assumed in the ER approach that they could be any values or answers appropriate to the question. analyses on the sensitivity of outcomes to simultaneous uncertain changes of different variables are normally referred to as global sensitivity analysis. the position of the user’s opinion (Figures 7. Xu performance assessments. It is labeled as unknown as shown in Figures 7. The distributed performance graph displays the aggregated performance distributions based on the assessments conducted by the current user and all the previous users as a group. The expected utility scores of the assessments conducted by the current user and all the previous users as a group are displayed in a different graph as shown in Figure 7. For users who have made a number of assessments over a period of time. If there are missing data.-L. Text summary explains what an assessment score means. The text summary giving an overview of assessment outcomes can be accessed by clicking on the top criterion in the criteria hierarchy displayed in the left pane of the browser (Fig- . in which uncertainty is considered for one factor at a time. but also how different their judgments are from those of all previous users on average.6). scores and sensitivity of the scores when there are uncertainties. The score ranges reflect the combined effects of all missing data. From Saltelli et al. The distributed assessment graph also displays the proportion of any missing information. By taking into account the full ranges in which the missing data may vary. The proportion of unknown elements will be non-zero and shown in the graph if the sum of belief degrees assigned to the answers of a question is less than 100 %.3 and 7. the lower and upper bounds of the scores are calculated using the built-in ER algorithm. Sensitivity of scores or rankings is displayed as grey areas in the graphs representing the scores (Figure 7. the ER approach provides a means for global sensitivity analysis in MCDA. the user knows not only the variations of the aggregated performance of the assessed alternative in any area (sub-criteria).4.3 and 7. as shown in Figure 7. The score ranges reveal whether alternative rankings could result from the missing data.6). Different formats of outcomes can be accessed from the same interface as follows. From the graph.4) and the key areas that the user needs to pay attention to regarding the performance of the assessed alternative (Figure 7.

The snapshots of the application of the tool to business innovation capability assessment are shown in Figures 7. the links hosted in the pane direct users to “answering the question” page. In the assessment stage when the focus of its users is on answering questions. supplier assessment and impact assessment of UK adopting the Euro. it is described how interfaces of the tool are designed to support users to take ad- . Further improvement and development are still under way.7 Concluding Remarks Real world decisions are often based on assessments and comparisons of alternative courses of action.6. the links direct the users to the “displaying outcomes” page. Those assessment schemes are developed through a number of research projects conducted in the Manchester Business School and tested by business users from the UK and other European countries [28]. 7.6 Applications The tool has been used in a number of areas including business innovation capability assessment.A Web Based Assessment Tool via the Evidential Reasoning Approach 199 ures 7. 7. Considering the data structure used in belief decision matrices and the new features provided by the ER approach. After all the questions are answered. To make a rational and reliable decision. It is demonstrated in this chapter by examples that the ER approach meets such crucial requirements by using belief decision matrices for problem modelling and the ER algorithm for information aggregation.6). The other types of information are available for any criterion in any level of the hierarchy and can be accessed by clicking the questions listed in the hierarchy. It is also crucial that uncertainty in data and its effects are taken into account so that risks of a decision associated with the uncertainty is fully understood and prepared for.3-7. The users have provided valuable feedbacks on how to improve the interfaces so that the tool can be more user-friendly and be used by users with different levels of computer literacy. it is crucial that information in all available data in different formats and from different sources is fully utilised and undistorted in an assessment process. Note that the hierarchy serves different purposes in different stages.4-7. They are complex in the sense that they often involve many criteria and uncertain knowledge. such as the “explain questions” and “explain answers” features. Some of the interface features described earlier are added by taking into account the feedback. A web based assessment tool is developed based on the ER approach.

impact assessment in supporting policy making. In the chapter the advantages offered by the tool are also illustrated using examples. the interfaces are made user friendly.200 D. The assessment criteria hierarchy on the left pane provides users with a holistic view of the assessment problem. The flexibility offered by its data input interfaces helps users to express their judgments more precisely and consistently. and Lin Yang of University College London for converting the pictures in the paper to the requested format by the publisher. such as through the “Explain Question” and “Explain Answers” links. business performance self-diagnosis in business management. Knowledge can be retrieved at where and when it is most needed by users. Domain expert knowledge in the applied areas are structured and stored in the assessment framework. such as data validation and automatic assignment of belief degrees. Xu vantages of the features. The tool can be used in a variety of applications such as contractor pre-qualification in supply chain and project management. The outputs of the tool include performance distributions and effects of uncertainties which allow users to gain more insights into the assessment problem. The web based assessment tool is also an effective and efficient knowledge management tool.-L. By using different web design techniques. and opinion surveys. . Acknowledgement The author would like to thank the support of the UK Engineering and Physical Science Research Council under grant No: EP/F024606/1.

5 Aggregated Assessment Outcomes – Key Areas to Consider for Performance Improvement. .A Web Based Assessment Tool via the Evidential Reasoning Approach 201 Figure 7.

-L.202 D.6 Aggregated Assessment Outcomes – Scores Ranges Showing Rank Sensitivity to Missing Data. Xu Figure 7. .

Chichester. 15. 329–356. 357–367. International Journal of Intelligent Systems. 1095–1107. [8] T. Decisions with Multiple Objectives. 37–50. IEEE transactions on Engineering Management. Technometrics. Pattern Recognition. 101–110. [10] R. Mathematical Theory of Evidence. “Decisionarium-aiding decisions. P. [6] K. [3] M. Srivastava (1995). Buchanan and E. [16] P. J. L. J. Morgan (2000). B. 6 (3). 28. Beynon. Expert Systems with Applications. A. 41 (1). [18] S. Decision theory . Multi-criteria Decision Making Methods: a Comparative Study. Multiple Criteria Decision Analysis — An Integrated Approach. Conti (2007). 30 (7). 1355–1368. Belton and T. F. [9] S. Marshall (2001). [19] M. R. Shortliffe (1984). Stewart (2002). I. 458–475. 20. P. “Knowledge discovery in distributed databases using evidence theory”. [15] A. A. “The Dempster-Shafer theory of evidence: an alternative approach to multicriteria decision modelling”. . Lee (2003). Marefat (2003). Hwang and K. Safety Science. [7] T. Yang. negotiating and collecting opinions a aa on the web”. N. H¨ m¨ l¨ inen (2003). Engineering Construction and Architectural Management. D. “An expert system for multi-criteria decision making using Dempster Shafer theory”. Yoon (1981). S. Yang and G. “Multiple criteria decision making in design selection and synthesis”. “The belief-function approach to aggregating audit evidence”. L. [22] E. 8 (3). G.Computational Intelligence in Complex Decision Systems 203 Bibliography [1] V. 745–761. [2] M. Sohn and S. J. Cosker and D. Saltelli. 10 (3). Journal of Engineering Design. The Accounting Review.: Princeton University Press. McClean (2000). Tarantola and K.J. “Addressing the contractor selection problem using an evidential reasoning approach”. Kluwer Academic Publishers. Ellis Horwood. H¨ m¨ l¨ inen (1995). Y. Pattern Recognition. M. 82. “Belief-function formulas for audit risk”. Ji and M. Sen and J. 207–230. [11] C. [13] R. H. M. The TQM Magazine. P. Springer-Verlag. European Journal of Operational Research. Triantaphyllou (2000). B. “A Dempster–Shafer approach for recognizing machine features from CAD models”. Holt (2001). Chin. Curry and P. “A history and review of the European Quality Award Model”. in press. 36 (6). P. “Analysis of evidence-theoretic decision rules for pattern classification”. “Preference programming through approximate ratio a aa comparisons”. Salo and R. Guo (2007). [5] D. Princeton. McTear and S. Yang (1995). Denoeux (1997). Journal of Multi-Criteria Decision Analysis. [17] G. ensemble and clustering to improve the classification accuracy for the severity of road traffic accidents in Korea”. 41. 1–14. 12 (2-3). Multiple Attribute Decision Making Methods and Applications. Omega. [14] A. 19 (2). 249–283. Raiffa (1976). Kluwer Academic Publishers. Shafer (1976). “An evidential reasoning-interval based method for new product design assessment”. Cambridge University Press. model independent method for global sensitivity analysis of model output”. 198–210. S. J. Reading. [21] R. 112–128. Beynon. ”A quantitative. 67 (2). Keeney and H. D. 39–56. Shafer (1992). B. [12] Q. French (1986). B.An introduction to the mathematics of rationality. Chan (1999). H. A. “Data fusion. Lam and M. [4] B. Sonmez. International Journal of Intelligent Systems. Srivastava and G. Rule-Based Expert Systems. MA: Addison-Wesley. Cai. [20] R.

Xie. Yang. J. Wang. in press. A. Yang (2001). 376–393. B. Yang (2006). L. and A. B. “An investigation into multiple criteria evaluation of consumer preferences under uncertainty”. R. 103–113. S. K. Kim and A. Xu. 113–127. 32 (1). P. “Use of advances in technology in marine risk assessment”.-L. New Zealand. Yang. Liu.L. Jenkinson and J. Ren (2007). Pillay. Ren and S. J. IEEE Transactions on Systems. and Cybernetics Part A: Systems and Humans. B. D. Wang. [31] J. 31–61. 32. European Journal of Operational Research. 42. Yang. B. “Self-assessment of excellence: an application of the evidential reasoning approach”. B. 32. . B. B. 664–673. Man and Cybernetics Part A: Systems and Humans. Journal of Marine Science and Technology. B. 131. [25] X. Xu [23] J. Dale and C. J. Risk Analysis. Yang. J. 39 (16). I. Yang and D. [27] D. L. Xu (2002b). McCarthy and J. Yu (2007). Maddulapalli (2007). J. H.204 D. [32] J. J. “On the evidential reasoning algorithm for multiple attribute decision analysis under uncertainty”. 52. Man. [24] J. “Intelligent decision system and its application in business innovation self assessment”. Reliability Engineering and System Safety. Auckland. 1011–1033. J. Wang. Siow (2001). 289–304. G. L. B. IEEE Transactions on Systems. Liu. [28] D. 24 (4). B. Yang and D. Xu. “Nonlinear information aggregation via evidential reasoning in multiple attribute decision analysis under uncertainty”. H. G. L. Submitted to 19th International Conference on Multiple Criteria Decision Making. J. Sen (1996). Xu. Xie. “Multi-person and multi-attribute design evaluations using evidential reasoning based on subjective safety and cost analysis”. “Ship selection using a multiple criteria synthesis approach”. L. Yang. Wang. Yang and P. Sii. G. Decision Support Systems. 3789–3812. Saajedi (2004). “Inference and learning methodology of belief-rule-based expert system for pipeline leak detection”. Xu. D. Expert Systems with Applications. “Rule and utility based evidential reasoning approach for multiple attribute decision analysis under uncertainty”. [29] J. S. [30] J. [26] X. J. International Journal of Production Research. Xu (2002a). B. L.

policy makers can objectively assess the nuclear proliferation intentions of states based on the current policy situation and also assess the effectiveness of new policy instruments in influencing nuclear intentions through simulation. This chapter suggests a new computationally intelligent nuclear policy simulator called DiNI (Discerning Nuclear Intentions) to objectively discern nuclear proliferation intentions of nation states in real time. 8. By using DiNI. “Valentina”.1 Introduction “It is easier to denature Plutonium than to denature the evil spirits of man” – Albert Einstein. While capability is a “tangible” D.1007/978-94-91216-29-9_8. 401. Thane (W). assessing intentions of nation-states is even more difficult. At the root of denaturing and predicting man’s ‘evil’ lies the assessment of his intentions. Computational Intelligence in Complex Decision Systems. DiNI is inspired from the well-validated theme of socio-biological evolution of altruistic behavior and works on a NetLogo platform. 400 607 India E-mail: sumanashokrao@yahoo.Chapter 8 An Intelligent Policy Simulator for Supporting Strategic Nuclear Policy Decision Making Suman Rao Risk Analytics and Solutions. Patlipada. Simulation conducted on DiNI using strategic decisions made under the historic “Atoms for Peace” program demonstrates that DiNI can indeed be used to proactively inform strategic policy decision-makers on unhealthy proliferation intentions of nation states. © 2010 Atlantis Press/World Scientific 205 . While assessing individual human intentions is in itself a difficult task. Hiranandani Estate. Conventional intelligence methods to detect proliferation intentions have been found to be inadequate/controversial in such cases [1]. Atlantis Computational Intelligence Systems 2.in One of the biggest challenges for strategic nuclear policy decision-makers is to discern nuclear proliferation intentions of nation states despite information deficiencies. These methods are usually centered on capability versus intent considerations. Ruan.co. The dilemma faced by strategic policy makers regarding nuclear intentions of Iran and North Korea is a case in point as to how ambivalent this assessment can be. DOI 10.

Consequently.206 S. is often not conclusive. the key elements of the international policy responses towards such states and the Decision Support Systems (DSS) for these policy responses rely on an unstructured combination of technical capability assessments and strategic intelligence information rather than on any objective modeling basis that integrates these elements.S.. the ethical and security aspects of simulation modeling as . Problem Statement: Considering the above practical circumstances.. as there have been many instances in the past where such assessments (e. the Bush Administration in the U.g. the strategic/military intelligence information to strategic policy decision makers on capability and intention is (in many cases) sketchy. how can strategic policy decision makers objectively assess the proliferation intentions of nation-states in real-time under given policy situations? In particular. By itself. Rao characteristic and can be objectively assessed to some extent by various physical means. objective assessment of intent which is an “intangible” is a challenge. this chapter proposes a nature-inspired real-time Policy Simulator called DiNI (Discerning Nuclear Intentions). though belonging to a tangible domain. For instance. In addition. International Atomic Energy Agency (IAEA) safeguard investigations) have been unable to detect clandestine developments of military applications of nuclear technologies. incomplete and vague. National Intelligence Estimate report on Iran was released in late 2007 with a ‘high-confidence’ that Iran had halted its nuclear weapons program in 2003 itself. was almost close to war with Iran based on Iran’s suspected nuclear ambitions[ibid]. the chapter gives a brief background of the existing methods. An additional issue consists of time-lag of information feeding the policy decision-makers. technical capability assessment. In order to demonstrate the practical applicability of the policy simulator.S. by the time the U. DiNI is a NetLOGO Agent Based Model (ABM) built using the socio-biological phenomenon of genetic evolution of altruistic behavior. First. followed by a discussion of the socio-biological background of the policy simulator and then the simulator construction in NetLogo is explained. Sections on validity and model verification. Subsequently. the experience of the USA vis-` -vis India in the a historic Atoms for Peace program is simulated and the results are discussed.g. can influence of strategic policy instruments e. the simulation runs are performed using hypothetical parameter values and the insights that were obtained to proactively support decision makers are highlighted. deterrence/incentives on nuclear intentions of target states be objectively evaluated prior to their deployment? As a response to the above question.

If they do not . Two such models are briefly explained below: Discourse Based Models [4]: E. by design. Sagan’s Security. Domestic Politics and Norms Model. States will seek to develop nuclear weapons when they face a military threat to their security that cannot be met through alternative means. intentions. Advanced models such as Real-Time Adversarial Intelligence and Decision-Making (RAID) of Defense Advanced Research Projects Agency (DARPA). As per the security model. economics. opinions and perceptions lies in applying interdisciplinary quantitative and computational social science methods from mathematics. use architectures based on game theoretic and cognitive approaches to discern the intention of opponents in war action. political science.g. military and strategic intelligence agencies apart from multilateral safeguard agencies such as the IAEA. But are these methods sufficient to address complex problems such as discerning the nuclear proliferation intentions of nation-states? According to a senior defense expert [3] “What technologies must we develop to understand and influence nation states. Type (i) is a normal architecture for most Dynamic Data Driven Simulation systems. sociology. this type does not fit most social science scenarios and an architecture that is somewhere in the spectrum between (i) and (ii) is recommended [2]. .. as reflected by the choice of model variables. However. However. neuroscience. .2 Existing Intelligence Methods in Discerning Proliferation Intentions of Nation-states Application of ‘intelligent’ methods for policy decision support is not new.An Intelligent Policy Simulator for Supporting Strategic Nuclear Policy Decision Making 207 well as further directions for research are also included in this chapter. these models rely on the world view of the analyst(s). Social simulations to support policy decision making are generally regarded as belonging to two categories: (i) Simulation pertaining to a single observed system and (ii) Simulation of a “typical system” – not of one particular instance. e. their cultures. and modeling and simulation.” Are existing models geared to meet the complex and multi-disciplinary nature of the above requirements? A limited survey of existing methods of ‘modeling’ nuclear intentions was performed and reveals different insights. 8. The DiNI Policy Simulator is designed to serve DSS needs for national strategic policy decision makers as in the national defense. statistics.. .g. the path to understand people. WMD proliferators. motivations. . cultural anthropology.

This model focuses on norms concerning weapons acquisition and postulates that States behavior is not determined by leaders or parochial bureaucratic interests. carefully delineated questions about which outcome each player would prefer. The domestic policy model focuses on the domestic actors who encourage or discourage governments from pursuing the bomb building. Rao face such threats. The history of the conflict.g. This is especially reflected in the choice of model variables and specific pre-determined questions posed to experts for their opinion.. While existing methods have distinct advantages. how important the issue is to each player. (iii) Model construction depends on the world-view of the analysts and experts. and how much influence each player can exert.208 S. Based on these inputs the model was run and the outcomes concluded. Sagan model on International theories and Mesquita models on game theory/rational choices. they will willingly remain non-nuclear states. de Mesquita (USA) based on rational choice/game theoretic approaches. (ii) Real-time capabilities of the models are limited. (iv) Capability to incorporate emergence properties of the problem is limited. De Mesquita models are reported to be accurate 90 % of the time.3 A Suggested Policy Simulator Based on the Genetic Evolution of Altruism Overcoming the lacunae (as described in Section 8. Computer Models [5]: These are typically Type (i) Models such as the ones being built by Prof. but rather by deeper norms and shared beliefs about what actions are legitimate [5]. Important domestic actors within the state are: 1) State’s nuclear energy establishment. DiNI takes inputs not only from experts but also .. DiNI is a policy simulator for strategic policy support which accepts real-time inputs from multiple sources/disciplines and helps objectively inform policy decision makers in assessing and influencing the nuclear intentions of target/suspect nation-states. or what the experts think will happen is not part of the model [7].2) of existing models for predicting nuclear intentions. this chapter proposes a nature-inspired real-time Agent Based Model (ABM) called ‘Discerning Nuclear Intentions’ (DiNI). Experts were asked narrow. 2) Units within the professional military. E. in the case of Iran this model assumed game theoretic interactions between 80 players. These players were stand-ins for real-life people who influenced a negotiation or decision. the cultural norms of the area. they have a few limitations as outlined below: (i) They lay an excessive emphasis on single disciplines –e.g. 3) Politicians or parties who favor nuclear weapons. 8.

they increase their own chance of being attacked. Parasiting on this spread of peaceful intentions are the unhealthy intentions of target/suspect nation-states that intend to take undue/illegal advantage of dual-use nuclear technologies and proliferate. Evolutionary socio-biologists hold that altruism triumphs due to the phenomenon of ‘inclusive fitness’. is a Dynamic Equilibrium .2 Peace as a State of Dynamic Equilibrium: Altruism in the Nuclear World In the nuclear world of dual-use technologies. this altruism should have actually wiped itself out in subsequent generations. this does not happen in reality. even though in doing so. where c = cost of altruism. the spread of nuclear cooperation is essential for constructive applications of nuclear energy.1 Inspiration from Nature: The Genetic Evolution of Altruism In socio-biological parlance. Policing multilateral agencies and individual/groups of nation-states are involved in detecting proliferation intentions and keeping such intentions under control with a ‘suitable’ policy package of incentives and/or deterrents. wherein the altruistic gene. Vervet monkeys warn fellow monkeys of predators. DiNI is inspired from a natural phenomenon: the socio-biological evolution of altruistic behavior and is therefore not dependent on the world-view of the analysts. 8. multibillion-dollar loans and economic reconstruction packages to national security guarantees. which is a function of continuous interactions of peaceful and proliferation intentions. this peace. In a world comprising of both altruists and their selfish counterparts.3.An Intelligent Policy Simulator for Supporting Strategic Nuclear Policy Decision Making 209 from other sources/data bases in real-time. the altruists are successful if c < b∗ r (called Hamilton’s inequality) [8]. For example. The objective of the policymaking is to overcome proliferation intentions of the target state and achieve the end objective of peace. These policy package incentives may range from restricted technology transfers. A brief background of the phenomenon of socio-biological altruistic behavior is given below. Under the conventional Darwinist intuition ‘survival of the fittest’. an organism is said to behave ‘altruistically’ when its behavior benefits other organisms at a cost to itself. However. The costs and benefits are measured in terms of reproductive fitness. b = altruistic benefit conferred and r = Sewall’s coefficient of relatedness. 8. though it reduces the reproductive fitness of the bearer. leading to the proliferation of the ‘selfish’ members of the species. The deterrents may range from diplomatic admonition and United Nations (UN) sanctions to an outright attack on installations/war. From a modeling perspective.3. survives due to an increase in the overall group fitness enabled due to altruistic behavior [7].

Also. particularly by way of drawing analogies between biological genes and human intentions. especially pertaining to interdisciplinary collaboration. it becomes essential to validate and verify the model and its premise very carefully. therefore. analogous to the DSE of this system based on the policy-driven change in the parameter input values. the same cannot be said for an agent-based simulation program” [10]. Simulation by itself has been described as a “third way of doing science” [9]. However. These are typically built on ABM platforms. The DE of peace exists in the system as long as the peace agents triumph over the proliferation agents. However. Therefore.3. Assessing the effectiveness of various policy instruments is. Is there an inconsistency involved? Some experts argue that the biological models of animal altruism are based on reproductive fitness consequences where no ‘intention’ as such in ‘helping’ other animals exists and. can nature’s altruism model abstraction be regarded as a valid inspiration for DiNI? 8. A more complete explanation of this analogy cuts across disciplines ranging from biology to social-psychology and is beyond the . the DiNI model has been ‘translated’ from animal data/modeling to humans and this translation involves making adjustments to human contexts. Nature’s altruism model is based on the socio-biological evolution of altruistic genes. it has also been acknowledged (by the same author) that “While a mathematical proof can usually be checked for accuracy without great difficulty. The DE of peace is achieved by overcoming the spread of unhealthy nuclear intentions (analogous to the selfish agents) by the peaceful intentions (analogous to the altruists) who offer the policy package at a cost to themselves. The following discussion gives key aspects of DiNI model validity/verification and is not exhaustive due to space limitations.210 S.3 DiNI Policy Simulator: Validity and Verification Policy simulations are used in a variety of situations ranging from agricultural/ land use planning to HIV policies. Information on the conditions/events/changes in motivations is then input in real time into this system. ABM-based simulations are widely regarded to offer tremendous advantages. whereas DiNI’s analogy to nature has been drawn with (abstract) intention agents. While the above seems to be a plausible model construction basis. caution attributing such behaviors to intentions. it needs to be clarified here that such claims also have their criticisms. therefore. The analogy between the socio-biological evolution of altruism and the nuclear proliferation context are now fairly evident. Rao (DE) of a complex dynamic nuclear world system requiring Dynamic State Estimations (DSE).

Agent rules are as follows: • Each patch first calculates its “fitness” and then contributes to a genetic “lottery” for its space. . • Fitness of selfish patch= 1 + b (No. • Based on the outcome of this lottery. of altruistic neighborhood agents)/5. There are four input parameters: • cost to altruist (c) • benefit of altruism (b) • harshness (h) and • disease (d). • Fitness of void agents is a constant value = h. Figure 8.1 A single ASASA VNN in the model with fitness representation.4 Modeling Basis for DINI .1). 8. of altruistic neighborhood agents). Further details on the modeling are provided in the next section. this model has been verified and validated from a computational modeling perspective [11].Nature’s Altruism Model in NetLogo [12] In this model. DiNI is based on an abstraction of the altruism model as built on NetLogo. 8. The model simulations provide insights into the circumstances under which the pink “altruistic” agents survive and the green “selfish” agents are contained in an environment that is made tough with Black “Void” agents. Fitness of an altruist patch = 1 − c + b (No. The NetLogo altruism model is freely available for download from the internet. Abstract agents called “patches” interact in a von Neumann Neighborhood (VNN) under conditions of harshness and disease (see Fig. a patch “survives” or “fails”.An Intelligent Policy Simulator for Supporting Strategic Nuclear Policy Decision Making 211 scope of this chapter. Based on a paper of Mitteldorf and Wilson.

for ma1t + Ma1t < nt < 1.1 above. 1] is generated for agent a1 ’s genetic lottery at time t. there are no “void” agents in the neighborhood. political motives of target Nation State. patch turns green.5. policy and instruments. Self wt).1) to the nuclear world analogies (see Section 8. f1 = 1 − c + b(3/5). 8. b. h. All Parameter values are dynamic and will change real-time.1 Calculating Parameter Values c. 8. for ma1t < nt < ma1t + Ma1t . b. f3 = 1 − c + b(4/5).) • Conduct genetic lottery: Let Major seed Ma1t = Max (Alt wt. b. h and d for DiNI The choice of parameters c.4) based on the existing circumstances. Self wt = ( f 4 + f 5)/F. patch turns pink.3) and computing input parameter values c. f5 = 1 + b(2/5). d (conforming to the altruism parameter ranges). patch turns black (void). If the model is maintained static with limited periodic updates to input values. h and d (Section 8. 8. Void wt = (0 + d)/F. . A random number nt ∈ [0.5 DiNI Model Construction The DiNI Model construction involves suitably adapting NetLogo Altruism model parameters (see Table 8. • Use a suitable method to normalize the multi-disciplinary inputs and compute values for c. Steps in the genetic lottery: (Illustrations for the VNN depicted in Fig. Rao In Fig. Suppose Ma1t = Alt wt for patch a1 at time t. sociological. If 0 < nt < ma1t . h and d has been made based on an inspiration from nature’s altruism model and does not depend on analyst’s world-view. • Calculate Alt wt = ( f1 + f2 + f3 )/F. f4 = 1 + b(3/5). Self wt) and Minor Seed ma1t = Min (Alt wt. The multi-disciplinary inputs are normalized based on a combination of PR (Perception Rating) and CI (Confidence Index). b. f2 = 1 − c + b(2/5). (In this case Figure 8. These values are determined based on the following steps: • Tabulate driving motivations (Motivational Indicators –MI) for the parameters based on assessment of economic.1. 8. then its utility will be greatly limited.212 S. cultural.1) • Calculate Fitness sum F = f1 + f2 + f3 + f4 + f5 + d.

expert opinions. PR is an objective assessment of motive strength affecting the model parameter. In strategic/military contexts.2-8. the comprehension of their meaning and their projection in the near future” [13]. Note: Situation Awareness is described as the “perception of elements in the environment within a volume of time and space. PR and CI are derived relative to the policy-making state.5 present an illustration of the base case value tables for parameter b. 5 reflecting the highest strength) and weighs the MIs using . Agent patches represent units of Nation State intention under particular situations of nuclear policy (either peaceful or proliferating) Peaceful intentions of policy making Nation state Represented by P agents Unhealthy intentions of suspect/target Nation State Represented by U agents Policy intentions Represented by V agents Net costs of incentivisation/deterrence policy packages (for P agents) Benefit of altruism (b) Net benefits of the policy package to the target state (for U agents) Strength of policy package provisions Harshness of environment (h) Disease (d) Tables 8.1 Altruism Model Elements Agent patches represent genetic units of social behavior (either altruistic or selfish) Pink altruist agents Adapting altruism model parameters to DiNI. The CI is an index number with which the confidence of the PR of the MI needs to be weighed in the DSS. Level 1 of the commonly known Endsley Model of Situation Awareness). Various advanced information synthesis methods are available with which to arrive at the PR and CI numbers as also their methods of weighting and normalization. The DiNI model illustration below takes hypothetical PR and CI numbers (on a scale of 1-5.g..An Intelligent Policy Simulator for Supporting Strategic Nuclear Policy Decision Making 213 Table 8. intelligence agency estimates etc. this confidence is an integral part of intelligence that is considered ‘actionable’. Both PR and CI are based on a synthesis of information pertaining to key motive indicators. CI denotes the confidence with which these perceptions are expressed. Altruism Input Parameters Cost of altruism (c) DiNI Parameters DiNI Elements Green “Selfish” Agents Black Void Agents Conditions under which both the P agents and U agents will die Initial proportion of altruists and selfish agents ∼Initial proportion of positive (p) and unhealthy (u) intentions. c. h and d. The PR is in fact a number arrived at as an outcome of the perception stage assessment of Situation Awareness (e.

37 12 + 9 + 6 = 75 0. h and d is thus obtained after taking into account a complex assessment based on multi-disciplinary MIs e. and sociological etc. Table 8.2-8.41 20 + 25 + 15 = 75 0.73 Table 8. Normalization of PR × CI 5 3 2 2 National/International popularity gain in as an upholder of peace Reduced risks of nuclear terrorism in own country Net Cost value c = 0.80 − 0.5 have been the outcome of learning from various world nuclear proliferation experiences in the past.g. MIs that reduce the perceived benefits received from the policy packages (b) National security and pride Regional competition Level of Mis-Trust in policy-making state Wtd.3 Value table for c – estimated net policy costs to the policy-making nation state. Rao a simple weighted average method of normalization. economical. cultural.36 = 0. Wtd. .39 = 0. Each value of c.2 Value table for b – estimated net policy package benefits to targeted nation state. MIs that increase the cost to the policy-making Nation State (c) Financial Cost of putting together the incentive/deterrent policy package Costs of Policing PR CI PR CI MIs that reduce the perceived cost of the policymaking Nation State (c) Long gains term Economic 4 5 4 4 5 5 3 3 NPT status and adherence to the treaty obligations by the target state..39 Where MI = Motivational Indicator. PR = Perceived Rating and CI = Confidence Index as explained above All the factors/motivations in Tables 8.214 S. political.80 16 + 9 + 4 = 75 0. Normalization of PR × CI PR CI PR CI MIs that increase the perceived benefits received from the policy packages (b) 4 3 2 3 3 3 4 5 5 5 4 3 Economic Admission into the world order Energy Security Net Benefit value b = 0. They are clearly not exhaustive in their scope and have only been mentioned here by way of illustration.73 − 0.36 20 + 20 + 15 = 75 0. b.

31 Table 8. b. The values p.6 and are interpreted in Table 8.73 − 0.73 12 + 1 = 0.26 = 0. End use verification provisions PR CI PR CI Factors/Forces/Features that reduce the perceived strength of the policy package (h) Reputation of the policy-making country in punishing package violations in the past Lack of widespread support to the package in own country Lack of support for the package from other key international actors in the policy-making arena Net strength of policy h = 0. The survival of pink P agents signifies effectiveness of the strategic policy package instruments and their extinction signifies policy failure.An Intelligent Policy Simulator for Supporting Strategic Nuclear Policy Decision Making 215 Table 8.11 4 3 4 2 3 3 5 1 5 2 Wtd.5 Value table for d – conditions under which both agents will die. Factors/Forces/Features PR that can increase the disease (d) Natural calamities Internal strife in target country An overriding influence of a third player Wtd. Factors/Forces/Features that increase the strength of the policy package (h) Quality and Quantum of Deliverables (policy covenants) of the target state according to the policy package. d and h in Table 8. Normalization of PR × CI 12 + 9 = 0.31 = 0.42 − 0.7 are hypothetical.47 8. u. c.7. Normalization of PR × CI 4 5 5 CI PR CI Factors/Forces/Features that reduce the disease (d) Weak political governance/administration Global Economic depression 5 4 3 4 1 3 1 20 + 20 + 15 = 75 0.26 50 Net Disease value d = 0.6 DiNI Model Simulation and Results The DiNI simulation results (time ticks > 200) are summarized in Table 8.42 50 8 + 5 + 10 = 75 0.4 Value table for h– strength of (existing) policy package provisions. The above hypothetical case illustrations of DiNI simulation results show how to discern .

7 Policy Simulation on DiNI: The Experience of USA vis-` -vis India under a the Atoms for Peace program 8.26 0..6 DiNI simulation run cases. u = 0.47 0.33 0. The DiNI Policy simulator is now illustrated with a real-life nuclear policy situation in order to demonstrate its validity and utility 8.1 Nuclear Cooperation Between USA and India .53 Case 3 p u c b d h = = = = = = 0. Case 1 (Base case)∗ p u c b d h = = = = = = 0.e.26 with assessed values of c.13 0.216 S. (based on Tables 8. h and d based on existing policy.5) The policy maker has now introduced new policy instruments (incentives) in such a way that c < b∗ 0.37 0. and a reduction in disease conditions (as compared to base case) . Case 1 Case Description Base Case: Equal probabilities of unhealthy intentions of target states and positive intentions of policy-making states i.7.26 0.20 0.2-8. 3 nuclear intentions of target nation states under various policy situations and evaluate the proliferation (intention) response to specific policy measures. Rao Table 8.31 0.88 Table 8. b. the United States actively promoted nuclear energy cooperation with India from the mid-1950s.26 0.10 0. building nuclear reactors .26 0.48 0.26 0. with changes in d. The policy maker has increased strength of policy provisions h though c and b do not satisfy Hamilton’s inequality Simulation results and Interpretation Result: Green U Agents survive. Green U agents extinct Success for the positive intentions of the policy-making nation state with the revised policy package of incentives and deterrents keeping proliferation risks under control Result: Pink P Agents survive. 2 Result: Pink P Agents survive.7 Illustration of how a policy maker can use DiNI for strategic decision-making.70 0.11 Case 2 p u c b d h = = = = = = 0. Pink P agents extinct The Unhealthy intentions dominate indicating that current state of the system of policy-making is ineffective in containing unhealthy intentions of the target Nation State.26 0. h. p = 0. Green U agents extinct Despite c > b∗ 0.26.2 (Hamilton’s inequality).A Brief Background As part of the 1950s-era Atoms for Peace (AfP) program.2. the new policy package still succeeds by getting proliferation risks under control through suitably increasing policy provision fitness (h).

Problem definition: On the basis of available information at that time (1950s–1960s) could the policy makers have objectively predicted the outcome of the cooperation of US with India under the AfP program? What further potential insights could DiNI provide for policy making countries such as the USA? As a solution construct to the above problem. Motivational Indicators (MIs) as elicited from historical published Strategic Intelligence information available during the era of AfP until the 1974 Indian explosion would be input into DiNI to objectively demonstrate how the policy simulation insights compare with reality as it unfolded. demonstrating that nuclear technology transferred for peaceful purposes could be used to produce nuclear weapons. 8. India did not join the NPT on the basis that it was discriminatory [14]. the United States refused nuclear cooperation with India for more than three decades and influenced other states to do the same. nuclear laboratories.S. Altruism Model Elements Agent patches represent genetic units of social behavior (either altruistic or selfish) Pink altruist agents DiNI Elements Altruism Input Parameters Cost of altruism (c) DiNI Parameters Green “Selfish” Agents Black Void Agents Conditions under which both the P agents and U agents will die Initial proportion of P and U agents ∼Initial proportion of positive (p) and unhealthy (u) intentions. and allowing Indian scientists to study at U. Agent patches represent units of Nation State intention under particular situations of nuclear policy (either peaceful or proliferating) Peaceful intentions of USA Represented by P agents Unhealthy/Proliferation intentions of India Represented by U agents AfP Policy intentions Represented by V agents Net costs of incentivisation/deterrence policy packages (for P agents) Benefit of altruism (b) Harshness of environment (h) Disease (d) Net benefits of the policy package to the target state (for U agents) Strength of policy package provisions . providing heavy water for the CIRUS reactor.8 DiNI Atoms for Peace (AfP) policy simulator – characterization.8 characterizes the DiNI AfP policy simulator constituents.7.2 Construction of the DiNI Atoms for Peace Policy Simulator Table 8. However. As a result. India exploded a “peaceful” nuclear device.An Intelligent Policy Simulator for Supporting Strategic Nuclear Policy Decision Making 217 (Tarapur). This policy stand has now been dramatically reversed through the recent (2008) Indo-US nuclear deal under the GNEP program. Table 8. In 1974.

b. The intelligence information is limited to the extent of declassification by the U. to feed into the Motivation Indicators.g.3 Calculating Parameter Values for c. sociological. and [19] • Declassified strategic intelligence documents of the government of USA (NIE) have been used for the purposes of this illustration. 17. Rather the perceptions of USA on the motives of India is what is input into the policy simulator. 18]. some necessary assumptions have been made: (i) Default PR and CI is 4 on a scale of 1-5 (5 being the strongest) . Government. cultural.218 S. • There is scope for including information from civilian sources e. and to test the ability of DiNI Policy Simulator to proactively inform decisionmakers: (a) no information source post the 1974 ‘peaceful’ explosion of India has been included. (b) no Indian interpretation of the events has been included since DiNI simulation in this example is run from the view point of USA. In order to avoid a “wisdom in hindsight” bias. b. PR and CI are derived relative to the viewpoint of USA. Since the information is over 50 years old and it is now not possible to obtain a Perception Rating and Confidence Index without wisdom in hindsight bias. 16. Rao Note 1: This illustration is for the sole purpose of demonstrating the utility and validity of DiNI and is not to be construed as throwing adverse light on either India or the USA.5) and values for c. 8. The multi-disciplinary inputs are normalized based on a combination of PR (Perception Rating) and CI (Confidence Index).S. h. d (conforming to the simulator parameter ranges) computed. However this has not been included for present illustration purposes due to constraints in space and time. • The strategic intelligence documentation included in this simulation is illustrative and not exhaustive. Note 2: Information sources used for this illustration are cited under [15.7. Then a simple method is used to normalize the strategic intelligence inputs (as was discussed in section 8. economic databases. h and d for the DiNI Atoms for peace policy simulator First driving motivations (Motivational Indicators – MI) for the parameters based on assessment of economic. political motives of target Nation State are tabulated time-period wise (In this case year wise)..

Missions and Info programs Value table for c – Estimated net policy costs to the USA.9 and 8. 8. ‘very clearly’ are taken to indicate high confidence and get a CI of 5. Table 8..Base Case a Tables 8. Parameter +/− c − Perception Confidence PR∗ CI Rating Index 5 5 −25 Year 1957 Source Joint progress report c − 5 5 −25 1957 Joint progress report c − 5 5 −25 1957 Joint progress report Joint progress report Joint progress report Joint progress report c − − 5 5 −25 −25 1957 c 5 5 1957 c − 5 5 −25 1957 c − − 5 5 c 5 5 Joint progress report Joint −25 1957 progress report Continued on next page. which are indicators of weak confidence get a CI of 3. a. CI rating is then varied.g...An Intelligent Policy Simulator for Supporting Strategic Nuclear Policy Decision Making 219 (ii) Depending on the qualifying adjectives for the MI.4 The AfP Experience of USA vis-` -vis India . likely etc. −25 1957 . b.7. E.9 MI Description Maintain US’s Leadership Position in the development of peaceful uses of nuclear energy To promote peaceful uses of nuclear energy in free nations abroad as rapidly as possible to promote cohesion within the free world nations to forestall Soviet Union’s similar objectives to assure continued access of nuclear materials to the US to prevent diversion of fissionable materials to nonpeaceful uses Distribution of nuclear material Conferences. The words probably. The 1957 joint progress reports submitted by the State Department and the AEC on implementation of NSC 5507/2 “Peaceful uses of atomic energy” have been used in framing the base case values [20].10 present an illustration of the base case value tables for parameter c and b. the words ‘strong’.

Missions and Info programs Power reactor agreements and prospects of nuclear commerce Power reactors required for use abroad to be developed in US to maintain US technological leadership Aid for Research reactors to be provided for select countries Educate and train foreign students b + + + 5 5 25 1957 b 5 5 25 1957 b 5 5 25 1957 b + 5 5 25 1957 b + 5 5 25 1957 Joint progress report Joint progress report b + 5 5 25 1957 .10 MI Description Value table for b – Estimated net policy package benefits to India. Parameter +/− Perception Confidence PR∗ CI Rating Index Year Source Joint progress report Joint progress report Joint progress report Joint progress report Distribution of nuclear material Conferences.9 – continued from previous page Perception Confidence Parameter +/− PR∗ CI Rating Index c − 5 5 −25 Year 1957 Source Joint progress report Joint progress report c − 5 5 −25 1957 c − 5 5 −25 1957 Joint progress report Joint progress report c − 5 5 −25 1957 Overall value for c is −1. Rao MI Description Power reactor agreements and prospects of nuclear commerce Power reactors required for use abroad to be developed in US to maintain US technological leadership Aid for Research reactors to be provided for select countries Educate and train foreign students Table 8. Table 8.220 S.

7.An Intelligent Policy Simulator for Supporting Strategic Nuclear Policy Decision Making 221 Applying scale translation. then the overwhelming categorization was considered.11 Year b − Value tables for year 1961. c.The 1960s a For this evaluation. Prestige and ability to be neutral (due to potential Chinese nuclear program) + Cost and reluctance to divert resources from current economic programs Leadership decision on program not to go the weapons way Psychological and political factors opposing nuclear weapons c − Lack of technical capability (snags) + Developing plutonium separation plant Developing indigenous uranium resources h − Expressed desire to avoid safeguards . thereby restricting the complexity in calculations. 8.9 above to be repeated as benefits in Table 8. (However DiNI provides flexibility to include multiple classifications as well.05 for d have been assumed based on literature study of the crucial AfP decade (1957–1967).) Overall net values of b. Note: There is no specific mention on the benefits for recipient countries in the Joint progress report. c. 1961 25 −50 −25 −25 75 25 25 25 15 −25 −25 40 20 20 −20 −20 −20 Neighbor (China’s) foreign policy Security.5 illustrations.2 for h and 0. h and d were calculated in year-wise value tables (Tables 8. There is scope to further validate these two values with more information pertaining to that decade.11-8. Table 8. A conservative (subjective) value of 0. Subsequently the author has assumed those relevant points from Table 8. h or d category. The MIs were first classified as belonging to either of b. If the same MI could have been classified under two categories.5 Evaluating the AfP Experience of USA vis-` -vis India . a total of 77 MIs were elicited from the strategic intelligence information (NIE Reports).16) similar to the methods demonstrated in Section 8. The individual year wise value tables are provided below for information (1962 data were unavailable to the author).10. c = −1 + 1 = 0 Overall value for b is 1.

Year b − As China’s arsenal improves the views of the military leaders are likely to change Chinese test in late 64 Development costs of nuclear weapons not a sufficient deterrent to India Economic burden in developing few simple fissile weapons is not great and only small parts of these would involve foreign exchange If India tested.14 Value tables for year 1965. .12 Year b − Value tables for year 1963. 1963 20 −25 −25 45 20 25 50 50 25 25 Chinese threat may affect the progress and direction of the Indian nuclear program + political leadership unlikely to go ahead with the weaponisation Psychological and political factors opposing nuclear weapons c If dependency on US for heavy water is eliminated then can independently produce plutonium without safeguards India improving overall nuclear capabilities Table 8.13 Year b − Value tables for year 1964. Rao Table 8. 1964 −40 −40 −20 −20 20 20 20 Changes in Sino-soviet relations scope and pace of the Chinese nuclear program c + Basic facilities existing including plutonium separation plant Table 8. it will claim it was for peaceful purposes thereby avoiding damage to prestige and not violated its agreement with Canada on the CIR Inability to stand up to a nuclear-armed China 1965 −270 −375 −15 −20 −25 −15 −25 −25 −15 −25 −25 Indian prestige and acknowledgment as a great power is not possible without nuclear weapons Political leadership will support technical efforts to shorten time between decision and detonation of a test device Public sentiment for nuclear weapon building −25 Severe economic difficulties might deter the testing decision −15 Indo-Pakistani war has strengthened shift towards nuclear weapons −25 Threat of Chinese intervention in Indo-Pak fighting has given a boost to weapons proponents −20 No scientific spectacular achievement to counter Chinese achievements of nuclear tests −25 Continued on next page..222 S..

14 – continued from previous page Year Pace and scope of the Chinese nuclear program A renewal of war with China supporting Pakistan might cause New Delhi to opt for the bomb Declining legacy of Gandhi (Non-violence and Nehru) Costs would be prohibitive (as per some Indian quarters) + Apprehensions of Setting up a nuclear weapons race with Pakistan Build up of conventional weapon strength is preferred by military leadership Conventional arms are what is required the most in view of the Indo-Pak war Indian military focussed on borders and not on strategic capabilities and hence not pressing for nuclear weapons Reversal of Nehru’s position will lead to damage of international prestige Political leadership’s position strengthened who has non-proliferation leanings c − No evidence of advanced weapons design and test site identification Limited work on missile technology so far and sourcing missile systems will be a challenge + India has capability to test a device within a year of the decision India has the capability to develop nuclear weapons Plutonium production facilities make it doubtful whether it is for only peaceful purposes New Delhi will probably detonate a nuclear device and proceed to produce nuclear weapons within next few years d + If China attacks India with nuclear weapons. h − India regards that Moscow won’t cut off aid to influence nuclear policy Moscow’s silence when China threatened to intervene in the Indo-Pak war Moscow’s support will ensure Western support as well Not willing to insure security through a test ban treaty Political leadership will avoid committing to international agreements that will curtail India’s options stiff stand on non-proliferation treaty Suspension of US military aid to India and US’s failure to prevent Pakistan’s use of US weapons against India New Delhi won’t accede to NPT since China is not limited in weapons devpt as per NPT and China won’t accept such restrictions China has not suffered any set-backs on account of its nuclear capability and its status has been enhanced + Meaningful international progress in disarmament might deter the testing decision Political leadership will weigh assurances. the US and USSR may become inevitably involved. inducements and pressures forthcoming from other nuclear powers Political leadership’s immediate course of action is to keep diplomatic and technical options open Securing international gurantess 1965 −25 −15 −15 −20 105 15 15 20 15 15 25 35 −40 −25 −15 75 15 20 15 25 15 15 15 −100 −180 −15 −20 −15 −15 −25 −20 −20 −25 −25 80 15 25 25 15 .An Intelligent Policy Simulator for Supporting Strategic Nuclear Policy Decision Making 223 Table 8.

b. There is a data availability gap on NIE sources the author faced for 1962-hence there is a break in the yearly time series).224 S. .16 Year b − Value tables for year 1967. Rao Table 8.15 Year b − Value tables for year 1966. h and d values were cumulated and DiNI was run again. b.18 below for all the years ranging from 1961 to 1967. 1967 −95 −95 −20 −20 −25 −15 −15 −10 −10 −10 Fear of Sino-Pakistani collaboration War with Pakistan (Self-reliance in Defence matters) Chinese detonation of 1966 No sign of satisfactory security guarantee from USA/USSR Cessation of US assistance during 1965 Indo-Pak war c − Huge demand for cost. h and d values are mentioned in Tables 8. 1966 −85 −110 −15 −15 −15 −25 −25 −15 25 25 −75 −75 −25 −25 −25 China’s growing nuclear strength National Security Prestige Spectre of Pakistani-Chinese collaboration New Delhi won’t accede to NPT since China is not limited in weapons devpt as per NPT and China won’t accept such restrictions Strengthening its bargaining position + Political leadership believe that this is not reqd c − Food situation critical Present and prospective economic difficulties India has capability to produce nuclear weapons and test a first device within one year of decision Table 8. The cumulated c.17 and 8. foreign exchange and trained manpower a deterrent Every year c.

26 0.26 0.00 0.00 1.05 0.00 0.15 0.05 From Value Table −0.13 0.00 0. c.. p u c b d h = = = = = = 0.00 0.00 1966 After scale translation 0.00 0.90 0.40 0.26 0. 1966 and 1967.00 0.90 −0. b.20.20 p u c b d h = = = = = = 0.05 From Value Table −0.6 DiNI Atoms for Peace Simulation and Results The DiNI simulation results (time ticks > 200) are summarized in Table 8. c.19 and insights are interpreted in Table 8.15 −0.7.26 0.00 −0.55 −0.80 0.13 −0.00 0.20 −0. h and d values tailored to model scale From Value Table 0.80 0.63 0.00 0.90 0.05 0..27 −1.7.60 After scale translation 0.00 0. 1961 Initial Values Base case b c h d 0.13 1.00 0. Table 8.00 0.00 1967 After scale translation 0.05 1963 1964 Revised b.65 From Value Table −0.00 0.18 Cumulative value tables for years 1965.60 8.00 Cumulative 0. .60 0. d and h in Table 8.65 From Value Table 0.19 DiNI AfP simulation results.80 −0.80 0.00 p u c b d h = = = = = = 0.05 0.26 0. 1965 From Value Table b c h d −0.60 0.90 0.00 0.65 Cumulative −0.05 After scale translation 0.10 1.60 0.90 0. u.00 Continued on next page.00 0.05 After scale translation 0.20 0.00 0.An Intelligent Policy Simulator for Supporting Strategic Nuclear Policy Decision Making 225 Table 8.90 0.60 0.60 0.00 0.55 0.00 0.7 are as derived from Section 8.00 0.00 0.2 above.80 0.05 Table 8.80 0.55 −0.13 0.20 −0.76 −0.00 Cumulative 0.80 0.80 0. The values p.65 Cumulative −0. The survival of pink P agents signifies effectiveness of the strategic policy package instruments of USA and their extinction signifies policy failure.05 After scale translation 0.80 0.17 Cumulative value tables for years 1961.60 −0.00 0.26 0.00 Cumulative 0.65 Cumulative 0.05 0.60 0. 1963 and 1964.90 0.

Pink P agents extinct The US intentions dominate indicating that current state of the system of policy-making is effective in containing unhealthy intentions of the target nation states. Result: Green U Agents survive. Result: Pink P Agents extinct.65 0. Continued on next page.26 0.26 0.05 0. i.00 1967 p u c b d h = = = = = = 0.00 1966 p u c b d h = = = = = = 0. Proliferation intentions of India are stronger than the peaceful intentions of the USA.00 0.e.15 0. This results in a violation of Hamilton’s rule 1964 .26 0.00 Table 8. Green U agents survive Signifies policy failure. p = 0.90 0.55 0. 1961 Equal probabilities of unhealthy intentions of target states and positive intentions of policy-making states. Green U agents survive Signifies policy failure..00 0.26.26 with assessed values of c.90 0. Rao Table 8.20 Case Year 1957 Base Case Illustration of proactive DiNI insights on Atoms for Peace policies.226 S.26 0. (based on Tables 8.5) No change in Simulator parameter values despite cumulating parameter values and applying scale transitions (see Table above) 1963 There has been a disproportionate increase in costs with d and h not changing.2-8.26 0. Proliferation intentions of India are stronger than the peaceful intentions of the USA. u = 0.65 0.26 0. h and d based on existing policy.19 – continued from previous page 1964 p u c b d h = = = = = = 0.00 0.13 0. This results in a violation of Hamilton’s rule Further increase in costs with d and h not changing.00 1965 p u c b d h = = = = = = 0.26 0. Case Description Simulation results and Interpretation Result: Green U Agents survive.65 0. Pink P agents extinct The US intentions dominate indicating that current state of the system of policy-making is effective in containing unhealthy intentions of the target nation states Result: Pink P Agents extinct.. b.26 0.

(ii) Intensifying peaceful cooperation with India. DiNI can help evaluate the effectiveness of policy decisions and options. Providing unilateral security guarantee to India in the event of nuclear attack by China. Some of them were (i) Security guarantees [21. India’s leanings towards proliferation were already well known to the intelligence community. with Hamilton’s inequality still not prevailing. b. Collaborative project on recycling Plutonium for India’s nuclear reactors. Providing Joint-guarantee with USSR to India in the event of nuclear attack by China. 1974.20.7. Green U agents survive Signifies policy failure. c. h has now increased but the increase is not sufficient to offset the disproportions between b and c Result: Pink P Agents extinct. 22]. b. There would be a counter argument that whether or not DiNI was available. 8. Proliferation intentions of India are stronger than the peaceful intentions of the USA. DiNI has successfully simulated AfP policy situations and accurately predicted the proliferation intentions of India.7 Atoms for Peace: Some Policy Options for the USA During this period (1957–1967). .India conducted its first ‘peaceful’ nuclear explosion signifying policy failure and triumph of green U agents As has been practically demonstrated in Table 8. 1966. Collaborative reactor construction projects. military leaders and the like).20 – continued from previous page Case Description Simulation results and Interpretation Both cumulative costs and benefits are at their maximum.An Intelligent Policy Simulator for Supporting Strategic Nuclear Policy Decision Making 227 Case Year 1965. d. a. there were various policy options available to the USA to influence nuclear intentions of India. Cooperative “Plowshare” projects-nuclear explosions for civilian uses. Indo-US cooperation on peaceful uses of nuclear energy to enhance scientific prestige. 1967 Table 8. [23] a. Hence what is the value-add DiNI brings to the whole process of discerning nuclear intentions? Apart from the fact that DiNI provides an objective basis to assess nuclear intentions (which was hitherto based on subjective opinions of state leaders. scientists.

228 S. (ii) Provides an objective. Instead of the above policy options. (iii) DiNI Inputs are predominantly driven by analogies to nature’s altruism model and thus reduces the chances of analyst world-view biases. Such an evaluation can be performed by suitably reworking the c. Ambassador at Large etc. The US diplomatic channels were briefed to: [24] (i) Abandon rhetoric suggesting special prestige of nuclear-weapon nations (ii) Emphasize Wisdom of a national decision to forego development of nuclear weapons (iii) Exhort the benefits of civilian nuclear energy use (iv) As far as China’s nuclear attack possibility is concerned. strategic advisors to governments and nuclear strategic policy decision-makers.. b. Rather the policy decision not to implement the above went more on the basis of opinion held by key influencers in the US e. accentuate US’s capacity for retaliation Again the above were a result of discourse oriented strategic thinking and had no objective policy impact analysis basis. DiNI avoids the lacunae in existing policy simulation modeling (Section 8. Rao None of the above policy options were objectively analyzed for policy impacts. However due to limitations of space individual proliferation impact assessment of the above policy options of Atoms for Peace have not been considered in this chapter. emergent view of proliferation to serve the DSS needs of strategic policy decision-makers in real-time. the following policy decisions were taken which had little impact on influencing the Indian intentions (U agents). Joint chiefs. h and d parameter values and running them through the simulator under various policy scenarios. 8.3) and offers the following distinct advantages (i) Provides a platform to combine multi-disciplinary inputs with a scope to include a wide range of objective/mathematical methods of aggregation. (iv) The results of DiNI are easily ‘visible’ and conducive to interpretation even to a predominantly non-technical audience such as politicians. DiNI facilitates an objective evaluation of such policy options using simulation.8 Advantages of the DiNI Policy Simulator As demonstrated both by a hypothetical example and the Atoms for Peace Program illustration.g. .

9 Ethical and Security Aspects of DiNI Policy Simulator According to the code of ethics for simulationists. building low-level ‘intelligence’ into the agent patches... assuming that the latest real – time DiNI model is available to a rogue user: (i) The rogue user can take on the ‘role’ as the altruist and consider benevolent states as selfish (viz. infidel etc. In accordance with these maxims. (ii) The rogue user can identify himself/herself with the selfish agent and look for events/circumstances when the altruists lose and selfish agents win in the model thereby objectively and systematically countering strategies of the Altruist strategic policy makers. However DiNI is a dynamic model by design and multi-disciplinary inputs feed into the system real-time based on changes in events and circumstances e. However this goes against the very conceptualization of ‘helping others at a cost to oneself’ which is inherent in the model and hence will render the whole simulation exercise meaningless.). the entire platform of DiNI is based on the open architectural model of altruism available on the net freely for anyone to download. d values will increase in the case of terrorism) and prompt strategic policy decision makers/politicians to act in a manner so as to bring back the dynamic equilibrium of peace. . Would harm to humans be possible if DiNI falls into the wrong hands? The design philosophy of DiNI is based firmly on Claude Shannon’s maxim that “the enemy knows your system”.g. different neighborhood values within VNN. constructing and analyzing second-order emergence and self-organizing approaches to containment of nuclear proliferation.g. additional parameters etc. evil. 8. an ethical simulationist is expected to “support studies which will not harm humans (current and future generations) as well as environment” [25].An Intelligent Policy Simulator for Supporting Strategic Nuclear Policy Decision Making 229 8. From a model misuse point of view.10 Further Research Directions Further research directions include extending the DiNI Policy Simulator into different cellular automata neighborhoods.. an orchestrated terror attack will trigger real-time change in DiNI simulator parameter values (e. dealing with groups of agent behavior rather than individual agents. Also supported by DiNI design is Kerchoff’s design principle for military ciphers – “it must be able to fall into the hands of the enemy without inconvenience” [26].

uk/research/sgp/aimss/20051001 kennedy PolicyDecisionSupport. [4] Schimdt O.ncess. . Towards Intelligent Data-Driven Simulation for Policy Decision Support in the Social Sciences can be accessed at http://www.230 S. Simulation runs of DiNI with hypothetical parameter values demonstrated how DiNI can help discern current state of proliferation intentions as also evaluate how future policy changes will impact these intentions. By mimicking nature’s model of altruism. and Theodoropoulos G. Padma Prakash. who need to make key decisions on the basis of a reliable proliferation intention assessment of target nation states.ac. It serves DSS needs of strategic policy decision makers from intelligence agencies/governments /defense etc. DiNI enables simulation modeling which is fairly independent of the world-view of the analyst and also provides the flexibility to include inputs from multiple disciplines in a strategic decision-making environment. November 2007. Iran: Nuclear Intentions and Capabilities.) Da RUAN (SCK • CEN) for their encouragement in making this chapter happen. [2] Kennedy C. Intelligence Report of National Intelligence Council (USA).. Bibliography [1] National Intelligence Estimate. The ethical aspects of the DiNI Policy simulator and possible further research directions were also discussed in brief. Presentations of DARPA Tech 2005. Rao 8. Understanding & Analyzing Iran’s Nuclear Intentions – Testing Scott Sagan’s Argument of “Why do States build Nuclear Weapons” MA International relations Dissertation Lancaster University.. Utilizing Social Science Technology to Understand and Counter the 21st Century Strategic Threat.11 Conclusions This chapter suggested a nature-inspired policy simulator called DiNI (Discerning Nuclear Intentions) for objectively discerning unhealthy nuclear proliferation intentions of target nation states. Acknowledgement I am grateful to M. DiNI is based on the socio-biological evolution phenomenon of altruism and works on a NetLOGO ABM platform. I devote this chapter to the precious affection bestowed on me by my dear elder Brother and the pure sincerity of my dear younger Sister Mrs..pdf [3] Popp R. The experience of USA vis-` -vis India in the historic Atoms for a Peace program was simulated in DiNI and DiNI demonstrated the inevitability of India moving towards nuclear-weaponisation based on the contemporary strategic intelligence information. 2008. Ludo VEUCHELEN (SCK • CEN) and Prof (Dr. Aug 2005.

[9] Axelrod R. [20] Joint Progress Report (State Department and AEC) on implementation of NSC 5507/2 – “Peaceful uses of nuclear energy”... J.gwu. Declassified can be accessed at http://www...northwestern. No. ISBN 0805821341. [19] National Intelligence Estimate. 2002 . 1957 Declassified [21] Memorandum from the State Department. TI. Likelihood and Consequences of a proliferation of Nuclear Weapon Systems. Theor..unige. Lawrence Erlbaum Associates. 1967 Declassified. et. volume 2.edu/netlogo/models/Altruism [13] Endsley M.).. CRS Report for Congress. Nuclear Cooperation With India: Issues for Congress. 4-66. 1964. “Biological Altruism”.html [23] Memorandum from the State Department. (1998). [18] National Intelligence Estimate. 1964.edu/∼nsarchiv/NSAEBB/NSAEBB6/index.. 3. 1965 Declassified can be accessed at http://www. [14] Squassoni S. S. 9780805821345. [17] National Intelligence Estimate. Biol. 1963 Declassified. 2006.edu/∼nsarchiv/NSAEBB/NSAEBB6/index. 4-63. pp. Handbook of Computational Economics. 1965 Declassified can be accessed at http://www. 2003d. Situation Awareness: Analysis and Measurement.etu. 12. chapter 33.gwu.gwu. Why do states build nuclear weapons? Scott Sagan – 1996/97 as accessed at http://home.al. 1565–1584.” January 30.J.” in: Leigh Tesfatsion & Kenneth L. 1964 Declassified can be accessed at http://www. Evanston. [15] National Intelligence Estimate (NIE) 4-3-61 Nuclear Weapons and Delivery Capabilities of freeworld countries other than the US and the UK. Proceedings of the 2002 Summer Computer Simulation Conference. Prospects for a proliferation of Nuclear Weapons over the next decade. Ambassador at Large: “Indian Nuclear Weapons Capability. [10] Axelrod R. 1961 Declassified [16] National Intelligence Estimate.ch/∼grafale3/spo/spo-read/2b2 sagan mathias.D.html ¨ [25] Oren.” January 30.R. A Code of Professional Ethics for Simulationists.sciencenews.. Proliferation of Missile Delivery Systems for nuclear weapons.html [22] Memorandum from the State Department. 2000 Preview courtesy: Google book search.Computational Intelligence in Complex Decision Systems 231 [5] Bergundthal M. The likelihood of further nuclear proliferation 1966 Declassified. “Agent-based Modeling as a Bridge Between Disciplines. Dec. Public Stance Nuclear Proliferation. Northwestern University. NetLogo Altruism model. Advancing the Art of Simulation in the Social Sciences Journal of the Japanese Society for Management Information Systems. The Stanford Encyclopedia of Philosophy (Summer 2005 Edition).edu/netlogo/models/ [12] Wilensky U.edu/∼nsarchiv/NSAEBB/NSAEBB6/index. The genetical evolution of social behavior.gwu.. Web edition of ScienceNews can be accessed at http://www.. 4-2-64. Mathematical Fortune-telling How well can game theory solve business and political disputes?. edition 1.pdf [6] Rehmeyer J. 1966. Garland D. IL http://ccl. October 27. Elsevier. 4-67.” December 31. 7:1-16. Zalta (ed. Edward N. 1964 Declassified.northwestern. http://plato. Center for Connected Learning and ComputerBased Modeling.edu/archives/sum2005/entries/altruism-biological/ [8] Hamilton W. Judd (ed. Ambassador at Large: “New Delhi’s 1862 of December 31.S.html [24] US State Department Cable Regarding a U. [11] NetLogo Models Library download URL: http://ccl. Ambassador at Large: “Indian Nuclear Weapons Capability.edu/∼nsarchiv/NSAEBB/NSAEBB6/index. 2005.org/view/generic/id/9041/title/Math Trek Mathematical FortuneTelling [7] Okasha S.). Vol. I.stanford. U.

232 S.nationmaster. download available at http://www. Rao http://www.org/ethics/scsEthicsCode.pdf [26] Encyclopedia entry Shannon’s Maxim.com/encyclopedia/Shannon’s-Maxim .scs.

Ming Hsieh Department of Electrical Engineering. DOI 10.edu. According to him [2]. The former is explicitly pointed out by Wallsten and Budescu [3] as “except in very special cases. CWW is “a methodology in which the objects of computation are words and propositions drawn from a natural language. 9. CWW engine and decoder — are briefly described. dongruiw@usc. all representations are vague to some degree in the minds of the originators and in the minds of the receivers. Ruan. Los Angeles. Computing with Words (CWW) is a methodology in which the objects of computation are words and propositions drawn from a natural language. University of Southern California. Mendel and Dongrui Wu Signal and Image Processing Institute. it may be more accurate and less subjective. Its three components — encoder. Computational Intelligence in Complex Decision Systems.Chapter 9 Computing with Words for Hierarchical and Distributed Decision-Making Jerry M.1 Introduction Zadeh coined the phrase “computing with words” (CWW) [1.” D. A Journal Publication Judgment Advisor (JPJA) is designed as an example to show how the journal paper review process can be automated by the Per-C. and moreover. called Perceptual Computer — Per-C for short — is introduced in this chapter.” and they suggest to model it by type-1 fuzzy sets (T1 FSs). Atlantis Computational Intelligence Systems 2.usc. The JPJA has the potential to relieve much of the burden of the reviewers and the associate editors. The latter is pointed out by Mendel [4] as “words mean different things to different people” and Wallsten and Budescu [3] as “different individuals use diverse expressions to describe identical situations and understand the same phrases differently when hearing or reading them. USA Email: mendel@sipi. The main focus is on how the Per-C can be used to assist in hierarchical and distributed decision-making.1007/978-94-91216-29-9_9. A specific architecture for subjective decision-making using CWW. © 2010 Atlantis Press/World Scientific 233 .” There are at least two types of uncertainties associated with a word [3]: intra-personal uncertainty and inter-personal uncertainty.edu According to Zadeh. 2]. CA 90089-2564.

Wu Because an interval type-2 FS [4] (IT2 FS) (see Appendix A. the encoder1 transforms linguistic perceptions into IT2 FSs that activate a CWW engine.M. it can model both types of uncertainty. 9. rank. The decoder2 maps the output of the CWW engine into a recommendation. 3HUFHSWXDO &RPSXWHU 3HU& 3HUFHSWLRQV :RUGV. 6]. 4.1. A specific architecture [7] shown in Fig.234 J. hence.1 is proposed for making subjective judgments by CWW. which can be a word. In Fig. or class. The CWW engine performs operations on the IT2 FSs. Mendel and D.1) can be viewed as a group of T1 FSs. It is called a perceptual computer—Per-C for short. 9. we suggest IT2 FSs be used in CWW [5.

(QFRGHU .7 )6V &:: (QJLQH 5HFRPPHQGDWLRQ ZRUG UDQN FODVV.

g. each of the word’s data intervals is individually mapped into its respective T1 interior. outliers) are eliminated. for each word in an applicationdependent encoding vocabulary. after which the union of all of these T1 MFs is taken. c 2002. 12] (LWA) (see Appendix A. he calls this precisiation. In [8. 2] calls this linguistic approximation. This can be done with Liu and Mendel’s Interval Approach [10] (IA).g.7 )6V Figure 9. First.e..2). during which some intervals (e. The words and their FOUs constitute a codebook. which maps IT2 FSs into IT2 FSs. To operate the Per-C. There are different kinds of CWW engines. Then. (2) How to construct the CWW engine. e.. left-shoulder or rightshoulder IT2 FS. the encoding problem. . i. what are the end-points of an interval that you associate with the ? word After some pre-processing.1 Conceptual structure of the Perceptual Computer (Mendel [7]. each of the remaining intervals is classified as either an interior. 2] calls this constraint explicitation. The result is a footprint of uncertainty (FOU) for an IT2 FS model of the word. one needs to solve the following problems: (1) How to transform words into IT2 FSs. which is 1 Zadeh 2 Zadeh [1. AMCS). (a) The linguistic weighted average [11. 9] and some of his recent talks. left-shoulder or right-shoulder MF.. a group of subjects are asked the following question: On a scale of 0-10. [1. 'HFRGHU .

the sub-criteria (e. Fjk ) j=1 p (4) in which sJ (X j . . Fjk ) = ∑N min(X j (xi ). . X p ). . 12] that the upper membership function (UMF) of YLWA is a fuzzy weighted average [13] of the UMFs of Xi and Wi . where X j ( j = 1. . .g. e. numbers. i.e. data. the weights. which considers the following problem: Given a rulebase with K rules. and Wi . and a new input X = (X1 . . F k (xi )) i=1 i=1 j ∑N max(X j (xi ). 15] (PR). the decoding problem. It has been shown [11.. (3) How to map the output of the CWW engine into a recommendation.. F j (xi )) + ∑N max(X j (xi ). sJ (X j . Then y is Gk (2) where F jk and Gk are words modeled by IT2 FSs. i. recommendations. X j (xi ) and X j (xi ) are upper and lower MFs of X j . i. or T1 FSs. scores. . Another approach that uses firing intervals instead of firing levels is described in [17]. and x p is Fp . decisions. intervals. (b) Perceptual reasoning [14.. each of the form: k k Rk : If x1 is F1 and .Computing with Words for Hierarchical and Distributed Decision-Making 235 defined as YLWA = ∑N XiWi i=1 ∑N Wi i=1 (1) where Xi . and the lower membership function (LMF) of YLWA is a fuzzy weighted average of the LMFs of Xi and Wi . (5) where. Fjk ) is the Jaccard similarity for IT2 FSs [16].g. f k (X ) = ∏ sJ (X j .. they can also be special cases of IT2 FSs. . are usually words modeled by IT2 FSs. F j (xi )) + ∑N min(X j (xi )..e. then what is the output IT2 FS YPR ? In PR [14] one computes YPR = ∑K f k (X )Gk k=1 ∑K f k (X ) k=1 (3) where f k (X ) is the firing level of Rk . judgments. etc). however. Thus far. features. e. there are three kinds of decoders according to three forms of recommendations: .g..e. . . F k (xi )) i=1 i=1 j k k . p) are also words modeled by IT2 FSs.

.e. but is based on aggregating independently made recommendations about an object from other individuals. By “hierarchical and distributed decision-making” is meant decision-making that is ultimately made by a single individual. Because the performance of each alternative is represented by an IT2 FS obtained from the CWW engine.” and it is the aggregation of the distributed recommendations at a higher level that leads to this being called “hierarchical. The Jaccard similarity measure [16] described by (5) can be used to compute the similarities between the CWW engine output and all words in the codebook. Then. a military objective. (b) Rank: Ranking is needed when several alternatives are compared to find the best. In this chapter our attention is directed at the hierarchical and distributed journal publication judgment advisor (JPJA) in which for the first time only words are used at every level. This application is representative of other distributed and hierarchical decision-making ap- . and words modeled by IT2 FSs. the individuals. etc. and so a weight has to be assigned to each of them when they are aggregated. Vlachos and Sergiadiss subsethood measure [18–20] is useful for this purpose.236 J.” There can be multiple levels of hierarchy in this process. because each of the independent recommendations may also involve a hierarchical decision-making process. The independent recommendations can involve aggregating numbers. it must be possible to compare the similarity between two IT2 FSs. T1 FSs. This chapter applies Per-C to a hierarchical and distributed decision-making process. intervals. (c) Class: A classifier is necessary when the output of the CWW engine needs to be mapped into a decision category. a ranking method for IT2 FSs is needed. An object could be a person being considered for a job. judges). Wu (a) Word: To map an IT2 FS into a word. Additionally.M. group or organization. the final decision class is the one corresponding to the maximum subsethood. groups or organizations (i. Mendel and D. A centroid-based ranking method [16] for IT2 FSs has been proposed. Then. an article being reviewed for publication in a journal. One first computes the subsethood of the CWW engine output for each of the possible classes. the word with the maximum similarity is chosen as the Decoder’s output. The final recommendation (or decision) is made by a decision maker who not only uses an aggregated recommendation that is made across all of the judges but may also use the aggregated recommendation from each of the judges. It is the independent nature of the recommendations that leads to this being called “distributed. groups or organizations making their independent recommendations may not be of equal expertise.

Technical Merit and Presentation. The distributed and hierarchical nature of the decision–making process is shown in Fig. Organization.” each reviewer usually has to complete a form similar to the one shown in Table 9. Technical Merit has three sub-criteria Importance. The rest of this chapter is organized as follows: Section 9.1 in which the reviewer has to evaluate the paper based on two major criteria.5 draws conclusions.g. Usually. Observe that each of the sub-criteria has an assessment level that is characterized by a starting word and an ending word (e. Background materials on IT2 FSs and the LWA are given in the Appendices.. The AE then makes a final decision based on the opinions of the three reviewers. so its results should be extendable to them. Clarity and References. Observe that there are three levels in the hierarchy: (1) Aggregation of the sub-criteria for the two major criteria. Finally.2 The Journal Publication Judgment Advisor (JPJA) 3 When an author submits a paper to a journal [11].4 illustrates the performance of the JPJA using several examples. the reviewer may give an evaluation based on the reputation of the author(s) or randomly choose an evaluation from several comparable evaluations. The Editor uses this publication-recommendation to assist in making a final decision about the paper.Computing with Words for Hierarchical and Distributed Decision-Making 237 plications.2 introduces the traditional journal paper review process. A similar situation 3 The material in Sections 9.3 describes how Per-C is used to construct the JPJA. and (3) aggregation of the three reviewers’ recommendations. 9. In addition to the “comments for the author(s).3 will also appear in [19]. Sometimes a reviewer may feel it is difficult to give an overall evaluation of a paper because it gets high assessment levels on some of the sub-criteria but does poorly on the others. The reviewers send their reviews back to the AE who then makes a publication-recommendation to the Editor based on these reviews. the Editor usually assigns its review to an Associate Editor (AE). and Presentation has four sub-criteria Style. (2) aggregation of the two major criteria. This is very subjective because no one knows what words are associated with the middle three boxes. Section 9.2. Valuable is the starting word for Importance and Useless is its ending word) and there are five boxes between them. who then sends it to at least three reviewers. 9. In that case. Content and Depth. Section 9. . the reviewer is also asked to give an overall evaluation of the paper and make a recommendation to the AE. Section 9.2 and 9. A reviewer chooses one assessment level by checking-off one of the five boxes.

what should the final decision be? Table 9.1 Paper review form for a generic journal. each reviewer should only be asked to provide a subjective evaluation of a paper for each of the seven sub-criteria. . e.. it may be better to leave it to a computer.. Because this evaluation process is often difficult and subjective. i. Wu may occur at the AE level. IEEE). moreover. another suggests a revision of the paper. Once the opinion of all the reviewers are obtained. and a third reviewer suggests acceptance of the paper.e.238 J. Mendel and D.M. if one reviewer suggests rejection of the paper.2 The paper review process (Wu and Mendel [11].g. Technical Merit: Importance Content Depth Presentation: Style Organization Clarity Reference Overall: Overall Evaluation Recommendation: ( ) ( ) ( ) ( ) Valuable Original Deep Readable Precise Clear Complete Excellent ( ) ( ) ( ) ( ( ( ( ) ) ) ) ( ) ( ) ( ) ( ( ( ( ) ) ) ) ( ) ( ) ( ) ( ( ( ( ) ) ) ) ( ) ( ) ( ) ( ( ( ( ) ) ) ) ( ) ( ) ( ) ( ( ( ( ) ) ) ) Useless Derivative Shallow Incoherent Ambiguous Confusing Incomplete Dreadful ( ) ( ) ( ) ( ) ( ) ACCEPT: as written (editorial corrections as noted are necessary) REWRITE: needs substantial changes and re-review REJECT: materials is unsuitable for publication in this journal TRANSFER: more suitable for publication in $VVRFLDWH (GLWRU 5HYLHZHU  5HYLHZHU  5HYLHZHU  7HFKQLFDO 0HULW &RQWHQW . c 2007. and. another LWA would compute an overall aggregated opinion for the AE. after which LWAs would automatically compute the reviewer’s overall judgment of the paper. This automatic process has the potential to relieve much of the burden of the reviewers and the AE. it may be more accurate and less subjective.PSRUWDQFH 'HSWK 2UJDQL]DWLRQ 5HIHUHQFHV Figure 9.PSRUWDQFH 'HSWK 3UHVHQWDWLRQ 6W\OH &ODULW\ 7HFKQLFDO 0HULW 3UHVHQWDWLRQ 7HFKQLFDO 0HULW 3UHVHQWDWLRQ 6W\OH &ODULW\ 2UJDQL]DWLRQ 5HIHUHQFHV &RQWHQW .

3. 9. Adequate. (3) A reviewer is no longer asked to provide a recommendation.2 Modified paper review form for a generic journal. Marginal. . and that asks each reviewer to indicate their level of expertise for reviewing the paper. fewer or more than five words could be used.1 and 9. and different words could be used for each sub-criterion.2.Computing with Words for Hierarchical and Distributed Decision-Making 239 9.2. a modified version of the Table 9.3 Perceptual Computing for the JPJA This section explains how a Per-C can be used as a JPJA. CWW engine and decoder—must be considered. Good and Excellent. Of course. While it is doubtful that fewer than three words should be used. The modified paper review form (for a generic journal) is depicted in Table 9. Table 9. observe that: (1) The same words are now used for all seven sub-criteria. When perceptual computing is used for the JPJA each of the three components of a Per-C— encoder.1 Modified Paper Review Form To begin.1 paper review form is needed. Moderate and High. it is possible that more than three words could be used. (2) Three words are used for a reviewer’s level of expertise: Low. Poor Marginal Adequate Good Excellent Technical Merit: Importance ( ) ( ) ( ) ( ) ( ) Content ( ) ( ) ( ) ( ) ( ) Depth ( ) ( ) ( ) ( ) ( ) Presentation: Style ( ) ( ) ( ) ( ) ( ) Organization ( ) ( ) ( ) ( ) ( ) Clarity ( ) ( ) ( ) ( ) ( ) Reference ( ) ( ) ( ) ( ) ( ) Expertise: Your Expertise Low ( ) Moderate ( ) High ( ) Comparing the review forms in Tables 9. one that removes the uncertainty about the words that are associated with the three middle boxes for each sub-criterion. namely: Poor. These five words are linguistically appropriate for each of the seven sub-criteria.

g. The interval corresponds to the importance weighting that you assign to each criterion.3. Adequate.5 0.2. Mendel and D. • Sub-codebook R1: This codebook contains the five words (Poor. and each of these codebooks has sub-codebooks.3.4 FOUs for the three-word Sub-codebook R2. FOUs for these words are depicted in Fig. Good and Excellent) and their FOUs that are used to assess the seven sub-criteria. 9.1 Codebook words used by a reviewer The reviewer codebook has two sub-codebooks. all of the results that are in this chapter are synthetic.75 0. 4 Intervals could be collected from a group of subjects by stating: You are to assign an interval or a “range” of numbers that falls somewhere between 0–10 to each of following three criteria that are associated with judging the Technical Merit of a journal article submission: Content. u Poor Marginal Adequate Good Excellent 1 0. Marginal.3.240 J. Wu 9. AEs and reviewers.75 0. They were not obtained by collecting data from a group of subjects4 .25 0 x 0 1 2 3 4 5 6 7 8 9 10 Figure 9. Importance and Depth.. each of which is described next.2 Encoder Two codebooks are needed. a pool of knowledgeable AEs and reviewers should be surveyed and then their data intervals should be mapped into word FOUs using the IA [10]. as explained next. but instead were generated by the authors. 9. one for the words that will be used by a reviewer and the other for weights.25 0 Low Moderate High w 0 1 2 3 4 5 6 7 8 9 10 Figure 9. It is important to note that not all ranges have to be the same and ranges can overlap. e. u 1 0. Subjects are then asked a question like: “Where on a scale of 0–10 would you locate the ends points of an interval that you assign to weight the importance to Technical Merit of Content?” .3 FOUs for the five-word Sub-codebook R1.M.5 0. Consequently. In order to get more accurate results.

9. Content (WCo ) and Depth (WD ). whereas for the weight sub-codebooks each FOU is associated with a weight that has no word associated with it. with the help of Dr. Again. That ordering is: Content (Co) is more important than Importance (I) which in turn is more important that Depth (D). they are assigned symbols rather than words.75 0. WCo > WI > WD . Weights correspond to Importance (WI ). • Sub-codebook W1: This codebook contains labels and FOUs for the relative weighting of the three sub-criteria that are associated with the criterion of Technical Merit. 9. the relative orderings of the three sub-criteria were first established (in 2007) with the help of Dr. 9. Nihil Pal (Editor-in-Chief of the IEEE Trans. . They also were not obtained by collecting data from a group of subjects. FOUs for these weights are depicted in Fig. however. These FOUs were also generated by the authors.25 0 w 0 1 2 3 4 5 6 7 8 9 10 Figure 9.5 FOUs for the three-word Sub-codebook W1. hence. WS = WCl > WO = WR . and that Style and Clarity are more important than Organization and References.3. u W D WI WC o 1 0. so the admonition just given for the FOUs in Sub-codebook R1 applies here as well.2. 9. hence. it was decided that Organization (O) and References (R) were indistinguishable as were Style (S) and Clarity (C).5. Moderate and High) and their FOUs that describe a reviewer’s level of expertise.6. The weight FOUs are only used in LWAs and are not available to a reviewer. A major difference between the weights sub-codebooks and the two reviewer sub-codebooks is that for the latter each FOU has a word associated with it. hence.Computing with Words for Hierarchical and Distributed Decision-Making 241 • Sub-codebook R2: This codebook contains the three words (Low. FOUs for these words are depicted in Fig.4. on Fuzzy Systems).5 0. but instead were generated by the authors. FOUs for these weights are depicted in Fig. • Sub-codebook W2: This codebook contains labels and FOUs for the relative weighting of the four sub-criteria that are associated with the criterion of Presentation. Nihil Pal.2 Codebook for the weights This weights codebook has three sub-codebooks each of which is described next.

6 FOUs for the four-word Sub-codebook W2. do not change the methodology of the JPJA that is further explained below. W2 and W3.7. • Sub-codebook W3: This codebook contains labels and FOUs for the relative weighting of the two criteria Technical Merit and Presentation.5 0. i. 9. Wu u 1 0. 9. i. Again.3. Weights correspond to Style (WS ).3 CWW Engine For the CWW Engine of the JPJA.242 J.5–9. then the shapes of the FOUs as well as the amount that they overlap will be different from the ones in Figs. Clarity (WCl ) and Reference (WR ). 2. u WP WT 1 0.25 0 w 0 1 2 3 4 5 6 7 8 9 10 Figure 9. nR ): Y jT = Y jP = X jI WI + X jCoWCo + X jDWD WI + WCo + WD X jSWS + X jOWO + X jClWCl + X jRWR WS + WO + WCl + WR (6) (7) . Organization (WO ). Those details. Mendel and D. If interval data are collected from AEs and reviewers for sub-codebooks W1.5 0.e. Clearly Technical Merit (T) is more important than Presentation (P). 9.e. Weights correspond to Technical Merit (WT ) and Presentation (WP ). each of the two major criteria (Technical Merit and Presentation) has an LWA computed for it.12] (see Appendix A.. To begin.7. .25 0 WO = WR WS = WC l w 0 1 2 3 4 5 6 7 8 9 10 Figure 9. with the help of Dr. LWAs [11. . .75 0.2) are used because one of the unique features of this application is that all assessments and weights are words (or FOUs).M. WT > WP . while important for the application of the JPJA to a specific journal.7 FOUs for the two-word Sub-codebook W3.75 0. ( j = 1. . it was decided that the FOUs for these weights should be located as depicted in Fig. and the FOUs are determined using the IA.. Nihil Pal.

or Reject. Y jT is the LWA for Technical Merit in which X jI . Rewrite or Reject. the nR Reviewer LWAs are aggregated using the following AE LWA: YAE = in which WR j ∈ {Low. .. X jCl and X jR are the linguistic assessments provided by Reviewer. 9.Computing with Words for Hierarchical and Distributed Decision-Making 243 where nR is the number of reviewers5 .4. AEs are surveyed using a statement and question like: A reviewer of a journal submission must score the submission using a number between 0–10 where 0 is the lowest score and 10 is the highest score.4. it classifies the overall quality of a paper. The reviewer’s score must then be mapped into one of three classes—Accept. Rewrite and Reject. X jCo and X jD are the linguistic assessments provided by Reviewer-j for Importance.1 Construct the decoding codebook using a survey In this approach. nR ): YR j = Y jT WT + Y jPWP WT + WP (8) Finally. Organization. into one of three classes: Accept. 9.j for Style. A decoding codebook is needed to store the FOUs for these three words. and the final 5 Although Fig. Moderate. . the AE’s FOU for a specific journal submission. 9.3.e. YR j ( j = 1. Rewrite. respectively. Y jT and Y jP are aggregated using the following Reviewer LWAs. High}. Content and Depth.3. Reject)? R ∑ j=1 YR jWR j R ∑ j=1 WR j n n (9) The IA can then be used to map the crisp intervals into IT2 FS FOUs for Accept. i. X jO . . . each of which consists of nR reviewers’ completed review forms.4 Decoder The decoder for the JPJA is actually a classifier. YAE .2 shows three reviewers. 2.2 Construct the decoding codebook using training examples A training data set is a collection of training examples. Where on the scale of 0–10 would you locate the ends points of an interval that you assign to Accept (Rewrite. . Two approaches for constructing such a codebook are described next. Next.3. Clarity and References. respectively. and Y jP is the LWA for Presentation in which X jS . the number of reviewers is treated here as a variable. 9.

because the generic review form in Table 9.5 0.8 but their specific parameters that completely define each FOU are not fixed ahead of time. and Reject are specified (or updated using an optimization algorithm). Instead. The Estimated Recommendation for each training example is obtained as follows: (1) FOU parameters for Accept. Its FOUs were synthesized with the help of Dr. X jO . u Reject Rewrite Accept 1 0.3. Final Recommendation(k) Instead of pre-specifying the FOUs for Accept. ss(YAE . NT ) in the training data set. .244 J.4 Decoder Vlachos and Sergiadis’s subsethoods [18–20] are computed between YAE and Accept. YAE . X jR . and Reject are computed. the decoding codebook used in this chapter is shown in Fig. Re ject). because the AE may want to make some subjective adjustments to the publication class when two subsethoods are very close.8 FOUs for the three-classes decoding codebook. . Accept). X jD . a training example is k k k k k k k k k (X jI . Mendel and D. 9. namely ss(YAE . j = 1. . for the kth submission (k = 1.8. Rewrite) and ss(YAE . 9. 9. So.4. no training examples are available. Rewrite and Reject. nR . . Wu recommendation made for the paper by the Editor-in-Chief (Accept. Rewrite or Reject).4. The decoder also provides YAE and the values of the three subsethoods to the AE.M. only the approach described in Section 9. however. Rewrite and Reject. . X jCo . X jCl . Rewrite.75 0. WR j ). for the present. Nihil Pal. the FOU parameters are tuned using search algorithms to minimize the errors between the given Final Recommendation and an Estimated Recommendation for all NT elements in the training data set.3. . X jS . . . and (3) the paper is assigned to the class with the maximum subsethood. Rewrite.1 can be used for constructing the decoding codebook. their shapes are chosen like the ones in Fig. i. (2) Vlachos and Serk giadiss subsethoods of YAE in Accept.e. Because we did not have access to a pool of AEs and reviewers. 9.3.4.2 has yet to be used. .3 Remarks In principle either approach could be used.25 0 y 0 1 2 3 4 5 6 7 8 9 10 Figure 9. after which the decoder recommends to the AE the publication class of maximum subsethood..

XCo and XD .5. In all cases.2 Importance is Poor.1 Aggregation of Technical Merit Sub-criteria To consider all possible combinations of the inputs to the criterion of Technical Merit. Each of the left-hand figures in Fig. and Content varies from Poor to Excellent The FOUs of YT for these five cases are depicted in the right-hand figures of Fig. 9. XCo and XD are varied. Depth is Good. and only XI .4 Examples6 This section contains some examples that illustrate the automatic paper review process. Note that XI = Excellent and XD = Good are the ones in Fig.3) that can be chosen by a reviewer. 9. 9. 6 The examples in this section are similar to those in [19]. 9. . Importance is changed from Excellent to Poor. since in (6) and Sub-codebook W1 it has already been established that WCo > WI > WD . 9. and then three complete paper reviews are provided. To do this 15 of the 125 possible cases are studied.Computing with Words for Hierarchical and Distributed Decision-Making 245 9.4.9. If Content dominates the other sub-criteria of Importance and Depth. 9.1. 9. This is impractical for us to do.10 depicts the respective FOUs for XI .1 Importance is Excellent.4. Depth and Content in Fig. 9. 9.1.10. Observe that as the reviewer’s response to Content varies from Poor to Excellent. Each of the left-hand figures in Fig. then regardless of the words chosen for Importance and Depth. 9. 9. Only XCo changes in the left-hand figures. so instead our focus is on whether or not Content dominates the other subcriteria of Importance and Depth as it should. First the different levels of aggregation are examined to demonstrate that reasonable results are obtained.9. XCo and XD .3. one expects to see YT move from left-to-right as XCo moves from left-to-right. The FOUs of YT for the present five cases are depicted in the right-hand figures of Fig. and Content varies from Poor to Excellent Starting with the choices that were made for Importance. Depth is Good. and everything else is kept the same. one would need to examine a total of 53 = 125 cases. YT in (6) is computed for the three weight FOUs that are depicted in Fig. because each of its three sub-criteria has five possible linguistic terms (Fig.4. YT does move from left-to-right towards the maximal score of 10 which supports our expectation that Content dominates.3 for all five words. 9. and its FOUs match the ones in Fig.9 depicts the respective FOUs for XI .

246 J.75 0.25 0 YT y 0 1 2 3 4 5 6 7 8 9 10 (i) (j) Figure 9.75 0. which again supports our expectation that Content dominates. 1 0. 9.3.5 0.25 0 YT y 0 1 2 3 4 5 6 7 8 9 10 (e) u (f) XD = XC o XI u 1 0.5 0.75 0.5 0.25 0 YT y 0 1 2 3 4 5 6 7 8 9 10 (c) u (d) XD XI u 1 0. .25 0 YT y 0 1 2 3 4 5 6 7 8 9 10 (a) u (b) XD XI u 1 0.25 0 XC o x 0 1 2 3 4 5 6 7 8 9 10 1 0.5 0. (j): Output YT when a reviewer’s response to Content changes from Poor to Excellent.25 0 uX Co XD XI u x 0 1 2 3 4 5 6 7 8 9 10 1 0. As in Fig.25 0 YT y 0 1 2 3 4 5 6 7 8 9 10 (g) u (h) XD XC o = XI u 1 0.3 for all five words.5 0.9.75 0. and its FOUs match the ones in Fig. Mendel and D.5.9 Technical Merit when Importance is Excellent. (b). 9. (i): Input FOUs when a reviewer’s response to Content changes from Poor to Excellent.5 0.5 0. (g). (h).5 0.75 0.75 0.5 0.75 0.25 0 x 0 1 2 3 4 5 6 7 8 9 10 1 0. (c).25 0 XC o x 0 1 2 3 4 5 6 7 8 9 10 1 0. only XCo changes in the left-hand figures. Depth is Good and Content varies. YT moves from left-to-right towards the maximal score of 10 (but never reaching it). (e).75 0.5 0.25 0 x 0 1 2 3 4 5 6 7 8 9 10 1 0.75 0. (f). 9. The corresponding weights for the three sub-criteria are shown in Fig.M. 9. Wu Note that XI = Poor and XD = Good are the ones in Fig. (d). Observe that as the reviewer’s response to the sub-criterion Content varies from Poor to Excellent. (a).75 0.

9. observe that each of the FOUs of YT for when Importance is Excellent lies farther to the right of its respective FOU for when Importance is Poor.75 0.75 0.25 0 uX I (h) XD XC o u x 0 1 2 3 4 5 6 7 8 9 10 1 0.75 0.10. (h).25 0 uX I (f) XC o = XD u x 0 1 2 3 4 5 6 7 8 9 10 1 0.5 0.75 0.75 0.5 0. This also agrees with our expectation.25 0 YT y 0 1 2 3 4 5 6 7 8 9 10 (i) (j) Figure 9.5 0.25 0 uX I (b) XD u XC o x 0 1 2 3 4 5 6 7 8 9 10 1 0. .9 and 9.75 0.25 0 YT y 0 1 2 3 4 5 6 7 8 9 10 (c) 1 0.75 0. because improving the assessment of Importance should improve YT .10 Technical Merit when Importance is Poor.5 0. move YT further to the right.5 0.25 0 YT y 0 1 2 3 4 5 6 7 8 9 10 (g) 1 0.5 0.Computing with Words for Hierarchical and Distributed Decision-Making 247 1 0.5. Depth is Good and Content varies.25 0 YT y 0 1 2 3 4 5 6 7 8 9 10 (a) 1 0.5 0.5 0. i.25 0 YT y 0 1 2 3 4 5 6 7 8 9 10 (e) 1 0. (c).75 0. 9. (j): Output YT when a reviewer’s response to Content changes from Poor to Excellent. (g).5 0. By comparing each line of the two sets of right-hand figures in Figs. (d).75 0.75 0.25 0 uX I (d) XD u XC o x 0 1 2 3 4 5 6 7 8 9 10 1 0. (i): Input FOUs when a reviewer’s response to Content changes from Poor to Excellent. (e).5 0..e. (b).25 0 uX = X I Co XD u x 0 1 2 3 4 5 6 7 8 9 10 1 0. (f). (a). The corresponding weights for the three sub-criteria are shown in Fig.

(e).5 0.75 0.25 0 uX = X I D XC o u (d) 1 0.11 depicts the respective FOUs for XI . Depth and Content in Fig.3 Importance is Poor. and everything else is kept the same.75 0. Depth is changed from Good to Poor. Each of the left-hand figures in Fig.75 0.1. The FOUs of YT for the present five cases are depicted in the right-hand figures of Fig.75 0.75 0.25 0 uX = X I D XC o u (f) 1 0.5 0. XCo and XD .5.11 Technical Merit when both Importance and Depth are Poor and Content varies.25 0 YT x y 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 (g) 1 0. 9. Depth is Poor.75 0.4. Wu 1 0. (h). The corresponding weights for the three sub-criteria are shown in Fig.75 0.5 0. and Content varies from Poor to Excellent Starting with the choices that were made for Importance.25 0 uX = X = X I Co D x 0 1 2 3 4 5 6 7 8 9 10 1 0.75 0. (b). 9.5 0.5 0.M.10.25 0 uX = X I D XC o u (h) 1 0.5 0. (a). (c).75 0.5 0. (d). (g). (f).25 0 YT x y 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 (e) 1 0.25 0 uX = X X I D Co u (b) 1 0. 9.5 0.248 J. (i): Input FOUs when a reviewer’s response to Content changes from Poor to Excellent.25 0 uY T y 0 1 2 3 4 5 6 7 8 9 10 (a) 1 0.25 0 YT x y 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 (c) 1 0. 9.25 0 YT x y 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 (i) (j) Figure 9.75 0. Note that .5 0. (j): Output YT when a reviewer’s response to Content changes from Poor to Excellent.5 0.11. Mendel and D. 9.

9.10(j) and 9. Finally. 9. because each of its four sub-criteria has five possible linguistic terms (Fig.3. These deteriorations (which correspond to poorer reviews within the criterion of Technical Merit) are evident in all respective right-hand figure FOUs in Figs.11(j) the very left portion of YT only reaches y = 3 and y ≈ 2.4.11 represent an even greater deterioration of those cases. it never reaches). which again supports our expectation that Content dominates.3) that can be chosen by a reviewer.4 Conclusions These 15 cases have demonstrated that Content does indeed dominate Importance and Depth and that a higher assessment for any of the three sub-criteria will move YT to the right. 9. 9. one expects . Observe again that as the reviewer’s response to the sub-criterion Content varies from Poor to Excellent. one observes that in Fig. In all cases.3 for all five words. and its FOUs again match the ones in Fig. comparing Fig. This also agrees with our expectation.10 represent a deterioration of the respective cases in Fig. only XCo changes in the left-hand figures. 9. (j) in these figures. observe that each of the FOUs of YT for when Depth is Good lies farther to the right of its respective FOU for when Depth is Poor.11 by their moving towards the left.9–9.. i. 9. because improving the assessment of Depth should improve YT .4.9(j) the very left portion of YT reaches y = 6.6). but in Figs. observe that the five cases in Fig. so instead our focus is on observing the effect that sub-criterion Clarity has on YP when Style is fixed at Excellent and Organization and References are varied by the same amounts (recall that the weights for both Organization and References are the same – see Fig. To do this 15 of the 625 possible cases are studied.10. 9. Because WS = WCl . and the five cases in Fig. 9.e. As in Figs.6.4. This is even more impractical to do than it was to examine all 125 possible cases for Technical Merit.2 Aggregation of Presentation Sub-criteria To consider all possible combinations of the inputs to the criterion of Presentation.9. 9. 9. again.Computing with Words for Hierarchical and Distributed Decision-Making 249 XI = XD = Poor is the one in Fig.11. This all seems quite sensible in that one would expect poorer assessments of any of the three sub-criteria to move YT to the left. 9. 9. 9. 9. By comparing each line of the two sets of right-hand figures in Figs. 9. For example.10 and 9.1. one would need to examine a total of 54 = 625 cases. YP in (7) is computed for the four weight FOUs that are depicted in Fig. move YT further to the right.9 and 9. YT moves from left-to-right towards the maximal score of 10 (which.

3.3. YP does move from left-to-right towards the maximal score of 10 which supports our expectation that YP moves from left-to-right as XCl moves from left-to-right. as in Fig. 9. 9.3 for all five words.2 Style is Excellent.12 depicts the respective FOUs for XCl . Organization and References are Poor. 9. 9.3 Style. XS = Excellent but now XO = XR = Adequate and these FOUs are the ones in Fig.14. XR and XS .1 Style is Excellent. 9. Again.250 J. and its FOUs match the ones in Fig.12. 9. Each of the left-hand figures in Fig. 9. 9. and its . XO . XO . 9.14 depicts the respective FOUs for XCl . 9.M. and its FOUs match the ones in Fig. Observe that as the reviewer’s response to Clarity varies from Poor to Excellent.13 depicts the respective FOUs for XCl .2.13. This also agrees with our expectation. 9. Organization and References are Adequate and Clarity varies from Poor to Excellent The FOUs of YT for these five cases are depicted in the right-hand figures of Fig. Note that XS = Excellent and XO = XR = Poor are the ones in Fig. Wu to see YP move from left-to-right as XS moves from left-to-right and also as XO and XR simultaneously move from left-to-right.2. only XCl changes in the left-hand figures. 9. and Clarity varies from Poor to Excellent The FOUs of YT for these five cases are depicted in the right-hand figures of Fig. Only XCl changes in the left-hand figures. 9.2. 9. 9. Each of the left-hand figures in Fig. Organization and References are Excellent and Clarity varies from Poor to Excellent The FOUs of YT for these five cases are depicted in the right-hand figures of Fig. only XCl changes in the left-hand figures. Mendel and D. Each of the left-hand figures in Fig..12 and 9.4. Again.12 and 9. observe that each of the FOUs of YP for when Organization and References are Adequate are somewhat to the right of its respective FOU for when Organization and References are Poor. XS = Excellent but now XO = XR = Excellent where Excellent is depicted in Fig.4. 9.13. because improving the assessments of Organization and References should improve YP .13. Note that.3 for all five words. XR and XS . YP does move from left-to-right towards the maximal score of 10 which again supports our expectation that YP moves from left-to-right as XCl moves from left-to-right. i. Note that. as in Figs.e. XO . 9. By comparing each line of the two sets of right-hand figures in Figs.12. XR and XS . Observe that as the reviewer’s response to Clarity varies from Poor to Excellent.3. move YP further to the right.4.

9. 9.e. Finally. because improving the assessments of Organization and References should improve YP . 9.13(j) the left portion of YP begins at y ≈ 4.3 for all five words.12. comparing Fig.13 represent an improvement of the respective cases in Fig. observe that each of the FOUs of YP for when Organization and References are Excellent are somewhat to the right of its respective FOU for when Organization and References are Adequate.13 and 9. because improving the assessments of Organization and References should improve YP . move YP further to the right. 9..Computing with Words for Hierarchical and Distributed Decision-Making 251 FOUs match the ones in Fig. This also continues to agree with our expectation. These improvements (which correspond to better reviews within the criterion of Presentation) are evident in all respective right-hand figure FOUs in Figs.4 Conclusions These 15 cases have demonstrated that YP moves from left-to-right as XCl moves from left-to-right and also as XO and XR simultaneously move from left-to-right.12-9. observe that each of the FOUs of YP for when Organization and References are Excellent are somewhat to the right of its respective FOU for when Organization and References are Adequate. one observes that in Fig. (j) in these figures. and.3 and the very right portion of YP reaches y = 10. 9. 9. For example. move YP further to the right. 9. Observe that as the reviewer’s response to Clarity varies from Poor to Excellent. i.5 and not only does the very right portion of YP also reach y = 10. 9. Observe that as the reviewer’s response to Clarity varies from Poor to Excellent.e. By comparing each line of the two sets of right-hand figures in Figs.14 represent an even greater improvement of those cases.14(j) the left portion of YP begins at y ≈ 6. YP once again moves from left-to-right towards the maximal score of 10 which continues to support our expectation that YP moves from left-to-right as XCl moves from left-to-right. and the five cases in Fig. in Fig. This also continues to agree with our expectation. .14.. 9.5 and the very right portion of YP also reaches y = 10.2. 9.13 and 9.12(j) the left portion of YP begins at y ≈ 3. YP once again moves from left-to-right towards the maximal score of 10 which continues to support our expectation that YP moves from left-to-right as XCl moves from left-to-right.14. i. in Fig.14 by their moving towards the right. This all seems quite sensible in that one would expect higher assessments of any of the three sub-criteria to move YP to the right. 9. but YP has changed its shape from that of an interior FOU to that of a right-shoulder FOU.4. observe that the five cases in Fig. By comparing each line of the two sets of right-hand figures in Figs.

3 Aggregation at the Reviewer Level This section assumes that there are three reviewers for a paper submission.25 0 YP x y 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 (g) 1 0.25 0 uX = X O R XC l XS u (f) 1 0.5 0. leading to Y jT and Y jP ( j = 1.25 0 YP y 0 1 2 3 4 5 6 7 8 9 10 (a) 1 0.5 0.25 0 YP x y 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 (e) 1 0.75 0. (j): Output YP when a reviewer’s response to Clarity changes from Poor to Excellent. (i): Input FOUs when a reviewer’s response to Clarity changes from Poor to Excellent.5 0. Wu 1 0.25 0 YP x y 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 (c) 1 0. Mendel and D.75 0. (b). whose shapes in this section are pre-specified by us7 .75 0.75 0.25 0 uX = X X O R Cl XS u (b) 1 0. (h).5 they will be determined by beginning with three completed generic review forms.75 0.75 0. 9.4.25 0 uX = X O R XC l XS u (d) 1 0.5 0. (f).5 0.M. Y jT and Y jP .5 0.25 0 uX = X = X O R Cl XS u x 0 1 2 3 4 5 6 7 8 9 10 1 0.12 Presentation when Style is Excellent. are then aggregated using the LWA in (8) in which Y jT and Y jP are weighted using their respec7 In Section 9. 2. and Clarity varies. 3).252 J. (e).25 0 uX = X O R XS = XC l u (h) 1 0.75 0. The corresponding weights for the three sub-criteria are shown in Fig. (g). (c).5 0. .75 0.4.5 0. Organization and References are Poor.75 0. (d).25 0 YP x y 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 (i) (j) Figure 9.6.5 0. 9.75 0.5 0. (a). and that Technical Merit and Presentation have already had their sub-categories aggregated.

4. as previously agreed upon. Adequate and Excellent are in Fig.1 Presentation is Adequate and Technical Merit Varies In this case all of the reviewers agree that Presentation is Adequate.2 Technical Merit is Adequate and Presentation Varies In this case all of the reviewers agree that Technical Merit is Adequate. Observe that these FOUs are spread out indicating substantial disagreement among the reviewers. Here the focus is on . Fig. So..3. Reviewer 2 rates Presentation as Adequate. hence.3. 9. hence. 9.4. 9. 3. and that as the rating for Technical Merit improves YR j moves to the right. j = 1. In those figures Y jP is fixed at Good and only Y jT varies from one figure to the next. where Poor. and Reviewer 3 rates Presentation as Excellent. 9. this effect on YR j is nowhere as dramatic as the effect of an improved rating for Technical Merit has on YR j . 9.3.4 Aggregation at the AE Level This section builds upon the results in Section 9. In those figures.Computing with Words for Hierarchical and Distributed Decision-Making 253 tive weight FOUs that are depicted in Fig. YR3 > YR2 > YR1 . Reviewer 2 rates Technical Merit as Adequate and Reviewer 3 rates Technical Merit as Excellent. Reviewer 1 rates Technical Merit as Poor. 9. 2. even though there is considerable disagreement among the reviewers about Presentation. Y jT is fixed at Good and only Y jP varies from one figure to the next.16(d) depicts YR1 .” 9.15(d). but they disagree about Presentation. YR2 and YR3 . where > is used here to denote “further to the right of. 9. YR2 and YR3 . i. Fig. While it is true that an improved rating for Presentation moves YR j further to the right.4. 9. heavier weight is given to Technical Merit than to Presentation. respectively. respectively.e.3. The results are YR j . 9.7.1. 9.15(d).4. 9. this case is the opposite of the one in Section 9. this does not make as much difference in YR j as did the disagreement about Technical Merit in Fig.3. Reviewer 1 now rates Presentation as Poor. Y jT and Y jP are depicted in Figs. a higher rating for Technical Merit has a much greater effect on YR j than does a comparable rating for Presentation.16(a)–(c) for Reviewers 1-3.3 Conclusions As expected.15(a)–(c) for Reviewers 1-3.4.15(d) depicts YR1 . Y jT and Y jP are depicted in Figs.3.4. Observe that these FOUs are much more bunched together and have considerably more overlap than the ones in Fig. it assumes there are three reviewers of a submission and their reviews have been aggregated using (8). where. but they disagree about Technical Merit.

(d).75 0.5 0. High > Moderate > Low. The corresponding weights for the three sub-criteria are shown in Fig. . Organization and References are Adequate.13 Presentation when Style is Excellent.25 0 YP y 0 1 2 3 4 5 6 7 8 9 10 (a) u XC l XO = XR XS u (b) YP 1 0.75 0.5 0.4. 9. (i): Input FOUs when a reviewer’s response to Clarity changes from Poor to Excellent.5 0. (h).5 0.25 0 x 0 1 2 3 4 5 6 7 8 9 10 1 0.75 0.25 0 XO = XR x 0 1 2 3 4 5 6 7 8 9 10 1 0.254 J. (f). and Clarity varies. because according to Fig. One expects that reviews from reviewers who have high expertise will have more effect on YAE than those from reviewers who have moderate or low levels of expertise.5 0.25 0 y 0 1 2 3 4 5 6 7 8 9 10 (g) u (h) XS = XC l u 1 0. Wu what effects different levels of reviewer expertise have on YAE as computed by (9). (a). (e).5 0. (b). (j): Output YP when a reviewer’s response to Clarity changes from Poor to Excellent.25 0 y 0 1 2 3 4 5 6 7 8 9 10 (c) u XO = XR = XC l XS u (d) YP 1 0. (g).75 0.75 0.” 1 0.M.25 0 x 0 1 2 3 4 5 6 7 8 9 10 1 0. Mendel and D. 9.25 0 uX Cl XO = XR XS u x 0 1 2 3 4 5 6 7 8 9 10 1 0.25 0 x 0 1 2 3 4 5 6 7 8 9 10 1 0.5 0.75 0.75 0. (c).5 0.5 0.25 0 y 0 1 2 3 4 5 6 7 8 9 10 (e) u XO = XR XC l XS u (f) YP 1 0.75 0.5 0.25 0 YP y 0 1 2 3 4 5 6 7 8 9 10 (i) (j) Figure 9. where again > is used to denote “further to the right of.75 0.75 0.6.

75 0.4.25 0 x 0 1 2 3 4 5 6 7 8 9 10 1 0. Technical Merit Varies.75 0.5 0.25 0 YP y 0 1 2 3 4 5 6 7 8 9 10 (c) u (d) XS = XO = XR u 1 0.5 0.Computing with Words for Hierarchical and Distributed Decision-Making 255 1 0.5 0.25 0 YP y 0 1 2 3 4 5 6 7 8 9 10 (i) (j) Figure 9.6. their weights are depicted in Fig.14 Presentation when Style.5 0.75 0. Organization and References are Excellent. and (9) leads to YAE that is depicted in Fig. When Reviewer 1 is of high expertise.3.25 0 XC l x 0 1 2 3 4 5 6 7 8 9 10 1 0.75 0.5 0.5 0.4.15(d) and that are quite spread out. (b). (f). (h). and Reviewer Expertise Varies This is a continuation of Section 9.17(a). (j): Output YP when a reviewer’s response to Clarity changes from Poor to Excellent.5 0. 9. 9.1 Presentation is Adequate.17(b).25 0 uX Cl XS = XO = XR u x 0 1 2 3 4 5 6 7 8 9 10 1 0. (c).5 0.25 0 YP y 0 1 2 3 4 5 6 7 8 9 10 (a) u (b) XS = XO = XR u 1 0.5 0. when .75 0. (e). (d). (i): Input FOUs when a reviewer’s response to Clarity changes from Poor to Excellent. The corresponding weights for the three sub-criteria are shown in Fig. (a). YR2 and YR3 that are in Fig. Reviewer 2 is of moderate expertise and Reviewer 3 is of low expertise. 9.25 0 x 0 1 2 3 4 5 6 7 8 9 10 1 0. On the other hand. 9.25 0 YP y 0 1 2 3 4 5 6 7 8 9 10 (g) u (h) XS = XO = XC l = XR u 1 0.1.75 0.75 0.25 0 XC l x 0 1 2 3 4 5 6 7 8 9 10 1 0.75 0. and Clarity varies. (g). 9.75 0.75 0.5 0.25 0 YP y 0 1 2 3 4 5 6 7 8 9 10 (e) u (f) XC l XS = XO = XR u 1 0.4. and starts with YR1 .

25 0 uY 1P Y 1T u 0 1 2 3 4 5 6 7 8 9 10 1 0.75 0. respectively. 9. What can be concluded from all of this? First.7.75 0.25 y 0 YR 1 YR 2 YR 3 y 0 1 2 3 4 5 6 7 8 9 10 (c) (d) Figure 9. their weights are depicted in Fig.75 0.25 0 Y 3P 0 1 2 3 4 5 6 7 8 9 10 1 0. 2 and 3’s overall FOU.75 0. (a).5 0.25 0 Y 3T 0 1 2 3 4 5 6 7 8 9 10 1 0.75 0. 9.75 0.5 0.16 The three reviewers have the same opinion about Technical Merit (Adequate) and different opinions about Presentation. The corresponding weights for the two criteria are shown in Fig.5 0.5 0. and the resulting YAE is in Fig.5 0. (d): Reviewers 1.256 J.25 0 Y 1P u 0 1 2 3 4 5 6 7 8 9 10 1 0.17(e).7.5 0. 9.15 The three reviewers have the same opinion about Presentation (Adequate) and different opinions about Technical Merit. 9.5 0. and the resulting YAE is depicted in Fig. 2 and 3’s aggregated FOUs on Technical Merit and Presentation. 1 0. The corresponding weights for the two criteria are shown in Fig.75 0. 9. (a). Wu all three reviewers are of moderate expertise. so that it is very likely that the AE will recommend that .25 y 0 YR 1 YR 2 YR 3 y 0 1 2 3 4 5 6 7 8 9 10 (c) (d) Figure 9. Finally. (b) and (c): Reviewers 1. (b) and (c): Reviewers 1. Mendel and D.25 y 0 Y 2T = Y 2P y 0 1 2 3 4 5 6 7 8 9 10 (a) u (b) Y3T u 1 0.17(c). observe that regardless of the reviewer’s expertise YAE is very spread out. respectively. when Reviewers 1 and 3 are of high expertise and Reviewer 2 is of low expertise. 2 and 3’s aggregated FOUs on Technical Merit and Presentation. their weights are depicted in Fig.25 y 0 Y 2T = Y 2P y 0 1 2 3 4 5 6 7 8 9 10 (a) u (b) Y3P u 1 0. (d): Reviewers 1.17(d). uY 1T 1 0.17(f).M.75 0. 2 and 3’s overall FOU. 9.5 0.

75 0.25 0 YA E y 0 1 2 3 4 5 6 7 8 9 10 (c) u (d) W1 = W3 1 0.g.18(f). (a). (b). observe that a better YAE is obtained when a reviewer’s YR j in Fig.15(d). and starts with YR1 .Computing with Words for Hierarchical and Distributed Decision-Making 257 the paper be rewritten. Reviewer 2 is of moderate expertise and Reviewer 3 is of low expertise.75 0.15(d).3.2 Technical Merit is Adequate.5 0.4.5 0.5 0. as can be seen by comparing YAE in Figs. On the other hand.18(d).5 0. 9. 9.25 0 YA E y 0 1 2 3 4 5 6 7 8 9 10 (a) u (b) u 1 0. and (9) leads to YAE that is depicted in Fig..4. (d) and (f). and the resulting YAE is in Fig.16(d) and that are much more bunched together and have considerably more overlap than the ones in Fig. (d) and (f): the corresponding YAE . 9.2. 9.75 0.18(a).75 0. (c). u 1 0.18(b). their weights are depicted in Fig. . Next. 9.25 0 W3 W2 W1 u w 0 1 2 3 4 5 6 7 8 9 10 1 0. 9. the expertise of the reviewer is important.18(c). When Reviewer 1 is of high expertise.25 0 W1 = W2 = W3 w 0 1 2 3 4 5 6 7 8 9 10 1 0. So.75 0. (e): weights for the three reviewers. Finally. when reviewer expertise varies.25 0 W2 u w 0 1 2 3 4 5 6 7 8 9 10 1 0.75 0.17 Aggregation at the AE level for the three reviewers’ opinions shown in Fig.18(e). 9.25 0 YA E y 0 1 2 3 4 5 6 7 8 9 10 (e) (f) Figure 9.5 0. and Reviewer Expertise Varies This is a continuation of Section 9.5 0.17(b). when all three reviewers are of moderate expertise. Reviewer 3) and the reviewer is of high expertise. YR2 and YR3 that are in Fig. when Reviewers 1 and 3 are of high expertise and Reviewer 2 is of low expertise. 9.4. their weights are depicted in Fig. Presentation Varies. and the resulting YAE is depicted in Fig. 9. 9.15(d) is farther to the right (e. 9. their weights are depicted in Fig. 9.

Observe again that a better YAE is obtained when a reviewer’s YR j in Fig.5 0. 9. Wu u 1 0.18(b).5 0. Reviewer 3) and the reviewer is of high expertise.18 Aggregation at the AE level for the three reviewers’ opinions shown in Fig.5 0.16(d).25 0 YA E y 0 1 2 3 4 5 6 7 8 9 10 (a) u (b) u 1 0.25 0 W3 W2 W1 u w 0 1 2 3 4 5 6 7 8 9 10 1 0.3 Conclusions In practice when an AE receives three widely different reviews the result is a recommendation that the paper be rewritten. 9.75 0. Mendel and D.18 that. as can be seen by comparing YAE in Figs.75 0.75 0. The JPJA gives results that seem to be consistent with this.75 0. 9.25 0 YA E y 0 1 2 3 4 5 6 7 8 9 10 (e) (f) Figure 9. 9.75 0. so that again it is very likely that the AE will recommend that the paper be rewritten. (e): weights for the three reviewers. (a).4.4.258 J. YAE is very spread out (although not quite as much as in Fig. observe from Fig. again the expertise of the reviewer is important. .17. (c). regardless of the reviewer’s expertise. In all examples there are three reviewers.5 Complete Reviews This section provides three examples of the complete JPJA for different kinds of reviews so that one can gain confidence that the JPJA will be a useful assistant to an AE.5 0. when reviewer expertise varies.25 0 W1 = W2 = W3 w 0 1 2 3 4 5 6 7 8 9 10 1 0.4. one may ask: “What can be concluded from all of this?” Again.M.5 0. 9. (d) and (f): the corresponding YAE .16(d) is farther to the right (e.17)..75 0.g.25 0 YA E y 0 1 2 3 4 5 6 7 8 9 10 (c) u (d) W1 = W3 1 0. 9. As in Fig. So.25 0 W2 u w 0 1 2 3 4 5 6 7 8 9 10 1 0.5 0. 9. (d) and (f). (b). 9.

19(a)–(c).5. An AE will tend to put more credence on the reviews from Reviewers 2 and 3 than on the review from Re- . They are used together with the overall FOUs in Fig. 9.5.19(g). 9. 9.. Rewrite) = 0. however.20(a)–(c).19(d)–(f). 9.19(i).4. and their levels of expertise are High and Moderate. These FOUs were computed using (8).72 Because ss(YAE .26 ss(YAE . Observe that all of the overall FOUs are also towards the right side of the scale 0–10. Marginal or Adequate. Accept) = 0.1 A Submission that would be Accepted The reviewer’s three completed generic review forms are depicted in Figs. The three overall FOUs look very similar because the three review forms are very similar.8). so the JPJA has made the recommendation that agrees with our earlier stated intuition.19(h). Reject) = 0 ss(YAE . 9.4. Observe that all of the linguistic assessments are in the Good or Excellent columns suggesting that this paper probably will be accepted.Computing with Words for Hierarchical and Distributed Decision-Making 259 9. suggesting that things are looking even better for this submission. respectively. The reviewer’s overall FOUs are shown in Fig. The decoder computes Vlachos and Sergiadis’s subsethood measure between YAE and the three words in the decoding codebook (Fig. the AE can recommend “Accept” with high confidence. YAE . 9. Observe that all of the FOUs are towards the right side of the scale 0–10. These FOUs were computed using (6) and (7). Instead. i. 9. The weights for each of the reviewers are shown in Fig. Will the JPJA arrive at this recommendation? The three reviewers’ aggregated FOUs for Technical Merit and Presentation are depicted in Figs. but his expertise is Low. Reviewer 1 has assessed all of the sub-categories as either Adequate or Good. YAE is sent to the decoder. we are not going to leave things to a subjective “look” to make the final publication judgment. depicted in Fig. Observe that YAE is also towards the right side of the scale 0–10. Reviewers 2 and 3 have assessed all of the sub-categories as Poor. suggesting that things are looking very good for this submission.19(g) in (9) to obtain the aggregated FOU for the AE. Accept) is much larger than the other two subsethoods and it is greater than 0. 9. ss(YAE . suggesting that things are looking good for this submission.e. 9.5.2 A Submission that would be Rejected The reviewer’s three completed generic review forms are depicted in Figs.

25 y 0 W3 W1 W2 w 0 1 2 3 4 5 6 7 8 9 10 (g) u (h) Reject Rewrite YA E Accept 1 0.25 y 0 Y 3P Y 3T y 0 1 2 3 4 5 6 7 8 9 10 (e) u (f) Y R 3 YR 1 Y R 2 u 1 0. (a).75 0.25 0 0 1 2 3 4 5 6 7 8 9 10 1 0.5 0. (f): aggregated Technical Merit and Presentation FOUs for the three reviewers. (c): the completed review forms. Wu viewer 1. (e).75 0.5 0.25 0 y 0 1 2 3 4 5 6 7 8 9 10 (i) Figure 9. . Will the JPJA also arrive at this recommendation? (a) (b) u 1 0.260 J. suggesting that this paper will probably be rejected.M.25 0 0 1 2 3 4 5 6 7 8 9 10 1 0.75 0.19 The first complete paper review example.25 0 Y1P Y 1T y 0 1 2 3 4 5 6 7 8 9 10 (c) u (d) Y2T Y 2P u 1 0. (i): the aggregated AE’s FOU in relation to the three categories.75 0. (g): overall FOUs of the three reviewers.75 0. (b).5 0.5 0.5 0.75 0.5 0. (h): the weights associated with the three reviewers. Mendel and D. (d).

3 A Submission that would be Rewritten The reviewer’s three completed generic review forms are depicted in Figs. The three reviewers have assessed all of the sub-categories as Adequate.21(g). It seems that the AE is getting mixed reviews for this submission.21(h). 9. Comparing this figure with Fig. 9.21(d)–(f). 9. This suggests that things are not looking good for this submission. when YAE is sent to the decoder the following results are obtained: ss(YAE .21(i). 9. it is not clear from the actual reviews whether the paper will be accepted or will have to be rewritten. How (or) will this ambiguity be seen by the JPJA? The three reviewers’ aggregated FOUs for Technical Merit and Presentation are depicted in Figs.5. Observe that the FOUs for all three reviewers are towards the left on the scale 0–10.20(g) in (9) to obtain the aggregated FOU for the AE. Good or Excellent. whereas the reviewer’s overall FOUs for the present case cover a larger range on that scale. 9. Observe that YAE is also towards the left side of the scale 0–10. YAE . The reviewer’s overall FOUs are shown in Fig. Indeed. They are used together with the overall FOUs in Fig. Observe that YAE covers a large range on the scale of 0–10.30 ss(YAE . The reviewer’s overall FOUs are shown in Fig. They are used together with the overall FOUs in Fig.19(g). 9. the AE can recommend Reject with high confidence. 9. Reject) is much larger than the other two subsethoods and it is larger than 0. Accept) = 0 Because ss(YAE . however.Computing with Words for Hierarchical and Distributed Decision-Making 261 The three reviewers’ aggregated FOUs for Technical Merit and Presentation are depicted in Figs. The weights for each of the reviewers are shown in Fig.21(g) in (9) to obtain the aggregated FOU for the AE. it is clear that the reviewer’s overall FOUs for the accepted paper are all to the right of 5 on the scale of 0–10. Rewrite) = 0. YAE .74 ss(YAE . 9.5. Reject) = 0. The overall FOUs for all three reviewers are very clearly shifted to the left on the scale 0–10.20(g). 9. The weights for each of the reviewers are shown in Fig.20(d)–(f). 9. 9. 9.20(i). . suggesting that things are looking even worse for this submission.21(a)–(c). depicted in Fig. Once again the JPJA has provided a recommendation that agrees with our earlier stated intuition.20(h).4. suggesting that things are looking very bad for this submission. 9. depicted in Fig.

on Fuzzy Systems. the Per-C can also easily aggregate mixed inputs of numbers. 9. 9.5 Conclusions The examples in this chapter have demonstrated that the Per-C is a very useful tool for hierarchical and distributed decision-making. however. the AE can recommend Rewrite with high confidence. . and is being published in this chapter with the permission of IEEE Press and John Wiley. Rewrite) = 0. The authors would like to thank Dr. the Per-C has great potential in complex decision-making problems.06 ss(YAE . Accept) = 0. This suggests that the JPJA will be a very useful assistant to an AE. intervals and words. hence. Wu Indeed.5. Rewrite) is much larger than the other two subsethoods and it is larger than 0.4 Conclusions From these three complete examples. Reject) = 0. Acknowledgment Much of the material in this chapter will appear in the authors’ book [19]. Nihil Pal.M. when YAE is sent to the decoder the following results are obtained: ss(YAE . for his helps in establishing the FOUs used in this chapter. the Editor-in-Chief of the IEEE Trans.262 J.76 ss(YAE . Mendel and D.4.5. the JPJA is providing recommendations that agree with anticipated recommendations.20 Because ss(YAE . In this chapter. all inputs to the JPJA have been words that were modeled as IT2 FSs.

25 0 Y 1P Y 1T y 0 1 2 3 4 5 6 7 8 9 10 (c) u (d) u 1 0.5 0. (c): the completed review forms.5 0. (h): the weights associated with the three reviewers.25 y 0 Y 3T Y3P y 0 1 2 3 4 5 6 7 8 9 10 (e) u (f) u 1 0. (g): overall FOUs of the three reviewers.75 0.25 0 y 0 1 2 3 4 5 6 7 8 9 10 (i) Figure 9. .25 0 Y2T Y 2P 0 1 2 3 4 5 6 7 8 9 10 1 0. (a).5 0. (f): aggregated Technical Merit and Presentation FOUs for the three reviewers. (i): the aggregated AE’s FOU in relation to the three categories.5 0.25 y 0 W1 W3 W2 w 0 1 2 3 4 5 6 7 8 9 10 (g) u (h) Reject Y A E Rewrite Accept 1 0.25 0 YR 3 YR 2 YR 1 0 1 2 3 4 5 6 7 8 9 10 1 0.5 0. (d).5 0. (e).20 The second complete paper review example. (b).75 0.75 0.Computing with Words for Hierarchical and Distributed Decision-Making 263 (a) (b) u 1 0.75 0.75 0.75 0.

25 y 0 W1 = W2 W3 w 0 1 2 3 4 5 6 7 8 9 10 (g) u (h) Reject Rewrite Y A E Accept 1 0. Mendel and D.25 0 Y 2T 0 1 2 3 4 5 6 7 8 9 10 1 0. (g): overall FOUs of the three reviewers.75 0. .5 0. (e).5 0.25 0 YR 3 YR 2 YR 1 0 1 2 3 4 5 6 7 8 9 10 1 0.5 0.25 0 y 0 1 2 3 4 5 6 7 8 9 10 (i) Figure 9.25 0 Y 1T Y 1P y 0 1 2 3 4 5 6 7 8 9 10 (c) u (d) Y 2P u 1 0. (d). (b).M. (f): aggregated Technical Merit and Presentation FOUs for the three reviewers.75 0.25 y 0 Y 3T Y 3P y 0 1 2 3 4 5 6 7 8 9 10 (e) u (f) u 1 0. (c): the completed review forms. Wu (a) (b) u 1 0. (i): the aggregated AE’s FOU in relation to the three categories.75 0.75 0. (h): the weights associated with the three reviewers.75 0.5 0.5 0. (a).21 The third complete paper review example.264 J.75 0.5 0.

is to-date the most widely used kind of T2 FS. and u∈Jx 1/u is the secondary membership function (MF) at x.e.2) An IT2 FS is shown in Fig. 1].  X $ $ $H [ $  Figure 9. an interval in [0. u is the secondary variable. FOU(A) = x∈X Jx (A.Computing with Words for Hierarchical and Distributed Decision-Making 265 Appendix A A. It is described as [4. b] : 0 i. the membership grade of each element of an IT2 FS is an interval [μA (x). A.22. It is bounded by an upper MF (UMF) A and a lower MF (LMF) A. by the union of all of the primary memberships. Jx . called the footprint of uncertainty of A [A]. 21] A= 1/(x. Uncertainty about A is conveyed Note that (A. 9.22 )28 $.. (A. u) = 1/u x∈X u∈Jx x. both of which are T1 FSs.1) x∈X u∈Jx where x is the primary variable. a b 1}.1 Interval Type-2 Fuzzy Sets (IT2 FSs) An IT2 FS. The FOU is shown as the shaded region.1) means: A : X → {[a. μA (x)]. and is the only kind of T2 FS that is considered in this chapter. consequently. is the primary membership of x.

. 1]. Ae = ∑ ui /xi . Note that an IT2 FS can also be represented as A = 1/FOU(A) (A. namely u1 .e. an embedded T1 FS Ae has N elements. . Jx2 . . . . Ae is an embedded T1 FS (Wu and Mendel [11]. i=1 N ui ∈ Jxi ⊆ [0. . . x2 . c 2007. An IT2 FS. xN } and discrete Jx . JxN . For discrete universes of discourse X = {x1 . IEEE). uN .3) with the understanding that this means putting a secondary grade of 1 at all points of A. i. . . . one each from Jx1 . . u2 . . ..

X i ] Wi = 1/FOU(Wi ) = 1/[W i .4) Mendel and John [22] have presented a Representation Theorem for a general T2 FS. . Mendel and D. Then A is represented by (A.1.3) and (A.e. .3). . 9. Doing this is equivalent to discretizing with infinitesimally many small values and letting the discretizations approach zero.10) (A. . to compute YLWA . fuzziness.W i ] ∑i=1 Wi max (A. such as the centroid. also Fig. . uiMi .24. . 8 Although there are a finite number of embedded T1 FSs. (A. μA (x) ≡ x∈X [μA (x). . . respectively. Let Ae denote the jth embedded T1 FS for A. x2 .1 (Representation Theorem for an IT2 FS).23 and 9.6) where Y LWA and Y LWA are the LMF and UMF of YLWA . in which8 nA FOU(A) = j=1 j Ae = x∈X μA (x).M.9) (A. ui1 .’s subsethood measure [18].22. μA (x)] at each x.W i ] it follows that Y LWA = Y LWA = ∑n XiWi i=1 n ∀Wi ∈[W i . x1 . xN . ui2 . Because in (1) Xi only appears in the numerator of YLWA . is very useful for deriving theoretical results. FOU(YLWA ) = Y LWA .5) This representation of an IT2 FS. Note that if each Jxi is discretized into Mi levels. i=1 N (A. 9. which when specialized to an IT2 FS can be expressed as: Theorem A. μA (x)] ≡ A. Recall from the Representation Theorem [(A. cardinality. . and at each of these values j its primary memberships ui are sampled at Mi values. A . Assume that primary variable x of an IT2 FS A is sampled at N values. i. A.W i ] ∑i=1 Wi min ∑n X iWi i=1 n ∀Wi ∈[W i . the embedded T1 FSs.2 Linguistic Weighted Average (LWA) YLWA . the Jaccard similarity measure [16]. . ..Y LWA one only needs to compute Y LWA and Y LWA . as expressed by (1).8) (A. Rickard et al.7) (A.5)] that Xi = 1/FOU(Xi ) = 1/[X i . where nA = ∏ Mi . .266 J. Wu Examples of Ae are μA (x) and μA (x). So. see. is itself an IT2 FS that is characterized by its FOU(YLWA ). in terms of simple T1 FSs. . variance and skewness of IT2 FSs [16]. it is customary to represent A as an interval set [μA (x). there will be a total of nA Ae . as shown in Figs.

L E D E . c 2007. Because all UMFs are normal T1 FSs. and it is briefly introduced next. Denote the height of X i as hX i and the height of W i as hW i . hY LWA = hmin . 12] to compute Y LWA and Y LWA efficiently.23 [ Xi and an α -cut (Wu and Mendel [11]. LO L D .25 YLWA and associated quantities (Wu and Mendel [11]. min hW i }.Computing with Words for Hierarchical and Distributed Decision-Making 267 An α -cut [24] based approach has been proposed [11.11) Then [12].  X </:$ KPLQ \ /O D  </:$ \ 5U \ /U \ 5O </:$ \ Figure 9.24 Wi and an α -cut (Wu and Mendel [11]. LU L LO LU  Figure 9. IEEE). IEEE). the heights of Y LWA and Y LWA need to be determined. c 2007. hY LWA = 1. Let hmin = min{min hX i . c 2007. IEEE). .L X . First.  K D .  K D :L X : L F : L LO G LU F G : LU L LO  Z Figure 9. ∀i ∀i (A.

14) is computed as r ∑i=1 dil (α ) + ∑n ∗ +1 cir (α ) i=Lr (d) yRl (α ) in (A. dil (α )] be an α -cut on W i (see Fig.19) . α ∈ [0. yRl (α )] be an α -cut on Y LWA (see Fig.23 and 9. The end-points of the α -cuts on YLWA are computed as solutions to the following four optimization problems [11.e. [cil (α ). dir (α )] be an α -cut on W i . 9. so.20) . ∀ wi ∈[cir (α ).25). yRr (α )] be an α -cut on Y LWA .dil (α )] ∑n wi i=1 ∑n bil (α )wi i=1 yRl (α ) = max .2. ∀ wi ∈[cir (α ).16) (b) yRr (α ) in (A. [air (α ). bir (α )] be an α -cut on X i .dil (α )] ∑n wi i=1 Furthermore [11. they are only computed from the corresponding α -cuts on the UMFs of Xi and Wi .15) L∗ L∗ l . yRr (α ) = ∀ wi ∈[cil (α ). hmin ] (A. R∗ . [cir (α ).24 that yLl (α ) and yRr (α ) only depend on the UMFs of Xi and Wi . 9. Observe from (A. α ∈ [0. 1] α ∈ [0. ∑n W i i=1 (A. bil (α )] be an α -cut on X i (see Fig. and Figs. hmin ] (A. α ∈ [0. hmin ] R∗ l ∑i=1 cir (α ) + ∑n ∗ +1 dil (α ) i=Rl In these equations L∗ . α ∈ [0. L∗ and R∗ are switch points that are computed using KM or EKM r r l l l yRl (α ) = l ∑i=1 bil (α )cir (α ) + ∑n ∗ +1 bil (α )dil (α ) i=R R∗ Algorithms [4. Y LWA = ∑n X iW i i=1 .15) is computed as yLr (α ) = r ∑i=1 air (α )dil (α ) + ∑n ∗ +1 air (α )cir (α ) i=Lr L∗ L∗ .12) (A. Wu Let [ail (α ).dir (α )] ∑n wi i=1 ∑n bir (α )wi i=1 max .12) is computed as yLl (α ) = l ∑i=1 ail (α )dir (α ) + ∑n ∗ +1 ail (α )cil (α ) i=L l ∑i=1 dir (α ) + ∑n ∗ +1 cil (α ) i=L l α ∈ [0.17). (a) yLl (α ) in (A. 25].14) (A.2. 9.13) is computed as yRr (α ) = r ∑i=1 bir (α )cil (α ) + ∑n ∗ +1 bir (α )dir (α ) i=Rr r ∑i=1 cil (α ) + ∑n ∗ +1 dir (α ) i=Rr R∗ R∗ . ∀ wi ∈[cil (α ). 1] (A.13) (A.24).. i. [yLl (α ).M. hmin ] α ∈ [0. 12]: yLl (α ) = min Theorem A. where the subscripts l and L mean left and r and R mean right.268 J. 1] α ∈ [0. 12]: ∑n ail (α )wi i=1 .dir (α )] ∑n wi i=1 ∑n air (α )wi i=1 yLr (α ) = min .17) (c) yLr (α ) in (A.18) (A.23). and [yLr (α ). Mendel and D. 9. 1] (A.16). (A.

(d) Repeat Steps (b) and (c) for every α j ( j = 1. 1] into m − 1 intervals and set α j = ( j − 1)/(m − 1). To do this: (a) Store the left-coordinates (yLr (α j ). 2.19). α j ). (A. hmin] into p − 1 intervals and set α j = hmin ( j − 1)/(p − 1). . . (b) Store the right-coordinates (yRr (α j ).24 that yLr (α ) and yRl (α ) only depend on the LMFs of Xi and Wi . . . p). . . p. dir (α j )]. j = 1. The algorithm for computing Y LWA is: (1) Calculate yLl (α j ) and yRr (α j ). . p. m). . To do this: (a) Store the left-coordinates (yLl (α j ).11). j = 1. the minimum height of all X i and W i . . . . α j ). . which is a normal T1 FS.g. (e) Repeat Steps (c) and (d) for every α j .19). the height of Y LWA is hmin . . The algorithm for computing Y LWA is: (1) Calculate yLr (α j ) and yRl (α j ). and Figs. m. Similarly. (c) Use a KM or EKM algorithm to find yLl (α j ) in (A. . n) for each α j . 9. m. and hmin in (A. j = 1. and then go to Step 2. To do this: (a) Select appropriate m α -cuts for Y LWA (e. . (c) (Optional) Fit a spline curve through the 2m coordinates just stored. . p. ∑n W i i=1 (A. .17). (b) Find the corresponding α -cuts on X i and W i (i = 1. To do this: (a) Determine hX i and hW i . . m). j = 1. (d) Use a KM or EKM algorithm to find yLr (α j ) in (A.23 and 9. . j = 1. α j ).18).18) and yRl (α j ) in (A. . 2. . . j = 1. . (b) Store the right-coordinates (yRl (α j ). . . . j = 1. α j ).. . . bil (α j )] and [cir (α j ). . . . . denote the end-points of the α -cuts as [ail (α j ).g. . . (2) Construct Y LWA from the m α -cuts. (2) Construct Y LWA from the p α -cuts. i = 1. j = 1. divide [0. (b) Select appropriate p α -cuts for Y LWA (e. .21) Unlike Y LWA . and then go to Step 2. p. . bir (α j )] and [cil (α j ). respectively. . Y LWA = ∑n X iW i i=1 . n. . . . j = 1. observe from (A. Y LWA is also normal. . (c) Find the corresponding α -cuts [air (α j ). . . . . dil (α j )] on X i and W i. . . .Computing with Words for Hierarchical and Distributed Decision-Making 269 Because all X i and W i are normal T1 FSs. divide [0. . hence.. m.16) and yRr (α j ) in (A.

e. .. 2. Wu (c) (Optional) Fit a spline curve through the 2p coordinates just stored. In summary. . n) at which the slope of these functions changes. 6HOHFW S D FXWV IRU </:$ LQ > KPLQ @ 6HW M  )LQG DLU D M . . i. points on the LMFs and UMFs of Xi and Wi (i = 1. . Y LWA and Y LWA .270 J. For triangular or trapezoidal IT2 FSs. it is possible to reduce the number of α -cuts for both Y LWA and Y LWA by choosing them only at turning points. 9. A flowchart for computing Y LWA and Y LWA is given in Fig.M.26. computing YLWA is equivalent to computing two FWAs. Mendel and D.

 ELO D M .

 FLU D M .

 DQG GLO D M .

 L " Q 6HOHFW P D FXWV IRU </:$ LQ > @ 6HW M  )LQG DLO D M .

 ELU D M .

 FLO D M .

 DQG GLU D M .

 L " Q &RPSXWH \ /O D M .

E\ DQ (.0 DOJRULWKP &RPSXWH \ 5U D M .

0 DOJRULWKP &RPSXWH \ /U D M . E\ DQ (.

0 DOJRULWKP &RPSXWH \ 5O D M . E\ DQ (.

0 DOJRULWKP </:$ D M . E\ DQ (.

> \ /U D M .

 \ 5O D M .

@ M M  </:$ D M .

> \ /O D M .

 \ 5U D M .

@ M &RQVWUXFW </:$ ></:$  </:$ @ P 1R M S 1R M M  &RQVWUXFW </:$ <HV IURP </:$ D M .

 M  S </:$ <HV IURP </:$ D M .

(1996).26 A flowchart for computing the LWA (Wu and Mendel [12]. 158–164. 3rd Int’l ICSC Symposium on Fuzzy Logic and Applications. Budescu. 103–111. M. A. . (1995). c 2008. In Proc. (Prentice-Hall. Fuzzy logic = Computing with words. Upper Saddle River. IEEE). Bibliography [1] L. [4] J. [2] L. on Fuzzy Systems. [3] T. (1999). A review of human linguistic probability processing: General principles and empirical evidence. Mendel. Mendel. 43–62. The Knowledge Engineering Review. IEEE Trans. M. Computing with words. NY (June. From computing with numbers to computing with words – From manipulation of measurements to manipulation of perceptions. 1999). IEEE Trans. 46 (1). Rochester. 4. 10(1). on Circuits and Systems–I. Wallsten and D. A. Uncertain Rule-Based Fuzzy Logic Systems: Introduction and New Directions. S. V. NJ. Zadeh. 2001). Zadeh. when words can mean different things to different people. pp. [5] J. M  P Figure 9. 105–119.

M. similarity measures and uncertainty measures for interval type-2 fuzzy sets. on Fuzzy Systems. Mendel. 1503–1521. (2008). John. Proc. A. Wu. J. IEEE Trans. on Fuzzy Systems. M. Perceptual reasoning for perceptual computing. Wu and J. [10] F. 10 (2). (2008). Wu and J. Upper Saddle River. (2008).Computational Intelligence in Complex Decision Systems 271 [6] J. Mendel and R. I. M. 1664–1666. Mendel. 16(6). Zadeh. South Korea (August. Information Sciences. 12(3). (2002). 16(6). NJ. 1397–1411. Rickard. and R. Mendel. Information Sciences. [13] F. 172. and G. 16(6).pdf. Enhanced Karnik-Mendel Algorithms. 700–705. 3(3). Gibbon. [15] D. 50–60. Mendel. Mendel. 988–1006. IEEE Trans. Fuzzy subsethood for fuzzy sets of type-2 and generalized type-n. on Fuzzy Systems. (2007). (2009).win. Fuzzy Sets and Systems. 179(8). M. (2008). on Fuzzy Systems. Standard background material about interval type-2 fuzzy logic systems that can be used by all authors. M. on Fuzzy Systems. Zadeh. Mendel. Wu and J. A. 177. IEEE Computational Intelligence Magazine. IEEE Trans.t2. 15(6). Wu. [7] J. IEEE Trans. 1995). Vlachos and G. 117–127 (2002). Klir and B. USC-SIPI Report 398. as computed using the Karnik-Mendel algorithms. to appear in 2010). Encoding words into interval type-2 fuzzy sets using an interval approach. Signal and Image Processing Institute. M. pp. J. on Fuzzy Systems. entropy. [24] G. Aggregation using the linguistic weighted average and interval type-2 fuzzy sets. An architecture for making judgments using computing with words. Hagras. Mendel and D. (2005). Information Sciences. CA. on Fuzzy Systems. Interval type-2 fuzzy set subsethood measures as a decoder for perceptual reasoning. (2008). A comparative study of ranking methods. IEEE Trans. M. Int’l Journal of Applied Mathematics and Computer Science. http://ieeecis. 1–40. Wu and J. Aggregation using the fuzzy weighted average. M. 1550–1564. (2009). Subsethood. (2009). M. (2007). [18] D. 1–12. Mendel. John. 17(4). 325–335. 11–22. 2009). (2009). Mendel. Liu and J. [23] J. Wu and J. Los Angeles. 12(1). (Prentice-Hall. Toward human level machine intelligence – Is it achievable? The need for a paradigm shift. Sergiadis. Perceptual Computing: Aiding People in Making Subjective Judgments. on Fuzzy Systems. [22] J. M. Mendel. Mendel. 17(1). Jeju Island. (2007). IEEE Int’l Conf. [21] J. 1169– 1192. M. [11] D. 1384–1396. Corrections to “Aggregation using the linguistic weighted average and interval type-2 fuzzy sets”. I. 1145–1161. . Aisbett. Yuan. Type-2 fuzzy sets made simple. (Wiley-IEEE Press. Fuzzy Sets and Fuzzy Logic: Theory and Applications. University of Southern California. Perceptual reasoning for perceptual computing: A similarity-based approach. M. [19] J. Mendel. 923–934. (2010). IEEE Trans. Mendel. M. and cardinality for interval-valued fuzzy sets – an algebraic derivation. Wu and J. [14] D. Similarity-based perceptual reasoning for perceptual computing. IEEE Trans. [12] D. [20] I. on Fuzzy Systems. submitted for publication in IEEE Trans. [16] D. 158. Toward a generalized theory of uncertainty (GTU)–an outline. [8] L. M. Computing with words and its relationships with fuzzistics. [17] J.org/ files/standards. H. T. Liu and J. M. IEEE Trans. [25] D. on Fuzzy Systems. Mendel and D. Wu and J. [9] L. 17 (6).

encouraging new and creative ideas. © 2010 Atlantis Press/World Scientific 273 . all key stakeholders are involved at an early stage and throughout the policy’s development. An expression like “This policy has more importance with respect to a considered criterion than the other policy with respect to the same criterion” should be evaluated using fuzzy sets which are one of the main computational techniques. and where appropriate. because of incomplete or non-obtainable information. civil servants involved in policy development not only need all the ‘traditional’ attributes (knowledge of D. Computational Intelligence in Complex Decision Systems. 10. kayai}@itu. The advice and decisions of policy makers are based upon the best available evidence from a wide range of sources. Consequently fuzziness is a necessity that must be taken into account in policy making since it is based on human beings’ judgments. a fuzzy analytic hierarchy process technique is used to evaluate the proposed alternative policies to decrease traffic congestion in Istanbul.1 Introduction A policy is a course or principle of action adopted or proposed by a government.1007/978-94-91216-29-9_10. For policy making to be fully effective.edu. In this chapter. Department of Industrial Engineering.Chapter 10 Realizing Policies by Projects Using Fuzzy Multiple Criteria Decision Making Cengiz Kahraman and ˙ Ihsan Kaya ˙ Istanbul Technical University. party. the data (attributes) are usually vague. In real-case problems. Turkey E-mail: {kahramanc. Effective policy making is one of the most important services that the public service provides to government. DOI 10. The policy-making process takes into account of the impact on and/or meets the needs of all people directly or indirectly affected by the policy. Determining the best policy can be taken into consideration as a decision making problem which is the process of finding the best option from all of the feasible alternatives. questioning established ways of dealing with things. Policy making is the process by which governments translate their political vision into programmes and actions to deliver ‘outcomes’ . ¸ ˙ Istanbul. Atlantis Computational Intelligence Systems 2. making established ways work better. business or individual. Ruan. and involves key stakeholders directly. 34367 Macka.tr The policy-making process is flexible and innovative.desired changes in the real world.

uk).1 sets out the analyses and process which policy-makers need to apply (http://www.274 C. Florida [22]. when it (begins to) exhibit (i) computational adaptivity. ability to design implementation systems).gov. evolutionary programming. wetware. Kahraman and ˙ Kaya I.1 A policy cycle. Figure 10. In order to have a brief outline of history of the term the founding and most interesting definitions will be summarized below: The first one was proposed by Bezdek [2] as follows: “A system is called computationally intelligent if it deals only with numerical (low-level) data. Eberhart et al. etc. (ii) computational fault tolerance. The birth of computational intelligence (CI) is attributed to the IEEE World Congress on Computational Intelligence in 1994 at Orlando. [8] defined CI as follows: “Computational intelligence is defined as a methodology involving computing (whether with a computer. There are tools and techniques that can help in making a policy more effectively. and additionally. and (iv) error rates that approximate human performance. but also numerous explanations of the term have been published. has a pattern recognition component. relevant law and practice. and artificial life. fuzzy systems. and does not use knowledge in the artificial intelligence (AI) sense. Figure 10. (iii) speed approaching human-like turnaround.” Marks [18] defined CI by listing the building blocks being neural nets. genetic algorithms. but they must also understand the context within which they (and the policy) have to work. Since that time not only a great number of papers and scientific events have been dedicated to CI.ofmdfmni.) that exhibits an ability to learn and/or deal with new situations such that the system is perceived to possess one or . understanding of key stakeholders’ views.

MCDM approaches can be classified on the basis of the major components of multiple criteria decision analysis.2 Multiple Criteria Decision Making Decision-making problem. introduced by Saaty [23]. In almost all such problems there may be so many criteria and the process of evaluating the efficiency of alternatives is a complicated problem. It has been one of the fastest growing areas during the last decades depending on the changing in the business sector.” Computational intelligence. The distinction between MADM and MODM is based on the . is the process of finding the best option among all of the feasible alternatives. association. a multiple criteria decision making technique based on analytic hierarchy process (AHP). 10. and decisions under certainty versus decisions under uncertainty. such as making decision on order of the projects. and so on.5. Conclusions and future research directions are explained in Section 10. The rest of this chapter is organized as follows: The decision making process and some multiple criteria decision making approaches are summarized in Section 10. The readers can find more details about CI techniques in Chapter 1. CI techniques have been used to solve different problems successfully.4. for many such problems.2. In this chapter. fuzzy logic and evolutionary computing. the application of MCDM methods has considerably become easier for the users and the decision makers as the application of most of the methods are corresponded with complex mathematics. Especially in the last years. Generally the primary concern of MCDM problems is to choose the most preferred alternative. which consists of neural network.Realizing Policies by Projects Using FMCDM 275 more attributes of reason. is proposed under fuzzy environments to evaluate alternative projects which aim at decreasing traffic congestion in Istanbul. The proposed fuzzy multiple criteria decision making (FMCDM) methodology is explained in Section 10. The details of application that the proposed methodology is applied to decrease traffic congestion in Istanbul are illustrated in Section 10. Three different classifications can be made as follows: multiobjective decision-making (MODM) versus multi-attribute decision-making (MADM). and abstraction. where computer usage has increased significantly. such as generalisation. That is. There are many MCDM approaches which differ in how they combine and utilize the data. discovery. individual versus group decision-making problems. is a novel technology to bring intelligence into computation.3. or ranking alternatives in order of importance for selection problems. decision makers put multiple criteria decision making (MCDM) techniques into account. or screening alternatives for the final decision.

the quality of . In the literature some multiple criteria approaches are available to help decision making process as briefly explained below: Analytic Hierarchy Process Analytic Hierarchy Process is a structured approach to decision making developed by Saaty [23]. it has been applied to a wide variety of decision making problems. b) ∈ A × A: a S b if. An objective is a statement about the desired state of the system under consideration. including the evaluation of alternatives. Criterion is a general term and includes both the concepts of attributes and objectives. It is a weighted factor-scoring model and has the ability to detect and incorporate inconsistencies inherent in the decision making process. evaluation criteria which are the standards of judgments or rules on which the alternatives are ranked according to their desirability. or derived from. (c) Evaluating alternatives in terms of criteria (the values of the criterion functions). Kahraman and ˙ Kaya I. a set of attributes [10]. (b) Developing alternative systems for attaining the goals (generating alternatives).276 C. It indicates the directions of improvement of attributes. According to Roy [21] an outranking relation is a binary relation S defined on the set of alternatives A such that for any pair of alternatives (a. The underlying idea consists of accepting a result less rich than the one yielded by Multiattribute Utility Theory (MAUT) by avoiding the introduction of mathematical hypotheses which are too strong and asking the decision maker (DM) some questions which are too intricate. Objectives are functionality-related to. Therefore. (f) If the final solution is not accepted. The concept of an outranking relation is introduced by Bernard Roy who is the founder of outranking methods. Outranking Methods The outranking methods are based on the construction and the exploitation of an outranking relation. The main steps of multiple criteria decision making are as follows [12]: (a) Establishing system evaluation criteria that relate system capabilities to goals. (d) Applying a normative multiple criteria analysis method. gather new information and go into the next iteration of multiple criteria optimization. given what is known about the preferences of the DM. (e) Accepting one alternative as “optimal” (preferred). An attribute is a measurable quantity whose value reflects the degree to which a particular objective is achieved.

Technique for Order Performance by Similarity to Ideal Solution Technique for Order Performance by Similarity to Ideal Solution (TOPSIS) is one of the well known classical Multi-Attribute Decision Making (MADM) methods. The decision-maker’s preferences are modeled in order to obtain a multiattribute utility function.ti . It was developed by Hwang and Yoon [11]. It is a quite simple ranking method in conception and application compared to other methods for multiple criteria analysis. for instance U(ci . [5]. while at the same time there does not exist any strong reason to refuse this statement [3]. Each particular analytical form for this function has preferential independence conditions to be evaluated. It is well adapted to problems where a finite number of alternative actions are to be ranked considering several. which is also known as Weighted Sum Method [11] is probably the most commonly used MADM approach. The method is based on the concept in which the chosen alternative should have the shortest distance from the positive-ideal solution and the longest distance from the negative-ideal solution. in order to guarantee that the decision maker’s preferences are associated to the basic axioms of the theory [1. an analytical function is obtained which combines all criteria through a synthesis function. particularly in dealing with a . sometimes conflicting. This function aggregates utility functions for all criteria or attributes. 15]. criteria [29]. Then the method chooses an alternative with the maximum similarity to the positive-ideal solution. That is. The Preference Ranking Organization Method for Enrichment Evaluations (PROMETHEE) is a multiple criteria decision-making method developed by Brans et al. di ). there are sufficient arguments to state that the alternative a is at least as good as the alternative b. TOPSIS defines an index called similarity (or relative closeness) to the positive-ideal solution and the remoteness from the negative-ideal solution. Simple Additive Weighting Method The Simple Additive Weighting Method. Multiattribute Utility Theory Multiattribute Utility Theory (MAUT) allows the decision-maker to quantify and aggregate multiple objectives even when these objectives are composed of conflicting attributes.Realizing Policies by Projects Using FMCDM 277 the evaluations of the alternatives and the nature of the problem under consideration. TOPSIS views a MADM problem with m alternatives as a geometric system with m points in an ndimensional space.

Assuming that each alternative is evaluated according to each criterion function. Kahraman and ˙ Kaya I. we can list multi-objective linear programming (MOLP) and its variants such as multi-objective stochastic integer linear programming. Multiobjective Methods In multiple objective decision-making (MODM).278 C. Weighted Product Method For a multi-dimensional MADM problem involving different units of measurement. VIKOR The VIKOR method was developed for multiple criteria optimization of complex systems. or solution process (e.g. This approach is based on the hypothesis that in any decision problem. and the weight stability intervals for preference stability of the compromise solution obtained with the initial (given) weights. nonlinear. This method focuses on ranking and selecting from a set of alternatives in the presence of conflicting criteria. Among MODM methods. It determines the compromise ranking-list. satisfaction.. finite or infinite). characteristic of decision space (e. The method is characterized by the additive utility assumption.g. linear. closeness to an ideal point. there exists a real utility function defined by the set of feasible actions. such as form of model (e. which the DM wishes to evaluate. referring to the total value of each alternative being equal to the sum of the products of the criteria ratings and their weight from the respective alternatives. multi-objective goal programming (MOGoP). application functions are established to measure the degree of fulfillment of the DM’s requirements (achievement of goals. the weights become exponents with respect to each attribute value. the Weighted Product Method is supposed to be more appropriate for application when the attributes are connected by a multiplication operation. single dimensional problem. multi-objective geometric programming (MOGeP). interactive MOLP and mixed 0-1 MOLP.g.. the compromise ranking could be performed by comparing the measure of closeness to the ideal alternative [20]. the compromise solution.. multi-objective non- . which takes a positive value in the case of maximization and a negative value for minimization problems. MODM methodologies can be categorized in a variety of ways. It introduces the multiple criteria ranking index based on the particular measure of “closeness” to the “ideal” solution. Accordingly. stochastic).) on the objective functions and are extensively used in the process of finding “good compromise” solutions. prior specification of preferences or interactive). etc.

it is sometimes unrealistic or even impossible to perform exact comparisons. This is why it looks somewhat similar to human reasoning. Keeping this in mind. and it is relatively difficult for the decision maker to provide exact numerical values. Fuzzy set theory implements groupings of data with loosely defined boundaries. it is called a linguistic variable. if needed. The roots of fuzzy set theory go back to 1965 when Zadeh initiated the concept of Fuzzy Logic Zadeh [30]. many engineering and decision problems can be greatly simplified. Several geometric . It is therefore more natural or realistic that a decision-maker is allowed to provide fuzzy judgments instead of precise comparisons. In this chapter. “mediocre”. 10. The traditional AHP needs exact judgments. The main benefit of extending crisp analysis methods to fuzzy techniques is the strength in solving real-world problems. multi-objective dynamic programming and multi-objective genetic programming. In addition. fuzzy logic that provides a mathematical strength to capture the uncertainties associated with human cognitive process can be used. and “bad” are characterized by fuzzy sets defined in the universe of discourse in which the variable is defined. Analytic hierarchy process is one of the most used group decision making techniques.3 The Proposed Methodology Sometimes a decision maker’s judgments are not crisp. due to the complexity and uncertainty involved in real world decision problems. Therefore most of the evaluation parameters cannot be given as precisely and the evaluation data of the alternative project’s suitability for various subjective criteria and the weights of the criteria are usually expressed in linguistic terms by the DMs. a fuzzy multiple criteria decision making methodology is proposed to select the best project alternative to avoid traffic congestion. To achieve this benefit. where the words such as “good”. It uses approximate information and uncertainty to generate decisions.Realizing Policies by Projects Using FMCDM 279 linear fractional programming. Since knowledge can be expressed in a more natural way by using fuzzy sets. In this case. by generalizing the concept of a crisp set to a fuzzy set with blurred boundaries. any methodology or theory implementing “crisp” definitions may be “fuzzified”. which has imprecision in the variables and parameters measured and processed for the application. linguistic variables are used as a critical aspect of some fuzzy logic applications. AHP based on the fuzzy set theory is applied on decision making process. If a variable can take words in natural languages as its value. The fuzzy set theory was specifically designed to mathematically represent uncertainty and vagueness and provide formalized tools for dealing with the imprecision intrinsic to many problems. In this chapter.

the fuzzy number will be. on the real line. Unfortunately. This is the main reason making the investment decision evaluation extremely hard.2. with a piecewise continuous membership function. so its support will not be bounded. Kahraman and ˙ Kaya I. for c x d ⎪ ⎪ ⎪ ⎩0 for Otherwise b c d with the following properties and membership function (1) Figure 10. If sufficient objective data were available. decision makers do not have enough information to perform such a decision analysis. a) ∪ (d. d] (3) μ (a) = μ (d) = 0 and μ (x) = 1. c. asymptotically zero. by the left or by the right. both the benefits and the costs embedded in a very important strategic investment projects are mostly intangible. standardized trapezoidal fuzzy numbers (STFNs) are used. as illustrated in Fig. mapping functions have been widely adopted. convex fuzzy set. d). such that there are four points a shown in (1): (1) μ (x) = 0 for every x ∈ (−∞. Within this chapter. trapezoidal and S-shaped membership functions (MF) where the triangular MF is a special case of trapezoidal one. ∞) (2) μ is increasing on [a. As mentioned in the preceding sections.2 Membership function of an STFN. b. A trapezoidal fuzzy number. such as triangular. is a normal. A(a. the probability theory would be preferred in such a decision analysis. 10. for every x ∈ [b. The cases a = −∞ and d = ∞ are admitted and. then. c] ⎧ x−a ⎪ ⎪ ⎪ b − a .280 C. since probabilities can never be known with certainty and the decisions about strategic . for a x b ⎪ ⎪ ⎪ ⎨1 for b x c μA (x) = d − x ⎪ ⎪ ⎪ d − c . b] and decreasing on [c.

Later. the fuzzy set theory is used. The decision makers are required to provide their judgments on the basis of their knowledge and expertise for each factor at the bottom level in the hierarchy. 13]. A fuzzy AHP is an important extension of the typical AHP method which was first introduced by Laarhoven and Pedrycz [16]. As different decision makers having different perspectives could have different influence on the final decision. In this chapter. For m decision makers in the evaluation group. According to Zeng et al. In a typical AHP method. To deal quantitatively with such an imprecision or uncertainty. fuzzy multiplication and addition operators. some other fuzzy AHP approaches were developed and applied to some industrial problems [6. The group preference of each factor is then calculated by applying fuzzy aggregation operators. The decision makers can provide a precise numerical value. To deal with this drawback. The pair-wise comparison matrixes for the main and sub-criteria are built by a simple Microsoft Excel based evaluation form. i. influential factors can be decomposed by brainstorming or checklist techniques. fuzzy AHP is proposed. A modified fuzzy AHP method is applied to work out the priority weights of selected attributes. fuzzy aggregation is used to create group decisions. In this method.. a range of numerical values. a modified AHP method proposed by Zeng et al. decision makers should rely on decision makers’ knowledge in modeling projects which are investments to decrease traffic congestion. then defuzzication is employed to transform the fuzzy scales into crisp scales for the computation of priority weights. and c1 + c2 + · · · + cm = 1.1 to the pair-wise comparison so that the priority vector can be computed. [31]. a linguistic term or a fuzzy number. However. 1].3. ci where ci ∈ [0. one drawback of the current AHP method is that it can only deal with definite scales in reality. 7. Step 2: Compare the factors using pair-wise comparisons. attribute comparisons often involve certain amount of uncertainty and subjectiv- .e. scored by fuzzy membership functions and weighed by AHP. The attributes having impacts on the selection of a project are listed and classified in a hierarchical structure as can be seen in Figure 10.Realizing Policies by Projects Using FMCDM 281 level information technology investments are attributable to many uncertain derivations. a fuzzy eigenvector based weighting is used in the model to calculate the decision makers’ competence. However. Here are the steps of the methodology [31]: Step 1: Measure the factors in the hierarchy. In this situation. [31] combined with a different fuzzy ranking method is used for selecting the traffic congestion project best fitting the municipality’s policies. the ith decision maker is assigned a weight. decision makers would have to give a definite number from Table 10.

282 C. . Kahraman and ˙ Kaya I. Figure 10.3 Hierarchical structure of the alternative projects.

A modified fuzzy AHP is used in this study to overcome these shortcomings.. and define scores of the project alternatives against these attributes. a linguistic term or a fuzzy number. A crisp numerical value. In this case. but he/she cannot give a definite score to the comparison because of being not sure about the degree of importance for attribute 1 over attribute 2. Step 3: Convert preferences into STFNs. In this case. then j has the reciprocal value when compared with i. a numerical value. 8 Reciprocals of above non-zero Intermediate values between the two adjacent judgments If activity i has one of the above non-zero numbers assigned to it when compared with activity j. As described in Steps 1 and 2. a range of crisp numerical values and a triangular fuzzy number can be converted to an STFN as follows: .1 Intensity of importance 1 3 5 7 9 Definition Scale of relative importances [23] Explanation Equal importance Weak importance of one over another Essential or strong importance Demonstrated importance Absolute importance 2. 4. the decision maker may prefer to provide a range of 3–5 to describe the comparison as. a range of numerical value. a typical AHP method has to be discarded due to the existence of fuzzy or incomplete comparisons. 6. The decision makers are required to compare every attribute pair-wise in their corresponding group structured in the hierarchy attached in Figure 10.g. the decision maker cannot compare two attributes due to the lack of adequate information. e. “attribute 1 is between weakly more important to strongly more important than attribute 2”. a decision maker is sure that attribute 1 is more important than attribute 2. because the values of factors provided by experts are crisps.Realizing Policies by Projects Using FMCDM 283 Table 10. Two activities contribute equally to the objective Experience and judgment slightly favor one activity over another Experience and judgment strongly favor one activity over another An activity is strongly favored and its dominance demonstrated in practice The evidence favoring one activity over another is of the highest possible order of affirmation When compromise is needed ity. the STFN is employed to convert these experts’ judgments into a universal format for the composition of group preferences.3. (ii) Assume that. Here are two very important special examples to illustrate this situation: (i) Assume that.

b = c = n. . is converted to the STFN as A = (n. the aggregation of STFN scales is defined as ai j = ai j1 ⊗ c1 ⊕ ai j2 ⊗ c2 ⊕ · · · ⊕ ai jm ⊗ cm (3) where ai j is the aggregated fuzzy scale of attribute i comparing to attribute j for i.e. It should be noted that the aggregation should discard the absent scale while it comes with non zero scales provided by other decision makers under the same comparison. DM2 . c = d = m). n. then it is represented with A = (0. . Similarly. y. n) (i. . T = (x. In each case. . . . “about n”. (4) A triangular fuzzy number. DM2 . The decision makers are encouraged to give fuzzy scales if they are not sure about the exact numerical values or leave some comparisons absent as they cannot compare two attributes at all. . (3) A range. . m. 0) (i. a = b = c = d = n). c2 . ..e. d = z). n. . respectively. . . ai j2 . y. (1) A crisp number “n” is converted to the STFN as A = (n. a = b = c = d = 0).. is converted to the STFN as A = (x. This process can be defined as follows. Step 4: Aggregate individual STFNs into group STFNs. a single STFN is employed to convert these decision makers’ judgments into a generic format for the composition of group preferences. a = x. .. . ai jm are the corresponding STFN scales of attribute i comparing to attribute j measured by decision makers DM1 . . . Kahraman and ˙ Kaya I. 0. and c1 . n + 1) (i. j = 1. cm are contribution factors (CFs) allocated to decision maker. n. (2) A linguistic term. a = b = n. ai j = ai j1 ⊗ c1 + ai j2 ⊗ c2 + · · · + ai jm ⊗ cm 1 − ∑ cr (4) . .e. . DM1 . . . m).. Si2 . . y.e. ai j1 . . 0. m) (i. . DMm . n. . DMm and c1 + c2 + · · · + cm = 1. . . Si1 . respectively. (5) If a decision maker cannot compare any two factors at all. DMm . is converted to the STFN as A = (n − 1. This step aims to apply an appropriate operator to aggregate individual preferences made by individual expert into a group preference of each factor. n. d = n + 1). Sim are the STFN scores of the factor Fi measured by m decision makers DM1 . z). a = n − 1. The aggregation of STFN scores is performed by applying the fuzzy weighted trapezoidal averaging operator..e. n. as can be understood from the conversion information in the brackets above. z) (i. respectively. b = c = y. 2. . ⊗ and ⊕ denote the fuzzy multiplication operator and the fuzzy addition operator. .284 C. DM2 . the scale is likely between (n. which is defined by Si = Si1 ⊗ c1 ⊕ Si2 ⊗ c2 ⊕ · · · ⊕ Sim ⊗ cm (2) where Si is the fuzzy aggregated score of the factor Fi .

Assume an aggregated STFN. . Step 5: After having all the required aggregated STFNs. . . Let F1 . . . j = 1. j = 1. a ji = 1/ai j . The priority weights of factors in the matrix A can be calculated by using the arithmetic averaging method wi = 1 n n j=1 ∑ ∑n ai j k=1 ak j i. If none of the decision makers can evaluate a particular comparison. F4 1 a1n 1 a2n where aii = 1. ai j = aaj . n (6) aaj + 2 abj + acj + adj i i i i 6 (5) 1 a12 1 A = ai j = F2 a 1 12 F3 . . A2 . a1n . .. . .. . (7) where wi is the section weight of Fi . . Assuming A1 . a2n . acj . Step 6: Calculate the priority weights of factors. . j = 1. . abj . Fn be a set of factors in one section. 2. F2 . .. .. n. ai j be the defuzzified scale representing the quantified judgment on Fi comparing to Fj . adj .. 1 . n) are transferred into crisp scales ai j within the range of [0. . . . . . . 2.Realizing Policies by Projects Using FMCDM 285 where ∑ cr is the total weight of decision makers who provide zero scales. it is now time for defuzzification to find the representative crisp numbers. 2. . .. . . Consequently. . the final weight wi of Fi can be derived by wi = wi × ∏ wgroup i=1 (i) t (i) (i) (8) All individual upper section weights of wgroup can also be derived by (7) to prioritize sections within the corresponding cluster in the hierarchy. i. i i i i the representative crisp value ai j can be obtained as follows: ai j = where aii = 1. and wsection is the section weight of the ith upper section which contains Fi in the hierarchy. . . . . pair-wise comparisons between Ai and A j in the same group yield an n-by-n matrix defined as follows: F1 F1 F2 . all the aggregated fuzzy scales ai j (i. . An represent a set of attributes in one group. . this comparison should be left absent. and a ji = 1/ai j . F4 . 9]. . . Assume Fi has t upper sections at different level in the hierarchy.

Max and Dmin = D2 A. . Dmax = D2 A. The Max and Min are chosen as follows: I Max(I) Min(I) sup i=1 I s Ai (10) s Ai i=1 inf where s(Ai ) is the support of FNs Ai . . If only one of these conditions is satisfied. n. Then Dmax and Dmin of fuzzy number A can be computed as follows: ⎧⎡ ⎤ 2 ⎪ a2 + a3 1 a2 + a3 ⎪ ⎪⎢ −M + − M × [(a4 − a3) − (a2 − a1)] ⎥ ⎪ ⎪⎢ 2 2 2 ⎥ ⎪ ⎪⎢ ⎥ ⎪ ⎪ ⎢ 1 a3 − a2 2 1 a3 − a2 ⎥ ⎪ ⎪⎢+ ⎥ ⎪ × [(a4 − a3) + (a2 − a1)] + ⎪⎢ 3 ⎥ ⎪ 2 6 2 ⎪⎣ ⎪ ⎦ ⎪ ⎪ + 1 (a − a )2 + (a − a )2 − 1 [(a − a ) × (a − a )] ⎪ ⎪ 4 3 2 1 2 1 4 3 ⎪ 9 9 ⎪ ⎪ ⎪ ⎨ if A is a trapezoidal fuzzy number D2 A. Step 7: Calculate final fuzzy scores. . . The idea is that a FN is ranked first if its distance to the crisp maximum (Dmax ) is the smallest but its distance to the crisp minimum (Dmin ) is the greatest. i = 1. Hence. In this chapter a method proposed by Tran and Duckstein [26] is used to ranking final fuzzy scores.286 C. the attitude of the decision maker in a decision situation). . M = ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪⎡ ⎪ ⎤ ⎪ ⎪ ⎪ (a − M)2 + 1 (a − M) [(a + a ) − 2M] ⎪ ⎪⎢ 2 2 3 1 ⎪ ⎥ 2 ⎪⎣ ⎪ ⎦ ⎪ ⎪ ⎪ + 1 (a − a )2 + (a − a )2 − 1 [(a − a )(a − a )] ⎪ ⎪ 3 2 2 1 2 1 3 2 ⎪ 9 9 ⎪ ⎪ ⎩ if A is a triangular fuzzy number (11) where M is either Max or Min. . . the FN might be outranked by the others depending upon context of the problem (for example. the final fuzzy scores FS can be calculated by FS = ∑ Si wi i=1 ∼ ∼ n ∼ i = 1. 2. n (9) Step 8: Compare the FS values using an outranking method. Min (12) . The method is based on the comparison of distances from fuzzy numbers (FNs) to some predetermined targets: the crisp maximum (Max) and the crisp minimum (Min). . When the scores and the priority weights of factors are obtained. Kahraman and ˙ Kaya I.

the Byzantine Empire (395-1204 and 1261-1453). the population of Istanbul and changes with respect to the years 1990-2007 are presented. Istanbul served as the capital city of the Roman Empire (330-395). Istanbul has been suffering from traffic for over 20 years since it is the vital part of the industry and being overpopulated. The city was chosen as joint European Capital of Culture for 2010. In Table 10. The city covers 27 districts of the Istanbul province. The population of Istanbul has begun to rapidly increase as people from Anatolia migrated to the city since it is the most-industrialized city in Turkey.2 The Population of Istanbul [27] Year 1990 2000 2007 Population 7.4 An Application Istanbul is Europe’s most populous city and Turkey’s cultural and financial center. Also. 836 Change (%) – 39.7 million that 1.8 million that is 29 % of Turkey and the daily number of vehicles in traffic is 1. The historic areas of Istanbul were added to the UNESCO World Heritage List in 1985. 573. problems in parking. geographical features. infeasibilities in the existing structure. In its long history. Since the population’s excessive increase.23 25. the total length of highway is nearly 12. 836 residents according to the latest count as of 2007. excessive number of private cars. the Latin Empire (1204-1261). Istanbul has two bridges in which the first called Bo˘ azici Bridge and the second called g ¸ .773 10.2. It extends both on the European (Thrace) and on the Asian (Anatolia) side of the Bosphorus.50 10. Istanbul has a population of 12. and coordination problems with authority the traffic problem of Istanbul arises more and more.6 million of them are private cars and daily number of vehicles joining Istanbul traffic is 400.000 kilometers and approximately 15 million journeys are made every day in the metropolitan area. As a result of the lack in public transportation vehicles. and is thereby the only metropolis in the world which is situated on two continents. the city has been face to face by many problems [14]. and the Ottoman Empire (1453-1922). The total number of cars is 1. private cars carrying only 1-2 passengers.735 12. The present number of vehicles in Istanbul is 2. 573.195.018. having the 40 % of industry of Turkey.5 million that is 20 % of Turkey. unplanned expansion.Realizing Policies by Projects Using FMCDM 287 Table 10. Istanbul is located in the north-west Marmara Region of Turkey. employees. lacks of expert staff.

5 % is made by rail transportation which includes trains.3. tramcars.000 Fatih Sultan Mehmet Bridge to links Asia and Europe.5 % is made by waterborne which includes ships.700. sea transportation and railway transportation.000 10. IETT nostalgic tramcars and tunnel. respectively [4].000 2. minibuses and IETT Buses while the TDI ships are the most important vehicles among sea based transportation as can be seen from Table 10. According to Table 10. The governmental organizations are Turkish Republic Railways Organization (TCDD) and Turkish Sea Transportation Organization (TDI). and 17 % of which is made by waterborne [4]. Table 10. all vehicle types except tunnel and nostalgic tramcar are almost equally important among railway transportation vehicles. the most important part belongs to company services. Trams and Tunnel (IETT).3 The Distribution of Highway Transportation Type of Vehicle Number of Vehicles 1. taxis.000 1. private cars. and vehicles for company services. minibuses.600. minibuses. Istanbul’s traffic is managed by three local organizations administrated by Istanbul Metropolitan Municipality (IMM) and two central governmental organizations.000 1. private motorboats.000 35. The other two local organizations are Istanbul Fast Ferries Co. The largest and first local organization which controls highway and railway public transportation mostly is Istanbul Establishment of Electricity. . public buses.3-10.000 2.4.100.3 million that 34 % of which are made by private cars.000 590 18. 90 % of Istanbul transportation is made by land-based transportation which includes private cars. 3.000 80. Besides.500 1. fast ferries and private motorboats and 6. Tables 10.100.5 show distributions of highway transportation. metro buses.300 7. and public buses carry a high percentage of passengers. subways.230. Taxis. transportation is also met by the private sector in Istanbul.100.000 1.000. According to Table 10.500 1.664. The daily number of travels between Asia and Europe is 1.000 2.288 C. 49 % of which are made by public transportation vehicles.150.890 Daily Passenger Private cars IETT Buses Public Buses Minibuses Dolmus ¸ Taxis Company Services TOTAL 2. (IDO) which is in charge of sea transportation which was established in 1988.5. Kahraman and ˙ Kaya I. light metro.

European end. Marmaray Project is finished in 2012.000 Republic of Turkey Ministry of Transportation and Communication has managed two important projects which are Marmaray and Third Bosphorus Bridge in Istanbul. Istanbul Strait Rail Tube Crossing and Commuter Railway Upgrading (MARMARAY) Project comprises of three phases as following.5 Type of Vehicle The Distribution of Railway Transportation Number of Vehicles 72 19 11 9 5 116 Daily Passenger TCDD Trains Light Metro Tramcars Subway IETT Tunnel & Nostalgic Tramcar TOTAL 125. which is currently in the early stage of design and construction.4 is the upgrading of approximately 76 km of commuter rail from Halkali.3 km of cut and cover tunnels and four numbers of stations [24].6 km.000 190.000 440.Realizing Policies by Projects Using FMCDM 289 Table 10. (i) Constructing an immersed tube tunnel under the Istanbul Strait with approaching tunnels. Ideally.000 160. Asian end. mainly consisting of 1.1 km of bored tunnels. Rail travelers (and rail cargo) currently have to transfered to a ferry for a three-hour journey across the city. to Gebze. Then that will drop to less than two hours. The Marmaray project which can be shown in Figure 10. the tunnel will also relieve some of the pressure on those bridges.000 160. The total length of the project is 13. 1.4 km of immersed tube.000 11. (ii) upgrading the existing commuter rail system including a new third track on ground and completely new electrical and mechanical systems.000 30.000 130. (iii) procurement of the rolling stock [19].000 Table 10. And when . It includes the new railway system that will be constructed in tunnels under the Bosphorus and it is called bosphorus crossing (BC Project).4 Type of Vehicle The Distribution of Sea Transportation Number of Vehicles 74 26 236 336 Daily Passenger TDI Ships IDO Ferries Motorboats TOTAL 250.000 616. 10. three underground and one surface stations.

north of the other two bridges. o ¨ udar to Tepe¨ st¨ . Figure 10. was subsequently constructed in 1989 when the first bridge became inadequate after a short period of time. The municipality’s policy is that more investment should be made in railway systems. is currently spanned by two bridges.4 Marmaray Project Location [24]. the construction of a third bridge is considered at the northern end of the Bosphorus. ¸ ¸ • Project-III: Ferryboat line from Ortak¨ y to Beylerbeyi. wave at the Hagia Sofia. called as Fatih Sultan Mehmet Bridge.5 are evaluated with the proposed fuzzy decision making methodology to decrease traffic congestion in Istanbul: • Project-I: Subway Line from Ba˘ cılar to ˙ g Ikitelli. Currently they are both inadequate in providing the required capacity. The Bosphorus. Kahraman and ˙ Kaya I. u u • Project-VI: Tramcar line from Kabatas to Besiktas. These two projects and under construction-projects such as the subway line from “Fourth Levent” to “ITU” and the tramcar line from “ehitlik” to “Topkapı” are reckoned without when the alternative projects are discussed in this chapter. in Istanbul. you’ll be able to hop on a train in London. • Project-IV: Light Rail System from Usk¨ u u • Project-V: Light Rail System from Tepe¨ st¨ to Samandıra. a strait that physically divides Istanbul into two parts. Georgia. For this reason. and Turkey. • Project-II: Subway Line from Besiktas to Otogar. ¸ ¸ ¸ • Project-VII: Subway Line from Yenikapı to Ba˘ cılar. As a result the following nine projects which are located in Figure 10. g . the Marmaray Project is integrated into a planned railway through Azerbaijan. and step off days later in Beijing [25].290 C. The first bridge was constructed in 1973 and the second bridge.

o • Project-IX: Light Rail System from K¨ cukcekmece to Beylikd¨ z¨ .000 6.3 are taken into consideration when the proposed AHP approach is used to evaluate the alternative projects for decreasing traffic congestion in Istanbul. The four level hierarchy composed of 6 main criteria and 22 sub-criteria and nine alternatives is given in Figure 10.3.000 30. These criteria have been derivated from the related literature [17.000 70.5 The Suggested Projects to Decrease Traffic congestion in Istanbul.8 10 14.6 • Project-VIII: Light Rail System from Bakırk¨ y to E-5 Bridge.000 70. • Economic (C1).000 70. Management Faculty.000 140. Department of Industrial Engineering.000 140. 32. The following criteria whose hierarchical structure is shown in Figure 10. 28] and the interviews with four experts from the municipality and one professor from Istanbul Technical University (ITU). The estimated main characteristics of these projects are shown in Table 10. .Realizing Policies by Projects Using FMCDM 291 Table 10.000 Cost ($/million) 700 600 30 800 800 5 600 800 800 Length (km) 14.5 9. 9.6 Project I II III IV V VI VII VIII IX Main Characteristics of Proposed Projects Capacity (passenger/day) 70.5 1.5 11.6.000 70.2 14 2 11. u¸ ¨ ¸ u u Figure 10.

• Political (C6). Five experts evaluate the considered criteria to determine the most appropriate project alter- . • Ecological (C3). (a) Contribution to the general transportation system (C21). cultural and social (C5). (c) Dependency on foreign countries for construction (C63). (a) Set up cost (C11). Kahraman and ˙ Kaya I. (b) Demand (passenger/hour) (C42). (c) Economic Value (PW. (c) Safety (C43). B/C) (C13). the most appropriate project alternative is selected by a FMCDM technique. (c) Contribution/harm to social development (C53) (d) Labour impact (C54). (d) Aesthetic appeal (C34). (b) Compatibility with the municipality’s master plan (C22). and country (C52). region. • Transportation system (C2).292 C. (b) Financial possibilities of the government (C62). (a) Capacity (passenger/hour) (C41). (b) Compatibility with the development of the city. IRR. (b) Negative impacts on the ecosystems during the operation phase (C32). (c) Contribution/harm to green(C33). • Technical (C4). (d) Maintenance cost (C14). (e) Reliability (C45). (a) Negative impacts on the ecosystems during the construction phase (C31). (a) Negative impacts on the historical environment during the construction phase (C51). (a) Compatibility with the transportation policy of the government (C61). (d) Feasibility (C44). • Public. (b) Operation cost (C12). In this chapter.

0.7 provides the fuzzy weights of the criteria for Project VI.6. a possible range of numerical value. Each criterion of the hierarchy is evaluated by experts under the defined criteria. 8. a linguistic term. 7.20 ⊕ (3.0.0. 10.Realizing Policies by Projects Using FMCDM 293 native.5. (3). the STFN scale of comparing . 10. 5.00.7. 7. The pair-wise comparisons of “Economic Criteria” and the corresponding STFNs are shown in Table 10.6 Membership Functions for Evaluation.5. 7. or a fuzzy number.0. 9. Figure 10.0) ⊗ 0.0.0) ⊗ 0. For example.0.20 ⊕ (7.40. A score system is shown in Figure 10.0) ⊗ 0. Then these evaluations are converted into STFNs as defined in Table 10.0.20 ⊕ (5. 5.5.20 Ssetup = (5. 8. 7. The aggregation of STFN scales are calculated by Eq. 7.8. Table 10.5.7.20 ⊕ (8.0) ⊗ 0.0) ⊗ 0. the aggregation of “Set up Cost” under “Economic” section is calculated as follows: Ssetup = (5.80) The other values for aggregation are shown in Table 10. 9. For instance. 10. (2).0. 7. 9.0.0.0. 3. Experts may provide a decision about their judgment as a precise numerical value. The aggregations of the obtained scores are calculated by Eq.60. 8.

9. 6. 10. 9. 10) (5.2. 10) (5.8 6. 10. 9. 10) (5.5.9 10 8 7.5. Political C1 C2 C3 C4 C5 C1 C2 C3 C4 C1 C2 C3 C4 C1 C2 C1 C2 C3 C4 C1 C2 C3 VH VH H 9 VH L H VG L L L 9 9.1.9 8. 7. 3. 8.7. 7. 10) (8. 8) (8. Kahraman and ˙ Kaya I. 3. 8. 8.9 3.5. 8) (9. 7.8) (5. 9. 7. 7. 6) (3. 7.5. 8. 3.6.5. 8.8. 8) (7. 7. 7. 5. 9. 7. 10.7 E2 Score STFN E3 Score STFN E4 Score STFN E5 Score STFN Scores and Converted STFN for Project VI Aggregated E1 Score STFN Technological Economic Ecological Transportation System Public. 10.5. 8. 8. 6.5. 8. 7. 6. 8) (7.4.10 8. 6. 10) (7.5. 10.5. 6) (7.8 7 6 7 G 8 (8.8 9. 9. 6.5.4.5. 10) (8. 8.10 VH VH VL H H G 8 VG VL (7. 10) (8. 2. 8. 10) (8. 8) (7. 8. 10.8 9. 8. 7. 10) (6. 7. 9. 7. 7.6. 6.3. 10) (9. 7. 7. 7.8 7. 7.4. 7. 5. 8. 7. 8.4. 8) (5. 8. 9.5.7. 7. 10. 8. 9.8. 7. 4. 7.1. 7. 10) (5. 7. 3. 6. 7.9 8 G 9. 7. 7. 7.4) (6.5. 8. 9.5. 6.4. 10. 9. 7.4) (8. 5.5. 7. 8. 8.2. 8. 7. 9) A7 7 6 7 G A8 A6 A9 A9 A3 A3 A7 A6 A9 H A7 A5 A4 6.6.7. 9) (7.4) (8. 6. 7. 10) (8. 10) 10 10 9 9 9 L 5 7 8 A4 7 7 A7 8. 5) (7. 8.5. 8. 5. 8. 7. 7. 10. 7. 5) (7. 7. 6. 9. 8. Cultural and Social C. 7. 9. 9. 9. 7. 4) (2.6) (5. 6. 5.3. 9) (8. 9) (9. 8) (8. 10) (10. 8.9.5. 8. 8. 7. 7) (9. 8) (9.6. 7) (8. 10) (7. 8) (3. 10. 9. 10.9.9. 10) (9. 9. 9) (7. 7.294 Table 10. 10) (7. 9.5. 8. 9.10 7. 9) (3. 7.5.5. 9. 5.5. 9.9. 9. 9. 9. 7.5. 4) (6. 9. 8) (6. 9. 8.6.2) (7. 8) (7. 9) 9 8 9 8. 7) (5.6. 5. 8) (7.4. 9.1. 7. 10.2) (5.4. 7. 9) (8. 7.6 G L 8 7. 7. 9.6. 8) (7.5. 10) (7. 10) (7. 6) (7. 6. 10.5 5.6) (7. 9. 10.6. 8. 7. 8.9 8. 5. 10.4. 7) (8.9 A7 7.5. 7. 10.2. 10) (7. 10. 5. 8. 8) (5.1. 8. 8. 6. 10. 6) (6.5.5.2) (5. 9. 10. 6. 9.4. 7.5.3. 9. 8. 10) (7. 10) (8.5. 7) (8. 9.5. 8. 6) (5. 7. 6) (7. 10) (5.6. 7. 10. 10. 10. 8. 7.6.7. 8.4. 6. 9) (7. 9. 9. 9. 7. 10. 8) A10 8 8 9. 8. 6. 7. 10) (8. 10. 10) (0.1. 6.6. 9. 8. 8. 8) (9. 7.8) (6. 7.5. 6) (6. 9.5. 7. 10) (5.8. 9. 10) (8. 8) (7. 8) (8. 10. 7. 10. 5.5. 8. 9) (8.5. 7. 5. 7.5.5.2) (6. 8. 9) (8. 9. 10) (5.5.5. 10) (9. 9.10 10 9. 7.8.10 8 7. 9) (9.1. 10) (9. 9. 8. 7.6) (7. 9. 10) (6.9. 9. 10. 7) (8. 7.5. 10. 8. 7. 8. 9.5.9 7. 7.5. 9. 7) (5. 7.5. 10) (5. 8. 6. 7. 7. 7.5. 6. 9. 10) (5. 8. 7) (7. 10) (7. 10. 8.8) (7. 7.4) (4. 7. 7. 10) (5.9 (9. 8. 9.6.4) (6. 10. 9.8) . 9. 10) (2.8) (7. 10) (7. 8.5. 3. 10) (5. 5. 8. 10.8 7. 6. 10) (8. 8. 9. 7) (5.8) (4. 7. 10 ) (7. 9) (5. 9. 6. 10) (9.7 A9 G A7 (6.5. 9.6 A7 A7 9 A8 (10. 7.5. 7. 7) (6. 2. 7. 8) (9.1.5. 5) (6. 6.4) (7. 10) (10. 8. 8. 9) (5. 7) (6. 7. 8) (6. 8. 9) (5. 7. 8.8. 8) (9. 8. 4. 8. 9) (9. 8.4.1.10 8 5. 8. 5) (5. 10. 6. 9) (8.10 VH A10 7 9.4) (8. 10) (8.5.5. 7. 9. 9) (5. 9) (7. 10) (5. 7. 8.9.3. 8) (5. 9. 9. 8. 8. 5. 5) (7. 10. 4. 10. 7. 8.8.10 8. 8) (4.6.8. 4. 8) (6.

respectively is determined as the best project alternative to decrease traffic congestion in Istanbul. 6.0.2 = 4.0.0.10.0) ⊗ 0.0. 7.000 By taking into account this matrix and using Eq. 6.3 is used.417 0.2) Then STFN scale of comparisons should be defuzzified. To rank the fuzzy scores the method explained in Section 10.0.83. the STFN scale of comparing “Operation Cost” with “Maintenance Cost” is defuzzified as a24 = 4.9.3528.909 1.0.IX – VIII-III}.4. and maximum. 5. 0. The results are shown in Table 10.20 a24 = (4. 4.000 1. (8). 5. 5.0) ⊗ 0. Project VI whose Dmax and Dmin values are minimum.4 + 5. 3. 5.0.2356.0) ⊗ 0.0.700 FS = (6.20 ⊕ (4.0.000 5.20 ⊕ (5. 5.0.68.192 1. 0.200 ⎦ 0. 5. 7. The membership functions of these fuzzy scores are shown in Figure 10.20 ⊕ (5.100 4.105 1.208 0.0.IV . 5. (5).Realizing Policies by Projects Using FMCDM 295 “Operation Cost” with “Maintenance Cost” is aggregated as follows: a24 = (5.114 0.2. 5.400 8.4. 5.0.80 6 Using Eqs.0) ⊗ 0.20 ⊕ (3. According to Table 10. the fuzzy scores need to be ranked.7.2) + 5.4 + 2(4. 5.I . In the last step of the proposed methodology. The ranking results are summarized in Table 10. (6)-(7) the pair-wise comparisons matrix of “Economic Criteria” is obtained as follows: ⎡ ⎤ ⎢ ⎥ ⎢ 2. By using Eq.55. 5.0) ⊗ 0. 0.0509} The final weights of the criteria are calculated by using Eq. The ranking of project alternative is determined as follows: {VI – II .V VII . . (8) the weights of the sub-criteria of “Economic Criteria” are calculated as follows: w = {0. 5. (9) as follows: Pr −V I ∼ ∼ ∼ 1.3607.475 2. Then the FS of Project VI is calculated by using Eq. 4.800 ⎥ ⎢ ⎥ AEconomic = ⎢ ⎥ ⎣ 0.89) FS values of project alternatives are also calculated.0. 5.10.0.000 0. 8.0.

00 5.10) (9.00 1.2. 7) (4.00 5. 2. 5.25. 5.2) (2.4.00 5. 9) (7. 5.5. 5. 0.00 1.50 0. 2. 2.50 0. Kahraman and ˙ Kaya I.2) (2.3. 0. 5. 1.25 0. 4.00 2.2) (5. 9) (8.5.3. 7) (5. 6.00 7. 9.296 Table 10. 5) (3. 6) (5.00 2.00 2.5) (0.55) C12 1.00 1.00 5. 0. 0. 1. 4. 1.00 7. 5. 5.75 0. 8) (9.00 1.50 0.5.25 0.00 5.0. 5) (5. 9.00 (3. 9. 2.00 9.5.00 5.5) (0.00 2.00 10.00 6. 8.10. 1) (1. 8. 2. 5.75) (0. 1) (1. 1. 2.00 (0.75. 3.4.5.00 (9. 2. 6. 1. 6. 0.2. 5) (4.00 1. 9) (5. 0. 7.00 1.00 1.55.5) (0.00 5. 3.6) (1. 0.00 C13 5.00 3. 5.00 9. 5. 2. 8.50 0. 3.5.4.5.25. 9.50 0. 0.2) 5. 1.25. 1. 3.00 5.00 9.00 1.3) (2.00 8. 5) (4.4.50 3. 1. 0.00 9.5.00 6. 1.00 4. 1.25. 1) (1.00 7. 7. 0.00 2.8 Fuzzy Aggregation of Economic Criteria for Project-VI C11 Experts STFN STFN STFN Scale STFN C12 Scale C13 Scale C14 Scale C11 0.00 5.3) (2. 1. 0. 5.4.00 . 6) C14 E1 E2 E3 E4 E5 Aggregation E1 E2 E3 E4 E5 Aggregation E1 E2 E3 E4 E5 Aggregation E1 E2 E3 E4 E5 Aggregation C. 0.00 3.00 9. 5. 4. 5) (5.2.00 3. 6) (3.6. 0. 1.00 5. 0. 4.00 3.00 1. 5.2.5) (0. 1.00 1.5.00 4.00 1.5.4.00 8. 7.00 3.3) (2.4. 9) (8. 2) (1. 2. 0.2. 9.9. 0. 1) (1. 0.50 0.4.00 2.00 2.00 9. 5. 5) (4. 9.

5.11) Figure 10.71.47.91. 5.8.89) (4.98.26.13. 5.89.93.11.22) (3.05.84.71. 5.Realizing Policies by Projects Using FMCDM 297 Table 10.1) (3. 4.83.5 Conclusions Traffic congestion occurs when a volume of traffic generates demand for space greater than the available road capacity. 7. 4. 5. 4. 6. 5.93. 7. 7. 5.02) (6.58.83. 6. 6. 4. 5.59) (5. 6.55.94.1) (4. 5. 5.43.95. most of them reduce the capacity of a road at a given point or over a .9 Alternatives Project-I Project-II Project-III Project-IV Project-V Project-VI Project-VII Project-VIII Project-IX Fuzzy Scores of Alternative Projects Fuzzy Scores (4. 6. 7.03.76.99. 4.91) (5. 8.77.7 The Membership Functions of Fuzzy Scores for Alternative Projects 10.68. 7.01) (3. 7. There are a number of specific circumstances which cause or aggravate congestion.

one of them is related to ferry boat line and the last one is related to tramcar line.068 5. precise numerical values.68 4. After the Marmaray and Third Bosphorus Projects are finished. Kahraman and ˙ Kaya I. As a result the project.298 C.89 4.84 5.960 4. ˙ Istanbul Metropolitan Municipality and Republic of Turkey Ministry of Transportation and Communication are working on not only extending sea.014 5.10 6.26 5.05 4.00 a4 6. traffic congestion is expected to decrease.916 5. fuzzy numbers.580 3. and railway transportations and but also improving mass transportation. and ranges of numerical values.275 4.93 3.93 7.43 5.10 Alternatives Project-I Project-II Project-III Project-IV Project-V Project-VI Project-VII Project-VIII Project-IX a1 4.59 7.030 Dmin 5.91 7.13 4.58 5.22 5. They also discuss about precautions to avoid traffic congestion.749 5. highway.113 4. or increase the number of vehicles required for a given throughput of people or goods. The authorities of IMM should give the and “light rail system from Usk¨ u u . It has been successfully applied to determine the best project alternative to decrease traffic congestion in Istanbul.10 7. “tramcar line from Kabatas to Besiktas”.98 6.94 4.89 7. Recently. The alternative projects are evaluated by a multi criteria decision making technique which includes the fuzzy set theory that is one of the most important techniques of computational intelligence. In this study three out of the nine projects are related to subway. the proposed methodology has the ability of taking care of all kinds of evaluations from experts.093 2. which are evaluated as a result of the municipality’s policy that more investment should be made in railway systems. Traffic congestion in Istanbul is unavoidable event since around 2 million vehicles are in Istanbul traffic every day.933 7. The proposed fuzzy multiple criteria decision making methodology allows experts to be flexible and use a large evaluation pool including linguistic terms.99 The comparisons results for alternative projects a2 5.01 6. four of them are related to light rail system.013 4.03 5. These projects have been evaluated to determine the most suitable.47 6.71 4. Hence.83 a3 5. is deter¸ ¸ ¸ mined as the best alternative while the two next are “subway Line from Besiktas to Otogar” ¸ ¸ ¨ udar to Tepe¨ st¨ ”. The fuzzy set theory is a powerful tool to treat the uncertainty in case of incomplete or vague information.71 5.11 3.83 5.02 8.55 5.11 Dmax 4.754 6.296 4.892 4.76 6.80 4.997 Ranking 6 2 9 3 4 1 5 8 7 certain length.77 7.449 6.136 5. Table 10. The best policy to decrease the effects of traffic congestion in Istanbul is determined as investment for railway transportation projects.91 3.95 5.

Simpson. Multicriteria decision model for outsourcing contracts selection based on utility function and ELECTRE method. [4] Bozdo˘ an. [17] Lee G.M. Comg mission 4: Urban Mobility Management.. 2008. M.J.D.R. Kılıncarslan.F.. Wegen. Raiffa. J. 1998..H..html [5] Brans. 17... 228–38. K.M. H. 1996. Cebeci. J. 233–247.. N. Decision with multiple objectives: preferences and value tradeoffs. Multi-attribute comparison of catering service companies using fuzzy AHP: The case of Turkey. ˙ Kahraman.. Multiple attribute decision making methods and applications. Chan. Hosseinzadeh.Computational Intelligence in Complex Decision Systems 299 first priority of project for the “tramcar line from Kabatas to Besiktas”. 1985. 2005. Zaerpour.. 24. Rabbani. [10] Gharehgozli. A. and Yoon.. R. E. [16] Laarhoven.K.. D. H. 171–184.. C. W. P.J. G. 181.J. [6] Buckley. Fuzzy robust process capability indices for risk assessment of air I. 2009. (Eds.. 109–118. A. 229–241. L. 39. [7] Chang. Fuzzy Set Systems. [9] Gercek. A comprehensive decisionmaking structure for acceptance/rejection of incoming orders in make-to-order environments. Telgen.. 1976...stadtentwicklung. 1544–1555. New York: Springer-Verlag. 87.. [12] Jahanshahloo..L. A multiple criteria approach for the evaluation ¸ of the rail transit networks in Istanbul. The analytic hierarchy process (AHP) approach for asses- . 2004. R. D. B. 1016–1032. 1983. J. Vincke. H. T. 529-541.. In the future research. the other projects should be considered by taking into account the proposed order.. [15] Keeney. L. http://www. 1986. [3] Boer.. European Journal of Operational Research.de/archiv/metropolis2005/en/dokumentation/c4. Applied Mathematics and Computation. Transportation. Mareschal. Bibliography [1] Almeida. 1981. R..L.J. Stochastic Environmental Research and Risk Assessment. If it is not possible ¸ ¸ ¸ to perform the project. [13] Kahraman. Transportation Systems Of Istanbul. Pedrycz W. 203–228. Robinson... Imitating Life. 1–12. International Journal of Advanced Manufacturing Technology. L. Y. J. Marks. 8th World Congress Metropolis. P. [11] Hwang. 34. J. C. C. Karpak. Extension of the TOPSIS method for decision-making problems with fuzzy data. European Journal of Operational Research. 11. [2] Bezdek. P. 2007. 1996.. 1994. U. NewYork: Wiley. Outranking methods in support of supplier selection. Computational Intelligence.V. B.L.. 23(4). Boston. [14] Kaya... R. M.).. Fuzzy Sets and Systems. What is computational intelligence?. How to select and how to rank projects: The PROMETHEE method. Razmi.berlin. the proposed fuzzy multiple criteria decision making approach can be used to evaluate alternative projects for sea and highway projects in Istanbul.. Computational Intelligence PC Tools. European Journal of Purchasing & Supply Management. R. Academic Press.. 649–655. International Journal of Production Economics..P. Dobbins. Applications of the extent analysis method on fuzzy AHP. 95.. Izadikhah. Computers & Operations Research... Ruan.. 2006. 2004. A fuzzy extension of Saaty’s priority theory. C. 4(2-3). 3569– 3574.T.. IEEE Computer Society Press. [8] Eberhart. C. In: Zurada. J. 2008... 31. Fuzzy hierarchical analysis. pollution.

Crit` res multiples et mod´ lisation des pr´ f´ rences: l’apport des relations de e e ee surclassement. Topcu. Tunnelling and Underground Space Technology.. L. [23] Saaty.300 C. 175. B.J. [19] Marmaray Project. 4.tuik. Smith. A decision support approach for the evaluation of transport investment alterna¸ ¸ tives. Comparison of fuzzy numbers using a fuzzy distance measure.09. 589–600. International Journal of Project Management.. Information Control.. S.gov. 3. H. [27] TurkStat-Turkish Statistical Institute..ubak. Issue 15. Journal of the Operational Research Society. R. 2007. Fodor. Cognitive map: KBDSS integration in transportation planning. 331–341. ¸ I.gov. M.. 1980.. International Journal of Computers. New York: McGraw-Hill.ofmdfmni..L. www. 89(1). 2004... Article in press. 2007. D. An. T. [25] Smith. 1993. The analytic hierarchy process. 445– 455. 155–168. http://www. N. http://www. 8. 1965. 25. I. [29] Wang. 2002. ancient finds have europe-asia tunnel on nonstop delay. Duckstein..tr ¨ [28] Ulengin.uk/practical-guide-policy-making. 1–44. Application of a fuzzy based decision making methodology to construction project risk assessment. [22] Rudas.J. 2006.pdf [32] Calıskan. and Yang. Tzeng. [31] Zeng. 2007. Computers & Operations Research. Wired Magazine. http://www. [21] Roy. F. [26] Tran. J. 1. IEEE Transactions on Neural Networks. Quake fears. Intelligent Systems.gov. L. Computational versus artificial. Using a hybrid multi-criteria decision aid method for information systems outsourcing.wired.. Kahraman and ˙ Kaya I. Fuzzy Sets and Systems. 2009. Marmaray project: Tunnels and stations in BC contract. 737–739. 338–353. J. J. ˙ 1997. Social Indicators Research.tr/eubak/projects/marmaray [20] Opricovic..L. 48. L. J. Communications & Control. 1696–1704. G.... 2009. European Journal of Operational Research. Compromise solution by MCDM methods: A comparative analysis of VIKOR and TOPSIS.. [30] Zadeh. [24] Sakaeda. 2009. J.H. 3691–3700.com/science/planetearth/magazine/15-09/ff quake. 1065–1075. 130. 34. 2008. Revue d’Economie Politique.. Fuzzy sets. 132–138. 1974. [18] Marks. 156. . European Journal of Operational Research. ment of urban renewal proposals.

and a fuzzy-number-ranking-based optimization method for fuzzy multi-objective non-linear programming are developed. and Jie Lu 2 School of Mathematics and Physics.1 Models for Day-ahead Markets The load dispatch in a spot market is one of the kernel problems in an electric power market. The model provides theoretical foundations for the research on operations and decision making in the electric power market.1007/978-94-91216-29-9_11. Guangquan Zhang 2. D. yagao. a hybrid evolutionary computation method with a quasi-simplex technique. Atlantis Computational Intelligence Systems 2. A load dispatch model considers a ramp rate and valve-point-loading effects. DOI 10. Baoding 071003.edu. Australia {zhangg.Chapter 11 Evolutionary Computation Methods for Fuzzy Decision Making on Load Dispatch Problems Guoli Zhang 1. © 2010 Atlantis Press/World Scientific 301 . Computational Intelligence in Complex Decision Systems. This section introduces the basic concepts of electric power markets and builds up two load dispatch models for a day-ahead market. Ruan. a weight point method for multi-objective programming.uts. An environment/economic load dispatch model is presented to handle uncertainty factors.jielu}@it. 11. Broadway NSW 2007.au This chapter introduces basic concepts relating to a day-ahead market in a power system. China E-mail: zhangguoli@ncepu. reliability of the power system and order operation of the electric power market. but is also a key issue to assure safety. North China Electric Power University. Although a lot of achievements have been obtained. University of Technology. PO Box 123. there are still many problems to be solved for the power market operation. To solve load dispatch problems from day-ahead markets in power systems.cn 2 1 Faculty of Engineering and Information Technology. It not only relates to the benefits of every participant in the market. Ya Gao 2 . Sydney.edu.

5. we use this market structure and build two load dispatch models. Y. realize lower bid.. Therefore we propose the principle of the market clearing price determined by SMP and load dispatch calculated by PAB. the same quality electric energy should have the same electricity price in the same power grid.1. ISO takes responsibility for both network security and the auxiliary service. the congestion management and spinning reserve are controlled by ISO. A reasonable price should consider the fairness of both the merchandise pricing and the market competition. and the price is fixed for a long time. A conventional economic dispatch (CED) considers only the output power limits and the . since the previous pricing mechanism is unreasonable to represent fair trading and reflect the market status of supply and demand. many new price methods have been proposed. The basic structure of a power market [1. which is consistent with the purpose of an opening electric generation market. but they still have insufficiencies. decrease generation cost. one price corresponds to one unit. and finally benefit consumers.1 Introduction In a traditional generation electricity plan. PX takes charge of the spot trading in the day-ahead market. This mechanism combines the merits of SMP and PAB. In this study. In the practical process of electric energy exchange. 6]. Different models and techniques have been proposed in the literature [3. and J. the electricity price is determined by the generated electricity cost. SMP represents the fairness of merchandise price. Lu 11. 21] consists of power exchange (PX) and independent system operator (ISO). 11. 19. and is feasible and simple. Gao. with the basic objective of scheduling the committed generating unit outputs to meet the load demand at minimum operating cost.1. while satisfying all units and system constraints. Zhang. In other words. This mechanism encourages generation enterprise to uncover the inner potential. There are two typical electricity prices widely used in electric power markets. In this market structure. i.2 A Load Dispatch Model for a Day-ahead Market Economic dispatch (ED) is very important in power systems. One is the so-called uniform market clearing price (MCP) or system marginal price (SMP) which can be obtained by the highest bidding of the unit committed. It represents the fairness of market competition by using PAB to compute the fee of purchasing electricity and dispatching load. Both SMP and PAB are reasonable. with the main task to solve the dynamic economic load dispatch problem. The other is pay-as-bid price (PAB).302 G.e. increase competition capability. In general. G. Under an electric power market environment. Zhang. solves simultaneous fairness of the price of merchant and market competition.

Therefore. N represents the number of committed units. since the CED model does not take into account the ramp rate constraints. a power exchange-independent system operator model (PX-ISO model) has been adopted in the Chilean power market [19]. 19. its solution may not be real optimal. and the California power market [1. transmission congestion and spinning reserve. A non-smooth function sometimes has to be used to account for special factors. In general. PX administrates the day-ahead market and the ED is the major task for PX. 6. 19. a dynamic economic load dispatch (DELD) model for a PX-ISO power market can be described as follows: ⎧ T N ⎪ ⎪min f (Pj (t)) = ∑ ∑ F(Pj (t)) ⎪ ⎪ ⎪ t=1 j=1 ⎪ ⎪N ⎪ ⎪ ⎨ P (t) = P (t) + P (t) ∑ j D L j=1 ⎪ ⎪ ⎪P ⎪ j min Pj (t) Pj max ⎪ ⎪ ⎪ ⎪ ⎪ ⎩−D j Pj (t) − Pj (t − 1) R j (M1) (1) where Pj (t) is the output power of the j-th unit during the t-th time interval. Due to the inclusiveness of ramp rate constraints. The sharp increase of the number of variables implies the increase of searching dimensions. and the Australia and New Zealand power markets [1. such as the voltage rippling [23. CED problems usually formulate the objective function as smooth. which. which are solved by using equal λ rules [22]. 4. Among some of these power markets. however. 21]. 24]. In order to assure the optimization of solutions. which furthermore results in the difficulty of solving the problems of DED. T is the number of time intervals per dispatch cycle. the load dispatch model must consider the ramp rate limit. 14. In this model. Based on the analysis above. 6. such as the England and Wales power market. the number of decision variables involved in the problem will increase dramatically compared with the corresponding CED problem. constraints on the spinning reservation can be ignored in the DED model. 21]. and also including the ramp rate constraints would be highly desired. 17. Norway power market. Great efforts have been devoted to economic dispatch problems and various models have been proposed [1. Chile power market. 7. In addition. 22. On the other hand. 15. Hence. the model becomes the dynamic economic dispatch (DED). If ramp rate constraints are included. including network security. . there are different constructions and operation modes in power markets. A more accurate ED model that can account for special cost factors leading to a non-smooth objective. California power market. are not always adequate for real ED or DED problems. ISO will verify the dispatch schedule against a set of criteria. the DED model is needed. 23. 3. 13. 24].Evolutionary Computation Methods for Fuzzy Decision Making on Load Dispatch Problems 303 balance between the supply and the demand. 8.

which are not distinguished in this chapter . the first three options should be determined by the generation companies. a1 j . not by the regulatory authorities.3 An Uncertain Environment/Economic Load Dispatch Model A conventional economic dispatch problem is mainly concerned with the minimization of operating costs or purchasing electricity fee. The objective function in the above model (M1) can also be the expense of purchasing electricity. The emission of thermal units mainly includes SO2 . G. 17. NOx and CO2 . Gao. Pj min and Pj max are the minimum and maximum output power of the j-th unit respectively. 19] have been proposed to reduce atmospheric emissions. However. In other words. d j and e j are coefficients of the j-th unit. the environmental/ economic load dispatch problem considering emission of harmful gases is a kernel issue in electric power markets..1. while satisfying load demand and system constraints. and emission dispatching. and J. PD (t) and PL (t) are the load demand and network loss in the t-th time interval respectively. These include installation of pollutant cleaning equipment. Zhang. 11. The new algorithm to solve this problem will be given later in this chapter. In fact. and the emission dispatching option as an attractive short-term alternative [17]. The prospective algorithms for solving this model must have a stronger global searching capability. a2 j are constants. Therefore. i. switching to low emission fuels. Petrowski referred the first three options as the long-term ones.e. replacing the aged fuel-burners and generator units. subject to the diverse constraints in terms of units and systems.304 G. The desired long-term target is to reduce the emission of harmful gases. minimization of fuel cost and emission. A variety of feasible strategies [1. the emission of harmful gases required to generate electricity should be curtailed in accordance with laws and regulations. Zhang. d j sin[e j (Pj min − Pj (t))] represents the rippling effects caused by the steam admission valve openings. D j and R j are the maximum downwards and the maximum upwards ramp rate of the j-th unit respectively. especially in the circumstances of the electric power market. Lu and F(Pj (t)) is the generation cost function and can be formulated as F(Pj (t)) = a0 j + a1 j Pj (t) + a2 j Pj2 (t) + d j sin[e j (Pj min − Pj (t))] (2) where a0 j . Y. Some researchers pointed out that the environmental/economic load dispatch problem is to simultaneously minimize two conflicting objective functions. The model (M1) describes a non-linear programming problem with multiple local optimal points. an environmental pollution problem caused by generation has been presented in recent years.

2. it is not precise to describe these coefficients as fixed values. we present these coefficients described by fuzzy numbers. b j . Aimed at characterizing the cost and emission more precisely. The meanings of the other symbols are the same as the symbols in the model (M1). . we will develop some new algorithms to solve these problems. these harmful gases mentioned above are all functions of output power Pj . called the fuzzy dynamic environmental/economic load dispatch model (FDEELD). β j .2 Evolutionary Computation Methods and Fuzzy Decision Making The model (M1) built in the above section is a non-linear programming problem with multiple local optimal points. γ j are fuzzy emission coefficients of the j-th unit. such as: experiment errors. In Section11. and (M2) is a fuzzy multi-objective non-linear programming problem. Therefore. the coefficients of both cost function and emission function are constants. c j are fuzzy cost coefficients of the j-th unit. a hybrid evolutionary method and a fuzzy number ranking method to solve FDEELD. can be described as E(Pj ) = α j + β j Pj + γ j Pj2 where α j . However. α j . and their emission (ton/h). e is an emission function. and generally can be obtained by experiments. In general. These optimization problems are hard to solve. from which it is very hard to obtain an optimal solution. there exist many factors which affect these coefficients. is built as follows: ⎧ T N T N ⎪ ⎪min f = ∑ ∑ F(Pj (t)) = ∑ ∑ (a j + b j Pj (t) + c j Pj2 (t)) ⎪ ⎪ ⎪ t=1 j=1 t=1 j=1 ⎪ ⎪ ⎪ T N ⎪ ⎪min e = ∑ ∑ (α + β P (t) + γ P2 (t)) ⎪ j j j j j ⎪ ⎪ ⎨ t=1 j=1 N (M2) ⎪ ∑ P j (t) = PD (t) + PL (t) ⎪ j=1 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪P j min P j (t) P j max ⎪ ⎪ ⎪ ⎪ ⎩−D P j (t) − Pj (t − 1) R j j (3) (4) where a j . A new load dispatch model with uncertainty. written as E(Pj ). different operation situations. 11. β j and γ j are coefficients of the j-th generator emission characteristics. The model (M2) describes a fuzzy multi-objective non-linear programming problem. In a typical environmental/economic load dispatch model. the quality of coal and the aging of facilities.Evolutionary Computation Methods for Fuzzy Decision Making on Load Dispatch Problems 305 for reasons of simplicity. we will propose a weighted ideal point method.

On the other hand.1 Evolutionary Computation Conventional optimization methods suffer from local optimality problems and some of them require a function with good characteristics. and J. have drawn many researchers’ attention because the stochastic optimization techniques are capable of finding the near global optimal solutions without putting restrictions on the characteristics of the objective functions. G.2. Since the sum of probabilities of all the generic operations must be the unity. it suffers from premature convergence problems and usually consumes enormous computing time. which can be regarded as the exploration operation. This will weaken the capability of local search dramatically. how to create a balance between exploration and exploitation in . stochastic optimization techniques.1-5%. the level of capability of global search will increase when the probability of mutation increases. In recent years. 26]. Therefore. Zhang. the mutation probability has to be reduced to increase the crossover probability for a reasonable level of capability of local search. limit their application. crossover and mutation. such as differentiability. which can be referred to as exploitation operations. the velocity of local search increases when the probability of crossover increases. A typical CGA has three phases. A great amount of effort has been devoted to improving these methods and some of them have been successfully used in a variety of real world problems [17. which. the ability of local search mainly relies on the reproduction and crossover operations. Lu 11.e. Gao. The performance of CGA precedes the traditional optimization methods in aspects of global search and robustness on handling an arbitrary non-linear function. while the capability of global search is assured by the mutation operation. we refer to it as the classical GA (CGA). and evolutionary algorithms (EA).. However. In the process of balancing exploration and exploitation based on reproduction/crossover and mutation operations for a fixed population. with a range of 0. which consist of reproduction. Similarly. genetic algorithms (GA). although they require significant computing burdens and generally take a fairly long time to reach a solution. i. In the CGA. it is hardly possible to achieve a win-win situation for both sides simultaneously. Hereafter. to achieve a satisfactory level of capability of global search. Generally speaking. Zhang. such as simulated annealing (SA). GA was initially introduced by John Holland in the seventies as a special technique for function optimization [9]. the probabilities of reproduction and crossover have to be decreased to increase the mutation probability. evaluation and genetic operation. continuity. Y. This contributes to the fact that the probability of mutation in CGA is very low. slow down the convergence rate and make the global search ability unachievable eventually. initialization. to a certain extent.306 G.

the problem can be formalized as a pair of real valued vectors (s. which are very similar to the traditional genetic operation.2. f ). where S ⊆ Rn is a bounded set on Rn and f : S → R is an n-dimensional real-valued function. In each iteration. In short. f (xmin ) f (x) (5) . All four operations together will produce a new sub-population with the same size as the corresponding parent sub-group.1. f needs not be continuous but must be bounded. HRGAQT has the following aims: (1) we assure the capability of global search by increasing the probability of mutation. The new generation is the collection of all the newly produced sub-groups. HRGAQT first divides the population into a number of sub-populations and each sub-population is treated as a classical simplex. The problem is to find a point where f (xmin ) is a global minimum on S. HRGAQT applies four operations in parallel to produce offspring. while the local search is mainly implemented by the quasi-simplex technique and reproduction including the elitist strategy and crossover operations. The other three operations are reproduction. According to Yao and Liu. The first operation is the quasisimplex evolution in which two prospective individuals will be chosen as the offspring. the so called quasi-simplex technique: a new hybrid real-coded genetic algorithm with quasi-simplex technique (HRGAQT) is used. crossover and mutation respectively. and (3) we enhance the capability of local search by introducing the so-called quasi-simplex techniques into the CGA since the capability of local search will be significantly weakened by the probability of reproduction/crossover decrease as a result of increasing the probability of mutation. Then for every simplex. (2) mutation is implemented by using an effective real-value mutation operator instead of traditional binary mutation.Evolutionary Computation Methods for Fuzzy Decision Making on Load Dispatch Problems 307 GA-type algorithms has long been a challenge and retained its attractiveness to many researchers [10. except that the probability of mutation is fairly high. 17].1 Function optimization and quasi-simplex technique We consider the global minimization problem described by Yao and Liu [25] for the purpose of development of new search algorithms. it is required to find an xmin ∈ S such that ∀ x ∈ S. HRGAQT maintains the diversity of a population to enhance global search capability eventually because a higher diversity of population leads to a higher level of capability to explore the search space. More specially. We present a new method to enhance the capability of global search by increasing the probability of mutation operation while assuring a satisfactory level of capability of local search by employing the idea of simplex method. 11.

In this light. Zhang. the worst and the best points in terms of function values are denoted by xH and xB . Since it has been highly recognized that GA has no special request on the characteristics of the objective functions.1. To understand the HRGAQT algorithm. . we briefly introduce basic ideas of the simplex technique. denoted by xi . 2. and contraction. which can be referred to as the worst-opposite direction. Conventional simplex techniques mainly consist of four operations. the centroid xC of the polyhedron are all the points but the worst one needs to be calculated by ∑n+1 xi − xH i=1 (8) n A better point predicted by simplex techniques lies on the line starting from the worst point. and can be determined by f (xH ) = fH = max fi . The process will . The actual value ranges of α for different points are shown in Table 11. reflection. . which can speed up the local convergence rate. Gao. n + 1. such as the reflection point. . i i = 1. with function values denoted as fi . and the compression points. G.308 G. . which are not in the same hyper-plane. is to combine the CGA with conventional optimization methods. . the conventional optimization methods that go with GA should not require that the objective functions have special characteristics. i = 1. n + 1 where fH and fB denote the worst and the best function values. Y. Zhang. The simplex algorithm produces a new simplex by either replacing the worst point by a better point produced using the simplex technique or contracting current simplex towards the best point in each iteration step. Assuming there are n + 1 individuals. A simplex in an n-dimensional space is defined by a convex polyhedron consisting of n + 1 vertices. . 2. The actual location can be determined by the following formula: x = xC + α (xC − xH ) (9) where α is a constant and can be a different value for different points lying on the worstopposite direction. respectively. Simplex method is promising because it demands less function characteristics.e. To determine a better new point than the worst point xH . Lu On solving the above optimization problem by genetic algorithms. . xC = towards the centroid.. expansion. compression. respectively. . and J. 2. . . Simplex is a type of direct search method. which is a widely accepted search technique. Therefore. i = 1. . i. n + 1 i (6) (7) f (xB ) = fB = min fi . expansion points. . an effective method. we choose to combine the conventional GA with simplex technique to form a hybrid generic algorithm in which a real-value scheme and a dynamic sub-grouping are used.

11.1 Points obtained using the simplex techniques with different α α =1 x= α (xC − xH ) xC + reflection point Expansion point Compression point Compression point α >1 0<α <1 −1 < α < 0 Reflection point of xH respect to xC A point farther than the reflection point from xC Points between xC and reflection point Points between xH and xC be continuous until the termination criterion is satisfied. and the reproduction will retain the best individual by applying the elitist strategy and also produce some individuals based on the probability of reproduction (R). crossover and mutation. the simplex algorithm has a higher level of ability of local search. which can be regarded as guidance in the search landscape. The best individuals of population in the final generation will be taken as the optimal solutions. The crucial idea of the classical simplex techniques is to track the local optimal following the worst-opposite direction of each simplex. evolution starts a new iteration. The population starts to evolve.2 Hybrid real-coded GA with Quasi-Simplex techniques HRGAQT is established by combining a technique evolved from the traditional simplex technique.Evolutionary Computation Methods for Fuzzy Decision Making on Load Dispatch Problems 309 Table 11. HRGAQT initializes a randomgenerated population with μ individuals (real-coded chromosomes) and each individual has n components. This process continues until the termination criterion is satisfied. reproduction.1. At the end of each evolution iteration. In doing so. the generation is divided into a number of sub-populations (or sub-groups) with each sub-group having n+1 individuals. while its capability of local exploitation can also be reasonably high by using both reproduction/crossover and quasi-simplex techniques. Therefore. If the termination criterion is not met. HRGAQT has a number of outstanding features which enable both local exploitation and global exploration. HRGAQT adopts the dynamic sub-grouping idea to ensure each sim- . The crossover operation will also generate a number of pairs of individuals according to the probability of crossover (C) and the left-over individuals will be produced by mutation (M). Each sub-group will then evolve into a new sub-population of the same size by four operations in parallel. HRGAQT can achieve a substantially high level of global exploration by increasing the probability of mutation. all the new individuals from the sub-populations will merge together and evolution enters new generation.2. The quasi-simplex operation (QS) will generate two new individuals. which are quasi-simplex operation. The process of HRGAQT can be described as follows: First. At the beginning of each iteration. which is referred to as a quasi-simplex technique with the CGA.

11.310 G. The detailed discussion about the size of a sub-population and the number of sub-populations to be used will be presented in another paper. Y.2. We refer to this direction as the best-forward direction. Zhang. GA usually uses a very small mutation probability.05.3 A new mutation operator To guarantee the local search effect. Although the computation time for each iteration may increase due to the partition process. xm and xn can be determined by the value of β in (10) and the range of β is shown in Table 11. Lu plex consists of reasonably correlated individuals in the entire evolution process to enhance the convergence rate. Gao. A typical mutation probability ranges from 0. expansion and compression operations along with the worst-opposite direction. Secondly. Three prospective individuals xe . The quasi-simplex technique also expands the conventional simplex technique by looking at the prospective individuals lying on a line starting from the centroid towards the best point of the simplex. The mutation operators have two . HRGAQT implements population partition different to the strategies proposed in the literature by two methods. It produces four prospective individuals using the reflection. Zhang.2. To avoid a situation in which too many individuals are similar so that the diversity of the population decreases dramatically. HRGAQT employs the quasi-simplex technique with ancillary reproduction and crossover operation to assure the local exploitation. G. and J. HRGAQT divides a population into a number of sub-groups with each sub-group consisting of n+1 individuals to ensure the search validity and efficiency in terms of computing times. The quasi-simplex technique absorbs the idea of classical simplex techniques to perform a guided search.1.001 to 0. in contrast with the worst-opposite direction. HRGAQT selects the best one from the two prospective individual groups along the worst-opposite and the best-forward directions to produce two new individuals as a part of offspring. xm and xn will be produced along the best-forward direction by the expansion and compression operations using the following formula: x = xB + β (xB − xD ) be calculated by xD = n+1 i=1 (10) where xD denotes the centroid of the remaining points except for the best point xB and can ∑ xi − xB n (11) The points xe . the enhancement in the convergence rate could decrease the number of iterations needed. The other method is to make a partition for each iteration. One is to take into account the dimension of individuals on deciding the number of sub-groups.

xi j can be expressed as a sequence of decimal numbers including the decimal point: xi j = diw1 diw2 · · · di j p •di j1 di j2 · · · di jq j j w f f f (12) where superscript w and f denote the integer part and the fractional part respectively. . .Evolutionary Computation Methods for Fuzzy Decision Making on Load Dispatch Problems 311 Table 11. xm and xn Range of β Calculated point xe xm xn β >1 β =1 0<β <1 kinds in real-coded GA: one is to generate a new random real number within the domain and the other is to add a new random real number to the original one. μ . . In a real-coded scheme.2 Formula x = xB + β (xB − xD ) Range of β in (10) for xe . f2 . . . . or become smaller when they are big enough ( 5). Based on this idea. and p and q are constants representing the number of digits in the integer part and the fractional part for a given xi j . 2. we propose a new real-coded mutation operator. Both of these two operators lack support from the principles of biological natural mutation processes. w p . the new sequence after the mutation can be represented as: xi j = diw1 diw2 · · · di j p • difj1 difj2 · · · di jq j j where each decimal digit is determined by: dirj = 9 − dirj if di j is selected or dirj = dirj if di j is not selected where r = w1 . . n. Suppose xi j . a gene often changes dramatically after it is mutated. . In application. where μ is the size of a population and n is the dimension of each individual. . p is determined by the maximum value that this sequence can represent. According to the principles of a natural biological mutation. . which is described as follows. these digits should also change significantly after a mutation operation. w2 . . . 2. j = 1. f1 . In a process of biological evolution. If the digits in the sequence are randomly selected to undertake a mutation. q is determined by the precision required by the problems and its maximum value will be determined by the hardware used in computing. . (15) (14) w f (13) . . i = 1. represents the j-th component in the i-th individual. In real-coded GA. fq . decimal digits are used to represent genes. respectively. . . they should become bigger when they are small enough (< 5). In other words.

Lu 11. Step 2.3 Combine the individuals obtained from steps 2. . . .1 Produce two new individuals using quasi-simplex techniques. Step 2.1. . .4 Remove S from the original population. reserve the best one in the sub-population to be a part of offspring. Step 2 Divide the population X into K sub-populations with each sub-group consisting of n + 1 individuals. Step 3.2. .312 G. . . . .4 Crossover operation is processed as follows: Select [(n − 2)PC 2] pairs of parents randomly. and J. in the target sub-group (sorted by descending the fitness) calculated by the following formula Pi = 1 rank(xi ) − 1 η − 2(η − 1) · n+1 n (16) where η > 1. . The subscript m1 and m2 are two random numbers between .2 to form a subpopulation S. xm2 . Gao. where [] is an operator producing the maximum integer which is less than or equal to the operand.5 Repeat Steps 2. Step 3. . Step 2. xn (17) (18) where the superscripts i and j denote the i-th and the j-th individual in the population respectively. The reproduction probability of the i-th individual xi . xi . xi . xi = xi .2 Implement elitist strategy. Zhang. . . . xi . . i. G.3 Produce new individuals by reproducing by linear ranking. Step 2. Step 3 Each sub-population evolves into a new group. x2 . . . .. . Step 3. Step 2.4 for the sub-population until no individuals are left. Step 3. . Zhang. xm1 . . .4 HRGAQT algorithm procedure The HRGAQT algorithm can be outlined in the following steps: Step 1 Initialize a random population X with size μ = K(n + 1).1 Select the best individual x from the population X. xi n 1 2 m1 m2 j j j j j x j = x1 .1 and 2. Y. For every pair of the selected parent.1 – 2.e. which can be determined by the desired probability of the best individual. .2 Select n individuals which are most close to x in terms of their Euclid distances.

. . . . To describe the proposed WIPM method. xm2 . xi . . . which doesn’t require any interaction.2 A Fuzzy Multi-object Non-linear Optimization Method 11. It introduces the concept of a reference point suggested by decision makers and presents some desired values of the objective functions. .2. xi . . . fimin = 0. . xm1 .1 A weight idea point method of multi-objective optimization problems Both weighting and reference point methods are all powerful methods to achieve Pareto optimal solutions for multi-objective non-linear programming problems. . It is very hard to determine weightings and reference points in applications. we propose a weighted ideal method (WIPM) as follows: let g(x) = w1 where fimin = min fi (x). . . 2. k. . . It is hard to know the magnitude of effect of the set of weights to each objective function value. For each individual. xn m1 m2 (19) (20) Step 3. fk (x)) x∈S (21) where f1 (x). This section proposes a new weighting ideal point method (WIPM). . fk (x) are k distinct objective functions and S is the constrained set defined by S = {x ∈ Rn | g j (x) 0. . . . . .2. the weight method only represents the relative importance of goal values from an objective rather than from different objectives. . x∈S min f1 − f1 min f1 2 + · · · + wk min fk − fk min fk 2 (23) . f2 (x). . . xi . . a new individual will be produced by (12)–(15).Evolutionary Computation Methods for Fuzzy Decision Making on Load Dispatch Problems 313 1 and n. Strictly speaking. we write a general multi-objective non-linear programming problem as: min f (x) = ( f1 (x).2. . . m} (22) In this section. . . . . . . x2 . i = 1. . j = 1. and the interactive approach increases computing burden heavily. and can predict the magnitude of effect of the set of any weights to each objective function value. . . 11.5 The remaining individuals will participate in the mutation operation. . . The two new individuals will be: j j xi = xi . . The reference point method is a relatively practical interactive approach to multi-objective optimization problems. xi new 1 2 n j j j j xnew = x1 . .

In the model (24). One of the methods to solve fuzzy optimization problems is the maximum satisfaction factor method [11]. Lu min min f min = ( f1 . then ∗ min min f ∗ − f1 f2 − f2 ≈ 2 1 min . . the weights given in WIPM can reflect the trade-off rate information among the objective functions. To get the Pareto optimal solution of the problem (21). all objectives are converted into the same magnitude by the formula fi − fimin . We will give a solution based on the weight idea point method and the fuzzy number ranking. if w1 = 2w2 . fimin We can therefore predict the effect quantity of the set of weights to each objective function value. x∗ is the optimal solution of (23). w = (w1 . . Now we need to solve a single objective fuzzy programming problem (SOFPP). . wk ) > 0.2. Y. We then compare the function values of different solutions by the method of ranking fuzzy numbers. In this section we do not use the above methods. 10.2. 15. ∑k wi = i=1 1 is a weight vector. Zhang. fk ) is a so-called ideal or utopia point. . .2 A weight idea point method of fuzzy multi-objective optimization problems A problem becomes a fuzzy multi-objective non-linear programming problem if the objective function fi includes uncertainty represented by fuzzy numbers in the multi-objective non-linear programming problem (21). but directly apply HRGAQT to search for optimum solutions. . The definition below comes from Lee and Li [15]. . Gao. . In other words. 11. it can be transferred to solve the single objective optimization problem below min g(x) x∈S (24) Since the values of different objective functions in (21) can be very different. G. Different methods for ranking fuzzy numbers have been proposed [5. For example. . Another one is to convert a fuzzy optimization problem into several classical optimization problems. 19].314 G. When the non-linear objective functions are fuzzy functions. min f2 f1 where fi∗ = fi (x∗ ). i = 1. and J. 2. it is hard to know the magnitude of the effect of the set of weights to each objective function value. Zhang. we also use (23) to convert (21) into a corresponding single objective fuzzy optimization problem.

(2) In solving the single objective optimization by using HRGAQT. n). Let a.1. m. then directly compare the function values of different solutions by the fuzzy ranking method given above. . b ∈ F(R) be two fuzzy numbers. The definition of ranking two fuzzy numbers is as follows: a or m(a) = m(b) and σ (a) b if m(a) < m(b) (25) σ (b) (26) where mean m(a) and standard deviation σ (a) are defined as xa(x)dx m(a) = s(a) (27) a(x)dx ⎞1 2 ⎛ ⎜ σ (a) = ⎜ ⎝ s(a) x2 a(x)dx s(a) a(x)dx s(a) − m(a) ⎠ 2⎟ ⎟ (28) where s(a) = {x | a(x) > 0} is the support of fuzzy number a. for each individual we compute fuzzy function values according to the fuzzy number operation principle. Step2 Solve the single objective optimization (24) by HRGAQT. 1 m(a) = (l + m + n) 3 1 2 σ (a) = l + m2 + n2 − lm − ln − mn 18 (29) (30) The main steps of the weight idea point method for fuzzy multi-objective optimization problems are as follows: Step1 Convert problem (21) into a single objective optimization by using (23) and (24).Evolutionary Computation Methods for Fuzzy Decision Making on Load Dispatch Problems 315 Definition 11. Note: i (1) fmin can be given by a desired value or determined by solving the corresponding single objective optimization problem by HRGAQT. For a triangular fuzzy number a = (l.

Table11.6 8. It is obvious that the standard deviation is small and the results are believable.00284 0. PD (t) is the load demand. The best result occurred in the 7-th time. Gao.6 gives optimal power output of units and total cost corresponding to the best results. Table 11.00056 0.084 0. The experiments show that the proposed method with hybrid real-coded generic algorithms and the quasi-simplex techniques is very effective and the results are convincing. Because the objective function is a high-dimensional function with multi extremum points.74 8. Zhang.00324 0.74 7.10 8.00324 0. T is time interval.74 7. In Table 11.00028 0. .4.74 7.3.1.00324 0.6 8. and the optimal power output of units and total cost corresponding to the best result is listed in Table 11. Zhang. the mean value and standard deviation of 10 results obtained by the proposed algorithm running independently 10 times.00284 0. G.4.1 An example of load dispatch in a day-ahead market To test the effectiveness of the proposed HRGAQT in solving a DED problem (M1).74 7. Y.6 c 0.084 D 60 50 50 40 40 40 40 40 40 30 30 30 30 R 50 35 35 30 30 30 30 30 30 25 25 25 25 11.3 Illustrations on Load Dispatch for Day-ahead Market 11.3.00284 0.00324 0. the mean value and standard deviation of total cost corresponding with optimal outputs would be significant and convincing.042 0.74 7. and J.3.084 0.063 0.00056 0.3 Technical data of units Unit No. 1 2 3 4 5 6 7 8 9 10 11 12 13 Pmin (MW) Pmax (MW) a 00 00 00 60 60 60 60 60 60 40 40 55 55 680 360 360 180 180 180 180 180 180 120 120 120 120 550 309 307 240 240 240 240 240 240 126 126 126 126 b 8. The data of the unit techniques and predicted load demands in each dispatch period are listed in Tables 11.063 0.042 0.063 0.3 and 11.084 0.00324 0.063 0. In order to demonstrate the effectiveness of the proposed algorithms.00284 d 300 200 200 150 150 150 150 150 150 100 100 100 100 e 0.1 lists the optimal total cost.035 0. Experimental results are as follows.00324 0.6 8.063 0.10 8. it is unknown where the real optimal solution is.316 G.10 7.063 0. Lu Table 11. a typical dynamic dispatch case consisting of 13 committed units and 24 time intervals is chosen.

4 T PD (t) T PD (t) 1 1550 13 1910 2 1500 14 1830 3 1520 15 1850 4 1540 16 1880 Load demands in different time intervals 5 1600 17 1920 6 1680 18 2150 7 1780 19 2370 8 1880 20 2280 9 1950 21 2130 10 2010 22 1950 11 1970 23 1790 12 1970 24 1670 Table 11. Finally. we use the HRGAQT to optimize the Lagrange function.12 shows results obtained by the proposed algorithm through 10 independent runs. Tables 11. the mean value and standard deviation 1 461839 46113. Table 11. In addition. Table 11. In order to demonstrate the effectiveness of the proposed algorithm.Evolutionary Computation Methods for Fuzzy Decision Making on Load Dispatch Problems 317 Table 11. the fuzzy number ranking method is used to compare fuzzy function values of different points for the single objective function. HRGAQT and fuzzy number ranking methods.5868 6 461540 2 461138 3 460893 4 461382 5 460562 7 459898 8 461779 9 461134 10 461182 11. cost function.3. respectively. fuzzy emission and fuzzy total cost corresponding with the optimal outputs are tested. . fuzzy emission and fuzzy total cost.2 An Example of Uncertain Environment/Economic Load Dispatch In this section. we calculate three group weights.10 show the test data of the units output. emission function.47 582. We then use the Lagrange relaxation method to form a Lagrange function. and therefore it is hard to know where the global minimum point is.11 lists the means and standard deviations of fuzzy fuel cost. Penalty function h is a high-dimension non-linear function. In the process of the iteration.5 i-th time i-th result Mean Value Std Dev i-th time i-th result Mean Value Std Dev The total cost. we solve an uncertain environment/economic load dispatch problem (M2) by using the WIPM. in order to compare the magnitude of effect of the set of weights to fuzzy fuel cost and fuzzy emission. We convert (M2) into a single objective optimization problem by using WIPM.7–11. the mean and standard deviation of fuzzy fuel cost. and load demand.

96 112.80 55.97 159.02 114.92 429.56 92.74 113.79 108.11 152.87 109.64 519.40 80.40 120.00 55.57 448.94 224.00 93.27 77.58 60.00 55.17 109.02 123. G.90 77.20 78.14 124.46 G.04 110.36 230.85 61.50 108.31 76.73 234.88 120.40 236.67 55. Gao.15 81.00 55.11 224.14 112.86 122.01 40.36 209.33 450.00 77.22 219.55 238.67 113.42 98.76 89.29 109.87 117.49 244.86 113.41 129.86 359.62 110.12 76.95 147.71 170.72 79.64 120.42 118.40 92.02 224.88 98.11 233.20 116.75 77.43 448.85 224.56 113.40 224.68 155.38 114.22 111.03 113.00 55.35 111.81 78.91 150.86 94.37 447.55 245.06 214.12 55.47 98.49 46.45 121.32 55.41 97.07 359.86 115.90 99.22 112.40 77.59 111.06 153.48 57.76 147.40 77.86 109.27 98.96 470.80 448. Lu . Y.47 298.96 108.91 54.43 114.47 42.28 55.78 92.96 116.37 55.00 92.26 75.84 158.52 459898 185.11 143.75 95.27 191.71 114.23 234.88 103.98 120.31 92.77 448.66 239.02 92.43 147.66 159.31 360.50 72.51 110.03 95.95 218.82 77.65 67.74 194.02 40.11 110.75 120.80 449.27 113.86 109.85 78.85 93.68 96.98 92.12 234.78 96.34 91.47 40.54 60. Zhang.45 78.41 96.80 451.84 118.47 114.74 152.72 92.32 77.65 92.36 109.00 55.97 94.92 111.04 355.24 247.52 114.43 76.67 71.88 244.89 126.55 91.82 93.20 75.41 224.60 131.77 132.00 55.37 220.55 40.10 224.35 224.41 115. and J.20 77.68 115.05 93.66 78.60 109.91 114.66 109.65 107.06 110.99 84.89 113.00 48.57 94.78 84.71 120.43 113.39 112.19 136.57 154.42 109.90 42.15 539.15 4 5 6 7 8 9 10 11 12 13 Unit number 1 2 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Total cost 359.24 110.99 109.09 75.25 40.83 40.14 108.62 89.43 113.77 126.61 77.04 117.41 67.10 106.02 110.45 117.81 287.35 114.01 538.24 251.70 76.09 115.87 91.81 108.72 76.39 171.00 63.18 448.05 80.89 77.05 174.00 55.53 285.6 Optimal power output of units and total cost Time Interval 3 94.06 227.05 93. Zhang.07 240.23 440.63 112.44 64.44 111.86 104.82 109.95 94.28 112.14 150.00 75.22 236.31 188.58 247.36 95.31 99.01 109.02 55.87 109.05 122.68 56.67 51.07 55.04 121.35 89.76 195.66 109.80 116.318 Table 11.39 69.91 448.75 108.98 57.00 55.25 99.75 60.41 92.43 52.03 120.87 288.08 452.34 114.53 95.95 126.67 109.01 55.40 82.79 125.19 152.87 109.07 114.50 75.44 111.68 114.14 120.22 358.34 111.92 150.37 224.00 55.01 244.38 158.37 234.17 147.60 92.14 224.58 64.48 114.60 355.90 251.72 47.

12619 0.65170 a2 846.58126 38.00392 0.08062 37.53245 −0.84250 1610.32673 38. PD (t) represents the load demand of the correspondence to the time segment.10.03762 40.02013 c2 0.92147 α1 15.70591 37. the results are believable.00698 0.73960 1295.00712 0.29276 −0.00725 0. and MTC present the means of the fuel cost.51721 39.00464 0.00400 0.28456 0.50298 −0.46468 43.29979 0.18178 34. 1 2 3 4 5 6 7 Fuzzy coefficients of the emission function α0 15.02521 0. STDEV-FC.00712 0.00479 In Table 11.02478 0.02651 0.29901 36.89357 1555.86167 c0 0.03472 0.52011 b1 38. STDEV-EC and STDEV-TC present corresponding standard deviations.62881 38.53973 42. the emission. T represents a time segment.51116 −0.72578 645.29276 0.53762 40.00481 1269.00472 0.29432 43.01653 0.18178 15.49967 1168.52201 β2 0.75660 α2 16.38295 β0 0.58267 42.89553 41.83217 37.58267 35.51366 γ0 0.60992 37.54069 1315.00725 0.00382 0.09409 1519.36892 661.26542 b2 39.10148 b0 37.92507 39.8 Fuzzy coefficients of the cost function Unit No 1 2 3 4 5 6 7 a0 800.04260 36.01979 c1 0.00468 0.89710 1198.65132 35.83637 36.00392 0.Evolutionary Computation Methods for Fuzzy Decision Making on Load Dispatch Problems 319 Table 11. Pmin (MW ) Pmax (MW ) Dj Rj 1 20 125 40 30 2 20 150 40 30 Limits of unit output and ramp rate 3 35 150 40 30 4 35 210 50 40 5 130 325 60 50 6 120 310 60 50 7 125 315 60 50 Table 11.85286 37.02559 0.96538 1107.15813 0.53896 42.92147 40.03534 0.95401 625.02754 0.02705 0.54136 −0.52201 −0.86520 1586.69310 34.02043 Table 11.00472 γ2 0.11.51760 −0.7 Unit No.65920 1496.65132 15. The model (M2) is a new environmental economic load dispatch model which considers .52816 −0.53245 β1 0.69310 42.71867 a1 825.00461 0.16559 0.29979 −0.29432 36.9 Unit No.16218 0.00453 0.61504 1222.003822 0.46062 41.00400 0.32513 1135.51366 −0.12359 0. It can be seen that the fuel cost decreases and the emission increases when the weight of the fuel cost increases.52816 −0.28456 −0.51760 −0. In Table 11. MFC. MEC.00479 0.45826 1158.04260 16.01682 0.38295 42.32782 39.00464 γ1 0.52138 −0. and the total cost respectively.74602 1466.90654 36. As the standard deviations of every result are all significantly small.54136 −0.12050 0.00698 0.01707 0.75660 41.03403 0.

67 4.69 1. The fuzzy minimum tells not only the approximate fuel cost and emission.9 STDEV-FC MEC STDEV-EC MTC STDEV-TC .4 1073181 1098320 1118805 468.7 12041.5) 1061711 1086218 1106213 472.25 12596.3 1.8 377.7 310.10 Load demands in different time intervals T PD (t) T PD (t) 1 690 13 890 2 670 14 890 3 670 15 930 4 680 16 970 5 730 17 930 6 800 18 950 7 870 19 1070 8 840 20 1040 9 890 21 950 10 920 22 850 11 950 23 760 12 910 24 730 the uncertainty in coefficients of the fuel cost and emission functions. Lu Table 11. Zhang.5.4 (0. and J.55 2. Table 11. Zhang. G.9 (0.81 12539. Compared with other fuzzy multi-objective programming methods. and can get a more accurate FDEELD model(M2). 0.5 540.2 5.1 615.7.2 2.23 11993.w2 ) MFC The comparison of the results obtained for different weights (0.320 G.4 303.3) 1053936 1078110 1097695 57 58.8 312.5 1065537 1090295 1110440 55 58.6 607. hybrid genetic algorithms with quasi-simplex techniques and fuzzy number ranking method are developed and used to solve the optimization problem described in model (M2). Y.2 60.9 11466.11 (w1 .3 9.3 60 11600.7 1078780 1104148 1124838 290. Gao.7 300.5 2. the proposed method has three main advantages: (1) To describe the coefficients of the fuel cost and emission functions by fuzzy numbers can more precisely characterize the fuel cost and emission.4 11423. (2) The results described by using fuzzy numbers can provide more information than real numbers.7) 1067359 1092154 1112300 291. The weighting ideal point method. 0.3. but also the dispersivity of the minimal fuel cost and emission.12 12744.4 1. 0.89 12184.

96 153.79 4 88. because the solution of minimum dispersivity has been chosen when candidate solutions have approximately the same total cost.33 76.80 162.43 133. and based on the analysis of market competition mode.92 141.79 167.30 49.73 80.52 120.07 86.69 109.97 104.67 89.07 164.36 190.05 49.42 113.19 130.26 169.29 110.07 92.47 106.35 94.42 152.77 90.77 110.70 94.29 170.51 174.03 145.46 49.26 116.17 181.33 110.89 133.72 169.66 69.36 171.82 68.66 107.49 114.57 183.30 195.04 88.76 75.59 111.83 116.67 165.43 90.43 5 136.44 174. 0.85 176.80 98.06 68.42 195.53 82.21 112.29 165.45 160.88 132.3.20 171. this chapter proposes a competition mode of “SMP plus PAB”.17 98.68 93.24 89.41 131.20 153.75 194.67 180.20 132.66 112.12 140.27 181.11 66.07 174. Table 11.07 174.70 181.6 12544.97 62.36 100.02 89.22 87.12 Optimal power output of units for weights (0.30 7 134.57 99.99 95.96 133.51 174. and develop a fuzzy environment/economic load dispatch mode by .58 165.74 166. we analyze the uncertainty of power generation cost function and harmful gas emissions function.03 105.96 175.40 86.75 189.42 71.4 Conclusions and Further Research Aiming at power markets characterized by the PX and ISO structure.97 172.52 50.07 130.08 129.90 81.83 160.22 61.82 172.99 88.6 11998.55 105.17 89.73 116.83 Total cost 1078220 1103570 1124240 6 134.37 172.11 141.45 112.78 160.87 85.91 169.50 159. As the pollution caused by power generation becomes an increasingly urgent issue.41 57.22 175.86 97.59 50.79 103.25 94.84 165.09 87.66 119.12 116.49 141.19 168.41 80.99 132.38 93.04 115.8 3 91.49 147.96 88.13 86.15 111.19 111.03 87.Evolutionary Computation Methods for Fuzzy Decision Making on Load Dispatch Problems 321 (3) The optimum solution obtained is steady and trustworthy.23 Emission 11427.55 119.11 176.03 119.48 126.51 177.7) 2 53.45 110.65 78.10 141.36 106.30 11.27 56.44 168.15 200.12 Time segment 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Fuel cost 1066800 1091570 1111700 Unit number 1 51.22 169.78 176.85 163.94 89.94 111.48 140.83 118.21 90.39 114.23 128.34 131.33 86.92 161.93 147. and establishes a load dispatch mode which considers both ramp rate and valvepoint-loading effects.57 92.18 87.21 49.34 55.

To solve fuzzy dynamic environment/economic load dispatch problems. Goodwin. To solve load dispatch problems. [2] Albuyeh. Y... Scheduling Energy and Ancillary Service in the New Ontario Electricity Market. to develop bidding models. 161–165. (1999). Ma. Security-constrained Bid-clearing System for the New Zealand Wholesale Electricity Market. IEEE Engineering Society Winter Meeting.. A. Payman. pp. [3] Alvey. IEEE Computer Applications in Power. and Alaywan. K. Functional Requirements of Energy and Ancillary Service Dispatch for the Interim ISO New England Electricity Market. 411–416. IEEE Power Industry Computer Applications Conference. 17. This method imposes no special requirement on objective functions. pp. Environmental/economic power dispatch using multiobjective evolutionary algorithms. which can make the most of available information and give more powerful assistance to decision makers. G. T. and Sheble G. S. B. pp. B. [8] Gao F. Sydney. A new approach for ranking fuzzy numbers by distance method. [6] Cheung. We will apply multiple objective techniques and fuzzy non-linear programming technique on our current methods. pp. Probabilistic Methods Applied to Power Systems.322 G. 269–273. We also present a fuzzy-number-ranking-based optimization method. J. Lu fuzzy set theory to minimize power generation costs and harmful gas emissions. pp. A hybrid EP and SQP for dynamic economic dispatch with nonsmooth fuel cost function. Multiple leaders multiple followers bilevel game models will be studied for more practical methods. E. (1998). P. (2001). H. (2003). 30–34. 18. Fuzzy Sets and Systems. Zhang. Z. Tanaka. Kim. which converts the FDEELD problem into a single objective fuzzy optimization problem.. and Hasegawa. (1999). IEEE Transactions Power Systems. Stochastic Optimization Techniques for Economic Dispatch with Combined Cycle Units. which combines an evolutionary method and the quasi-simplex technique to improve the convergence rate and global search capability. J. Experiments reveal the effectiveness of this method. M. 13. we propose a weight point method. (2002). I. S. [4] Attaviriyanupap. and has very wide applications. and J. H. pp. [5] Cheng. and Cohen. 1529–1537. pp. (1998). 12. together with Game theory. D. 95. 307–317. Kita. 2006. PMAPS . we develop an evolutionary method. A. IEEE Transactions on Power Systems. IEEE Transactions on Power Systems. Zhang. [7] Danai. Gao. F. we will use bilevel programming technique. Bibliography [1] Abido. and Asteriadis.. X. C. we will focus our future research as follows: Considering the profit for both generating companies and power corporations at the same time. W. Based on the research in this chapter. 340– 346. California ISO formation and implementation.

P. Evolutionary Programming Made Fast. G. D. and Lu. (1996). F. IEEE Transactions on Power Systems. Vol. pp.. L. C. A. IEEE Trans. C. (2000). D. C. A clearing procedure as a niching method for genetic algorithms. pp. McCalley. pp. Bagchi. 13. J. P. (2003). Evol. pp. IEEE Evolutionary Computation. 2002 IEEE. M. P. [14] Li D.. and David. 112–118. and Davis. Economic dispatch with network and ramping constraints via interior point methods. R. Evolutionary programming based economic dispatch for units with non-smooth fuel cost functions. Comparison of fuzzy numbers based on the probability measure of fuzzy events. 251–261. Comput. 1325–1332. C. and Sheble. M. 3. T.. Towards more cost saving under stricter ramping rate constraints of dynamic economic dispatch problems-a genetic based approach. (1996).. [13] Li Y. pp. P. 2003 IEEE International Conference on. Formulation of fuzzy linear programming problems as four-objective constrained optimization problems. 24. (2002).C. E. International Journal of Electrical Power & Energy Systems. [15] Lee. 221–225. Comput. Risk-based Unit Commitment. B. 4. pp. .. IEEE Trans. Applied Mathematics and Computation. Math. Y. and Rodrigo D. J. Wu. [11] Huang. Application of the Power Exchange-Independent System Operator Model in Chile. R. [18] Rughooputh. 433–439. S. in Proc. P. Clements. pp. 2007. pp. 139. 39th North American. [10] Herrera. L. Power Symposium. 798–803. Industrial Technology. and Williams. MI: Univ. (1993). T. pp. IEEE Transactions on Power Systems. Das S. (1975). (2003). F. A. H. 1715–1721. W. and Wollenberg. X. International Conference on. M. and Padhy. S. pp. and Huang. 82–102. Genetic Algorithms In Engineering Systems: Innovations And Applications. Operation and Control.. Gnanadass. Pahwa A. (1997). Comput. J. A New Optimal Dispatch Method for the Day-Ahead Electricity Market Using a Multi-objective Evolutionary Approach. 15. F. and Huang. 18.K. pp. 43–63. A. N. 12. pp. IEEE Transactions on Power Systems. (2003). J.. G. 688–697.Computational Intelligence in Complex Decision Systems 323 2006. pp. R. Bi-objective power dispatch using fuzzy satisfaction-maximizing decision approach. 11. and Ryan S. H. Yang. 383–399.1–8. [19] Venkatesh. [12] Irisarri. [23] Walters. L. [17] Petrowski. NAPS ’07. (1999). IEEE Transactions on Power Systems. [25] Yao. R. and King. pp. T. D. (1988).. IEEE Transactions on Power Systems. A. (2007). Kimball. F. Environmental/economic dispatch of thermal units using an elitist multiobjective evolutionary algorithm.. B. G. 1–7. Coordination of bidding strategies in day-ahead energy and spinning reserve markets. 48–53. Power Engineering Society General Meeting. Evol. H. and Liu.. (1997). Gradual distributed real-coded genetic algorithms. 236–242. L. L. Ann Arbor. h. and Lozano. 3.. [21] Wen. Yang. Atienza. (1998). Comparison and application of evolutionary programming techniques to combined economic emission dispatch with line flow constraints. IEEE. H. [24] Yang. Power Generation.. P. [16] Morgan. Y. Genetic algorithm solution of economic dispatch with valve point loading. M.D. pp. Appl. 8. A. [26] Zhang. [22] Wood. pp... 1392–1396. pp. Remias. K. (1996). (2002). 1. [9] Holland. and Rudnick. 887–896. Power Engineering Society Summer Meeting. and Li. C. F. Adaptation in Natural and Artificial Systems. (2007). [20] Watts.

uk Ambient Intelligence (AmI) systems rely on the use of environmental information to seamlessly assist users in their daily lives. BT37 0QB Newtownabbey. © 2010 Atlantis Press/World Scientific 325 . and Juan Carlos Augusto 2 n ı Dpto. A revolution is taking place in Computer Science driven by the availability of affordable sensing and actuating devices that can help a computing system to diagnose specific situations of interest and act consequently. University of Murcia. we describe a smart TV service which combines different information about viewers’ presence and preferences and TV programs data to elaborate a list of recommending programs.1 Introduction Technological advances allow now the realization of a dream which decades ago was confined to the science fiction genre. University of Ulster. 12. Computational Intelligence in Complex Decision Systems.es 2 1 School of Computing and Mathematics.augusto@ulster. by means of argumentative techniques and several meta-decision criteria. Then. DOI 10.1007/978-94-91216-29-9_12. Juan A. the system is able to assist through mediation in the context of conflicting preferences or needs. the potential occurrence of conflicts derived from the use of this information becomes higher. juanbot}@um. the AmI systems need a mechanism to decide upon these conflicting situations in order to still give an appropriate response. Atlantis Computational Intelligence Systems 2. As the development of AmI architectures is involving a larger amount of users and devices information. These conflicts are usually represented as inconsistencies between different pieces of information. Bot´a 1 .ac. Spain Email: {amunoz. de Informaci´ n para la Ingenier´a y las Comunicaciones. o ı 30100 Murcia. As a result. Several arguments for and against recommending a same program are presented to the AmI system. all autonomously and unobtrusively [1]. United Kingdom jc. Ruan. More specifically.Chapter 12 Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants Andres Mu˜ oz 1. In this chapter we propose the use of Argumentation as a methodology for automatically detecting and solving these conflicting situations. These computing systems are grouped generically under the term D.

C. people will have a device connected to Internet such that they can take their blood pressure and transmit that to a doctor’s computer who will monitor the progress of that parameter. Let us consider Smart Homes [11] as this is one of the areas more intensively explored and the one (together with Smart Cars) that is succeeding to reach a significant number of people. Smart Offices. Smart Cars. Smart Classrooms to name those that so far have attracted more attention. Researchers in this area progressively came into contact with different types of sensors that opened up the range of problems they were taking care of. and J. One such common hypothesis is that there is only one person in the house or at . Pendants with an emergency call button where offered as a necklace or a pull down cord will be placed in the bathroom allowing people who felt unwell or unsafe to transmit an alarm via a phone line to a response center which will call by phone and/or send a qualified person to visit the location where the alarm was issued. There is a substantial body of literature reporting interesting progress in this area [12]. Governmental organizations at political and health-care level realized of the potential to restructure health-care provision and decentralize the service shifting the care from the hospital to the home. Mu˜ oz. Bot´a. Researchers in those areas were mainly focused on building isolated systems that will help people to measure vital signs or to provide some level of speedy response to a problem. Intelligent Environments understood in the sense described above can be succinctly defined as: a digital environment that proactively. 5. For example. 4. 8] and the physical infrastructure (that also has some degree of intelligence embedded) is often referred as Smart Environment [9]. J. This new paradigm can be exercised in many ways and examples of systems being explored at the moment by the scientific community and some leading companies are: Smart Homes. The development of this area has been so far largely dominated by a progression of pre-existing communities that were developing technology for telecare/telehealth. but sensibly.A. 6.326 A. Augusto n ı Intelligent Environments because at the core of each of these artificial systems a software module is providing both an intelligent analysis of the information gathered through the sensors and intelligent-decision making that will drive the actuation of the system upon the environment. For example a bed occupancy sensor allows to estimate how often a person was getting up from bed during the night and a combination of these occupancy sensors coupled with PIR (Passive InfraRed) and RFID (Radio Frequency IDentification) sensors allowed to gather substantial information of the whereabouts and habits of people in a house. Most of these solutions however are still at a seminal level and relies on important infrastructure constraints. 3. supports people in their daily lives [10]. 7. The intelligent component is often termed Ambient Intelligence (AmI) [2.

Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 327 least only one to care of (e. programs which have an advisory “+18” parental guidance (PG) could be discarded from the list. breaking news.). they must now cope with hundreds of different channels and programs. It is this latest scenario that will be the focus of this chapter. the viewers wear these tags and they are automatically detected by the recommending service when they stay in proximity of the TV. and the user selects the current viewers through the TV remote control. Options in this list will be based on the viewers’ preferences. This presents a challenge to the research community focused on decision-making and this is the problem that we want to consider in detail here. we propose here an intelligent system which is able to display a small list of recommended programs. Our proposal is focused on an Ambient Intelligence system which identifies the current viewers in front of the TV in a particular moment. a family with children. the service displays a list of the members living in the house. If Smart Homes are going to be massively adopted by society then they have to take into account that homes are inhabited by several individuals and each one will require a personalized service from the artificial system. magazines). viewers spend more time searching for a suitable option than watching a specific program. Although it represents a wide offer for viewers. For example. Moreover.g. being actively tracked by the system). the reality is that houses are inhabited often by more than one person and when caring for an elderly who is living alone at home is very important so it is that these systems are able to take care of an elderly couple. external reviews (e. live event. and they eventually get bored and stop looking for other proposals that could have fulfilled their preferences better. Consequently.. For example. As each individual has different preferences (some prefer TV news whilst others will mainly watch sports programs) and needs (some want the system to entertain them with music whilst others need a quite environment to study or to relax) then conflict seems unavoidable and the responsibility to harmonize will be on the artificial system to balance these preferences and needs. etc. it could take place when the user initiates the recommending service. program information such as the importance of the event (e. This identification may be performed in different manners. In that moment. or similarities with other recommended programs.g. In order to avoid viewers wasting time in unsuccessful searches. In this case.. Once the viewers are identified..g. Digital TV (DTV) has brought a revolution in the growth of the number of channels and programs broadcast. However. this system could be extended with ambient information to perform more intelligent decisions when recommending programs. etc. Another more advanced type of identification could be achieved by using RFID tags. the AmI system combines . if the system detects an under-18 viewer.

. Internet Web Services. recommended programs). a team of agents cooperating to solve a task (e. having an agent as responsible for finding reasons to recommend a program. An argumentation process takes place between two agents. 14. while another agent performs the opposite action.. The domain criteria is again used here to evaluate the strength of the arguments. a team of robots in a soccer team trying to score a goal) may have conflicting views on what to do and how to do it. To this end. Both these approaches fit well in a methodology explored by an area of Computer Science in the latest three decades: Argumentation Systems [13. 16. The goal of this dialog is to persuade each other about the validity of the exchanged arguments. Firstly.. Consequently. the AmI system must evaluate all the related justifications to take a final decision. the main goal of the AmI system is to display a list of recommended programs according to the evaluation process. a most preferable option within those being examined. the AmI system defines a hierarchical scale of crite- . the second approach enables the AmI system to build justifications in a distributed manner. by starting a persuasion dialog [19.A. while the domain criteria employed to evaluate justifications or arguments could be the same in our two approaches. J. the AmI system builds a set of justifications for and against recommending a program. Based on this combination of information. and J. They naturally model the existence of conflicting views and the need of preference criteria to select.e.g. 15. i. to find reasons against recommending that program.328 A. Mu˜ oz.C. On the other hand. we have devised two approaches to integrate the evaluation process into the AmI system. For example. This inconsistent pieces of knowledge within a system would have rendered useless a system based in classical logic but it is tolerated and managed through a meta-decision criteria within an argumentation system.g. Bot´a. when possible.. In this context. Taking this description as a starting point. 17]. a program could be recommended and discarded at the same time by means of different justifications. its implementation differs depending on the approach selected. Therefore. Both of them are introduced here in order to evaluate their advantages and drawbacks. This has provided a useful framework for the multi-agent systems and robotics communities [18]. 20] where they exchange their justifications (or arguments) about their different views (e. In the first approach. the AmI system acts as a central entity which builds favourable/unfavourable justifications for recommending a program and then it applies a set of domain criteria to rule out those justifications that do not fulfill some conditions.g. Augusto n ı the viewers’ preference and personal information with the TV programs data retrieved by e.

This design shows two components.1. . etc. an agent A could put forward an argument U1 defending a program as recommended according to a specific criteria C1 . An abstract architecture of the SmartTV system is depicted in Figure 12. This system will be referred as SmartTV system henceforth. However.3 shows the argument structure and how they are built in the AmI system. A could then put forward an argument U3 against U2 by using a new criteria C3 (and C3 is preferred to C2 ). where C2 is preferred to C1 .6. An argument is then accepted when it has not been attacked or its attackers have been defeated by other accepted arguments in the dialog.2 the abstract architecture of this AmI system is depicted. A comparison between these two approaches is offered in Section 12. during the persuasion dialog. In this level. those justifications that fulfill all levels in the scale are classified as acceptable. Finally. Section 12. The rest of this chapter is structured as follows. Hence.4 reviews the different domain criteria that are applied to the scenario of recommending TV programs.).1 is devoted to obtain an intelligent and contextaware system which recommends a set of TV programs according to different types of information (viewer’s presence and preference. and Section 12. the argumentation process follows until one of the agents runs out of new arguments for attacking the ones received previously. The next subsections describe in detail each module. the implementation of the criteria is integrated in the argumentation process itself. assigning one level to each specific criterion. Finally. Section 12.2 SmartTV System Architecture The AmI system introduced in Section 12.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 329 ria. the Information and Decision modules. Section 12. giving some examples in practice.5 explains the two approaches for resolving conflicts when several inconsistent justifications are found. Then. In Section 12. The disagreement between arguments is known as “attack”. Then. the justifications for recommending/discarding a program are evaluated according to the scale in a bottom-to-top manner. 12. In the second approach. which form the core of the SmartTV system.7 points out the conclusions and it gives some future research directions. program information. the system is conceived as a black box which collects viewers and programs data and returns a list of recommended programs. Hence. An agent B could show a disagreement with U1 by giving another argument U2 discarding the same program with a criteria C2 .

This module has three specific tasks: (1) To describe a knowledge model for representing the viewers and programs information. a semantically annotated model will enable the definition of domain criteria to be used when evaluating arguments for a program (see Section 12. as it will be used by different entities in the system (elements in the Information and Decision modules).330 A. Firstly. and J.C. In particular. a formal representation is needed for the entities in the SmartTV system in order to process and infer the knowledge given. J. (3) To deliver the knowledge model about current viewers and programs to the Decision Module.2).A. we adopt here the OWL-DL version1 of this language.2. A compelling knowledge model representation which fulfills the previous requirements is OWL (Web Ontology Language) [21].org/TR/owl-ref/ . Thus.1 An abstract representation of the SmartTV system architecture 12. Bot´a.1 Information Module The Information module acts as the input source of data in the system. the inference process will allow the system to find justifications for recommending/discarding a specific program. the standard Semantic Web [22] language. it should be expressive enough for modeling the targeted situations in the domain. Some characteristics must be taken into account when selecting the knowledge model representation. Secondly. Moreover. the model should be easily shareable. Augusto n ı Viewer1 SmartTV system Information Module Viewer2 Decision Module Viewer1 Information Domain Criteria List of Recommended Programs Viewer2 Information ViewerN ViewerN Information Decision Mechanism Program Information Internet Figure 12. This decision is based on 1 http://www. Mu˜ oz.2. (2) To collect the viewers and programs data through different kinds of interfaces and properly transform these data to the format established in the knowledge model. And last but not least.w3.

etc. The information related to TV programs could be obtained from the Internet..g. when the SmartTV system is requested to offer a list of recommended programs. Some examples of concepts and relationships in the SmartTV ontologies will be given throughout this chapter. Therefore. 23] for more details about OWL and Description Logic. the Personal and Program agents. These Web Services could even use OWL as the language to represent the information [26]. Therefore. it could easily be integrated into the knowledge model used in the system. etc. disjointness. age) and preference (preferred programs. (iv) OWL-DL ontologies can be explicitly augmented with semantic axioms (e. The collected data is stored in the SmartTV system according to the knowledge model representation selected (OWL-DL ontologies in our approach). Different kinds of relationships among these concepts are also possible when modeling ontologies (e. genres. union and intersection of concepts. Thus.). all the information about viewers and programs must be separately maintained and timely delivered when the Decision module is to process it. (iii) OWL-DL is based on Description Logic (DL) [23].g. DL empowers several types of automatic knowledge manipulation and reasoning processes (see Section 12. Finally.3). As a result. by using Web Services retrieving specific and detailed data [25] (program’s genre. (ii) OWL enables the conceptualization of both viewer’s personal and preference information and program data by means of ontologies.) data. The interested reader is referred to [21. GUIs and Internet files could still be part of the input sources. Graphic user interfaces (GUI) are the common means to obtain viewers’ personal (name. etc. the knowledge model in the SmartTV system is represented by means of OWLDL ontologies. other concepts or relationships may automatically be derived from them. a subset of First-Order Logic. cast and reviews. a program is similar to another program). similar programs. The collection of input data in the SmartTV system involves two types of actors: Viewers and programs.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 331 the next characteristics of OWL and the specific OWL-DL version: (i) OWL’s syntax is written in XML (Extensible Markup Language). subsumption. a viewer likes a program. which .. it first detects the current viewers to only obtain and use the information about those particular viewers. Then. the Information module may be composed of two software agents.). Other techniques such as preference learning [24] could also be studied. which enables an easy and standard manner of interchanging the knowledge model. actors.

Finally. Hence. an argument recommends a program because it is one viewer’s preference. for example.1.g. delivering it to the Decision module when that particular viewer is detected in front of the TV..g. the Program agent will maintain an up-to-date information about programs by subscribing to the correspondent Web Services. and J. To solve this situation. Let us see an example of a simple domain criteria. suppose that the concept FavouriteProg represents the type of all programs that are classified as preferred. (3) To solve conflicts between inconsistent arguments by employing a specific decision mechanism. Using this knowledge.A. y) between two types of information x and y (e. In this context. a decision mechanism is needed. Therefore. a relation moreRelevant(x. Thus.2 Decision Module The Decision module is provided with viewers and programs knowledge from the Information module. the Decision module has three specific tasks: (1) To allocate the domain criteria in the knowledge representation model. the module builds arguments for recommending or discarding a program. x and y could be concepts and relationships in the OWL ontology) expresses that any element belonging to the type of information x is more important than any element belonging to y when they are compared in different arguments. etc. On the other hand. sex. whereas an argument . the Personal agent will keep different viewer’s information. Moreover. 12. some conflicts may arise between arguments (e. Mu˜ oz. suppose an argument Ud discarding a program p because it is classified as ExplicitProg.C.g. information about parental guidance of a program is more relevant than information about preferences on that program) to evaluate and select which argument must prevail in case of conflict.332 A. J. since they contain violence. This mechanism uses some domain criteria (e. The addition of domain criteria to the knowledge model could be achieved by including semantic annotations in the model. (2) To build arguments for recommending and discarding a program according to the knowledge provided by the Information module. since some viewers like them. respectively.2. Example 12. while another argument discards the same program because it is classified as +18 and a teenager viewer has been detected). Bot´a.. Augusto n ı are responsible for the viewers and programs information.. Suppose an OWL ontology where the concept ExplicitProg represents the type of all programs that are classified as explicit content.

g. then new information about e can be inferred according to the THEN part (the conditions part of the rule are called antecedents. some inconsistency among these arguments could arise. As it has been commented along this chapter. Domain rules are of the form of IF {conditions on domain element e} T HEN {new information on e}.. The first proposal (see Figure 12. Two approaches are presented here. the central . Jena [27] or Pellet [28]) and these domain rules. a more practical solution will involve system administrators defining domain criteria according to users’ requirements. However. this process could be seen as the construction of arguments for supporting different conclusions. As a result.2(a)) shows the decision mechanism as a central entity. If the domain is well defined and easy to understand. but introducing the domain criteria moreRelevant(ExplicitProg.4. After receiving the viewers and programs knowledge from the Information module (steps 1 and 2).g. In turn. Section 12. when different arguments about the same domain element coexist.. Then. the users themselves may use GUIs for this task. Another question that must be resolved when managing domain criteria is related to who is the responsible for defining them. and therefore obtain new conclusions by means of a sequential deductive inference process. This inference process takes the viewers and programs knowledge from the Information module together with some domain rules introduced in the Decision module (see Section 12. argument Ud prevails over Ur since it contains more relevant information. the SmartTV system needs a decision mechanism in order to select which argument is more acceptable (if possible).3 shows the argument structure and inference process in detail. which in turn could be supported by other arguments.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 333 Ur recommends the same program p because it is classified as FavouriteProg. by using a reasoning engine (e. new conclusions are obtained from the viewer and program knowledge. while the new information part is called consequent). Ud and Ur are conflicting arguments. then users will select which kinds of criteria must be employed as default option or/and on-demand. meaning that whenever a set of conditions about a particular element e in the domain (e. The Decision module performs a deductive inference process for building arguments for and against recommending programs. Several domain criteria and their integration into the knowledge model will be developed in Section 12. the inferred conclusions could again be used as knowledge to fulfill the conditions of other rules. Once the criteria have been established. In these cases.3). a program) are true. FavouriteProg) in the OWL ontology as a semantic annotation between concepts.

Both decision mechanisms present advantages and drawbacks.C. The result of this argumentation indicates whether the initial claim is accepted by the opponent. whereas the other agent acts as the opponent. Mu˜ oz. Example 12. Then..1. the domain model in the SmartTV system is represented by means of OWL-DL ontologies. or otherwise the proponent must retract it. Let us see a simple example of each assertion. a is-a C). ontologies also contain assertions about these elements. Suppose an OWL-DL ontology containing the concepts Viewer and Program.2. Bot´a. b) satisfies that individual a is related to b by the relationship R.g.3 Arguments in the SmartTV System As previously stated in Section 12. 12. two types of assertions are found. Hence. One of the agents is the proponent of the dialog by claiming an initial proposal (e. Likewise.A. Augusto n ı entity builds all the arguments possibles (step 3). and the relationship likesProgram(Viewer. Then. J. Suppose that a and b are two domain elements (e. while DiscarderAg finds arguments for discarding them (step 3). Those arguments that fulfill all the criteria are evaluated as acceptable. suppose that C and R are a concept and a relationship in the ontology. respectively. while a role assertion R(a. The second proposal (see Figure 12...2.5.334 A. known as individuals. These ontologies establish the domain concepts and relationships. and J. where they exchange claims and their justifications to persuade each other about the validity of the claims (step 4). two specific TV programs). the role assertion likesProgram(Peter. where agent RecommenderAg uses the inference process to build arguments recommending programs. In OWL-DL ontologies. that a program p should be recommended). and they will be developed in Section 12. the concept assertion Viewer(Peter) indicates that Peter is an individual that belongs to the concept Viewer. Friday the 13th) states that Peter likes Friday the 13th. This relationship relates a viewer with a program which she likes. Program).2(b)) is centered on a distributed construction of arguments. A hierarchical scale of domain criteria is allocated in it (step 4). Then. and the arguments generated by this module are then filtered according to the scale (from Criterion 1 to Criterion n). a concept assertion C(a) states that individual a belongs to the concept C (i. an argumentation process is commenced as a dialog between the agents. representing the current state of affairs in the domain. trying to invalidate that claim with its own justifications.e. Furthermore. . which is a program according to the relationship likesProgram.g.

Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 335 (a) Centralized approach (b) Distributed approach Figure 12. or derived assertions from other arguments (e. the viewers’ preferences or the programs’ genre). and according to the deductive inference process. . The entailed assertion in the argument is known as conclusion.g. The assertions forming part of the premises could be categorical (or grounded) facts obtained from the input sources (e.g. Thus. and as an argumentation process between recommender and discarder agents (b) In this context. an argument is defined as the justification for an assertion (e. This support set is formed by premises. the rule’s consequent is entailed as the conclusion. RecommendedProg(p). Then.. where p is a TV program) through a deductive inference process [29]. .g. which are a group of assertions in the model and exactly one domain rule. while the justification is called support set. the assertion p is an explicit program could be derived by an argument). the assertions in the support set match the rule’s antecedents... . . an argument U supporting a conclusion could in turn be supported by other arguments V1 .Vn if some of the U’s .2 The decision mechanism as a central unity which builds arguments and then filters them using a hierarchical scale (a).

math. where • The assertion ρi represents a grounded fact stated in the model by an input source or the conclusion of another argument. J. V = (ρi .A. we define U as an argument supporting an assertion ϕ by means of a support set ΣU . Moreover. This distinction between the type of the rule distribution will be explained at the end of this section. . According to Section 12. ΣV ). ρn be assertions about the domain model. To this end. . These built-ins are functions that offer different operations on variables (comparison.2. 1 i n.2. . . Let ρ1 . Then. . namely Horn clause rules. . Mu˜ oz. ΣU ). .C. Bot´a.1 (Argument). • DomainRule is an IF-THEN rule whose antecedents are matched by assertions {ρ1 . . Domain rules are expressed by means of SWRL [30] (Semantic Web Rule Language). . .1. Definition 12. ρn } and whose consequent is instantiated to ϕ • ΣU ϕ . the grounded facts involved in arguments are delivered by the Information module. Augusto n ı premises are conclusions supported by V1 . represented as U = (ϕ . ρi ∈ O. . ρn . This rule language adds a new kind of axiom to OWL-DL. . which extends OWL’s syntax and semantics in order to express IF-THEN rules such as the one in Section 12.. SWRL rules are of the form of an implication between a conjunction of antecedents (body) and a consequent (head). . y)) assertions from the OWL-DL language—where x. Both parts of the rule consist of atoms: unary (C(x)) or binary (R(x. where is the modus ponens operator (performing the deductive inference process). . . respectively. . Regarding domain rules. ). . Let us see some domain rules for the SmartTV system. . y are variables or OWL-DL individuals—and Built-ins. DomainRule}. where: • ϕ is the conclusion in U • ΣU is a premise list {ρ1 . Let O be an OWL ontology representing a domain model. Let us put this argument definition by means of a symbolic definition. they could be part of the knowledge model (designed by the domain expert) when the decision mechanism is a central entity. or they could be partially held by the argumentative agents when the argumentation process is used to solve conflicts.2.Vn . and J.336 A. . an OWL-DL ontology must be previously defined to conceptualize the concepts RecommendedProg and DiscardedProg as recommended and discarder TV programs.

respectively: RuleLikeP : Viewer(?x) ∧ Program(?p) ∧ likesProg(?x. ?p) ⇒ FavouriteProg(?p) RuleLikeG : Viewer(?x) ∧ Genre(?g) ∧ Program(?p) ∧ ∧ likesGenre(?x. it can be used as a knowledge model for creating rules. Program) are available in the ontology.. condn ) T HEN conclusion rules are represented by the following syntax: cond1 ∧ cond2 ∧ . Once the SmartTV ontology has been defined. THEN is substituted by ⇒.. Likewise. but a small model of the logical rules employed to classify a TV program as recommended or discarded2: (A) These three rules classify a program as recommended whenever it is classified as favourite. This a convenient manner of grouping rules. Rules in group A are designed to classify a program as recommended. The design of the domain rules has been divided into two groups. Program). cond2 . Other concepts. ?g) ⇒ FavouriteProg(?p) RuleLikeA : Viewer(?x) ∧ Actor(?a) ∧ Program(?p) ∧ ∧ likesActor(?x..2) These two rules classify a program as interesting according to some magazine reviews about that program or due to its similarity with some other favourite prothis chapter. dislikesGenre(Viewer.. and ‘∧’ is a conjunction operator read as “and”. . ?g) ∧ hasGenre(?p. among others.e. therefore it is not necessary to explain them here. IF (cond1 . Variables in rules are represented by preceding their name with the symbol ‘?’. 2 In . The rest of concepts and relationships are referred in the domain rules.1) This set of rules classify a program as favourite according to the viewer’s preferences on the program. ∧ condn ⇒ conclusion. the following rules are necessary for classifying a program as favourite. i.. ?a) ∧ cast(?p. Program) and hasSimilar(Program. in order to clearly differentiate the proposal of each rule and easily add future rules to their corresponding group. relationships such as likesProgram(Viewer. where IF is omitted. whereas OWL-DL individuals start with a capital letter. whereas rules in group B are directed to discard a program. such as Viewer or Review have been also defined in the ontology. interesting or important: RuleFav : FavouriteProg(?p) ⇒ RecommendedProg(?p) RuleInt : InterestingProg(?p) ⇒ RecommendedProg(?p) RuleImp : ImportantProg(?p) ⇒ RecommendedProg(?p) Now.. they are contradictory concepts. ?a) ⇒ FavouriteProg(?p) (A. its genre or the program’s cast.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 337 DiscardedProg ≡ ¬RecommendedProg according to this ontology. Note that the following set of rules is not exhaustive. interesting or important: (A.

Bot´a. ?t) ⇒ ImportantProg(?p) (B) These three rules classify a program as discarded whenever it is classified as nonfavourite. ?v) ∧ greaterT han(?v. This built-in is matched if the rating value in ?v is greater than 5.3) This rule classifies a program as important when it has some relevant tags associated. as for example an “urgent” tag: RuleUrg : Tag(?t) ∧ Program(?p) ∧ tagValue(?t. Notice that when ?v is lower than 5.Urgent) ∧ ∧ hasTag(?p. the following rules are necessary for classifying a program as non-favourite. (A. non-interesting or containing explicit content: RuleNonFav : ¬FavouriteProg(?p) ⇒ DiscardedProg(?p) RuleNonInt : ¬InterestingProg(?p) ⇒ DiscardedProg(?p) RuleExp : ExplicitProg(?p) ⇒ DiscardedProg(?p) Now.338 A. 5) ⇒ InterestingProg(?p) RuleSimilar : FavouriteProg(?s) ∧ Program(?p) ∧ ∧ hasSimilar(?p. non-interesting or explicit: (B. Augusto n ı grams: RuleRev : Review(?r) ∧ Program(?p) ∧ hasReview(?p. J. ?g) ⇒ ¬FavouriteProg(?p) RuleDislikeA : Viewer(?x) ∧ Actor(?a) ∧ Program(?p) ∧ ∧ dislikesActor(?x.A. ?g) ∧ hasGenre(?p.1) This set of rules classify a program as non-favourite according to the viewer’s dislikes on the program. its genre or the program’s cast. the built-in is not matched and the rule will not be fired. ?p) ⇒ ¬FavouriteProg(?p) RuleDislikeG : Viewer(?x) ∧ Genre(?g) ∧ Program(?p) ∧ ∧ dislikesGenre(?x. ?r) ∧ ∧ rating(?r.2) These two rules classify a program as non-interesting according to some magazine reviews about that program or due to its similarity with some other non-favourite programs: . respectively: RuleDislikeP : Viewer(?x) ∧ Program(?p) ∧ dislikesProg(?x. ?a) ∧ cast(?p. ?a) ⇒ ¬FavouriteProg(?p) (B. ?s) ⇒ InterestingProg(?p) Observe the use of the built-in greaterT han in the rule RuleRev. Mu˜ oz. and J.C.

?r) ∧ ∧ rating(?r. whereas derived assertions have a dashed line. Notice that when ?v is greater than 5. whose preferences are represented by the agent PeterAg. likes(Peter. and a ProgramAg in the Information module stores information about that program. Then.3.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 339 RuleNegRev : Review(?r) ∧ Program(?p) ∧ hasReview(?p. {FavouriteProg(Lost). all this knowledge is passed to the Decision Module. This attack is known as rebut. This built-in is matched if the rating value in ?v is lower or equal than 5. ?s) ⇒ ¬InterestingProg(?p) Observe the use of the built-in lessEqT han in the rule RuleNegRev. Domain rules are represented by means of arrows next to the rule’s name. Lost). ?g) ∧ lessEqT han(?a. Moreover. the conclusion in argument U attacks the conclusion in V . RuleFav}). when an argument U supports a conclusion which is conflict with another argument V . Program(Lost). the built-in is not matched and the rule will not be fired. 5) ⇒ ¬InterestingProg(?p) RuleNegSimilar : ¬FavouriteProg(?s) ∧ Program(?p) ∧ ∧ hasSimilar(?p. A graphical representation of both arguments is shown in Figure 12. As mentioned in Example 12. In this case. Peter likes the TV series Lost. Example 12.1. RuleLikesP}) Note that argument U1 concludes Lost as a recommended program. while premises are drawn in an oval shape. the conclusion in U may . This notion of conflict is also known as attack. Each argument is contained in a box with a solid black line. {Viewer(Peter). U2 = (FavouriteProg(Lost). ?a) ∧ ∧ PG(?p. some examples of arguments can be given. This module builds the following arguments: U1 = (RecommendedProg(Lost).3) This rule classifies a program as explicit when it has a parental guidance (PG) and there is a viewer younger than the recommended age in the PG: RuleAge : Viewer(?x) ∧ Program(?p) ∧ hasAge(?x. This classification is justified by the Peter’s liking to Lost and the use of the rule RuleLikeP in argument U2 . After defining the argument structure and how they are built. it is said that U attacks V . based on the knowledge that the program has been classified as favourite.3. and vice versa. Suppose that Peter is a viewer using the SmartTV system. different arguments can support inconsistent conclusions. ?g) ⇒ ExplicitProg(?p) After defining the domain rules. ?v) ∧ lessEqT han(?v. (B. Premises being grounded facts have a solid line. we can now explore the type of conflicts between arguments [31]. Conclusions are represented with a rectangular box. Thus.

⊥ . Let us see a formal definition for both attacks. V = (γ . ΣU ). Let O be an OWL ontology representing a domain model.2 (Rebut). J. ΣV ) be two arguments in O. Next. Definition 12. U undercuts V iff ∃ρi ∈ ΣV such as • ϕ ∪O • ρi ∪ O ⊥. and ⊥ • ϕ ∪γ ∪O Note that a rebut between U and V is detected in the SmartTV system when the addition of the conclusions of both arguments entails an inconsistence in the knowledge model. Augusto n ı Figure 12. ⊥ . and ⊥ • ϕ ∪ ρi ∪ O The concept of undercut may also be applied to the domain rules employed in the support set.340 A.3 (Undercut). Let U = (ϕ . ΣV ) be two arguments in O. although this attack is beyond the scope of this chapter and it will not be explored here. U rebuts V iff • ϕ ∪O • γ ∪O ⊥.3 Argument U1 supports Lost as recommended program. Definition 12. the definition of the other type of attack between arguments is given. This attack is known as undercut. V = (γ .C. while U1 ’s premise is supported by U2 attack the premises in the support set of the argument V .A. and J. and besides this inconsistence is not generated when adding each conclusion separately. . Then. Mu˜ oz. Then. Let U = (ϕ . Let O be an OWL ontology representing a domain model. ΣU ). Bot´a.

The status of an argument (i. Let Ω be the set of all arguments built in O. domain rules will be distributed as well. Observe that Definition 12. and vice versa. They will be explored in Section 12. In order to establish such a status. Definition 12. and besides V is not defeated. U defeats V .2. a decision mechanism is employed. the domain criteria could be defined on the rules when a central entity keeps . Let U = (ϕ . Finally. DiscarderAg is not aware of group A rules. And secondly. In the SmartTV system.5. However.4 (Defeat). U rebuts V and the premises in U are more relevant than the premises in V according to a particular domain criteria. The domain criteria used to determine the preferable relationship between arguments will be introduced in the next section. Contrarily. as the second condition of the definition states. when the argument U is evaluated as more preferable than argument V .2(a)) is in charge of building the arguments. If a central entity (Figure 12. Thus.2(b). we consider that U defeats V if the two following conditions hold: Firstly. defeated or undecided) must be defined based on the interactions among an argument and the rest of arguments in the domain. Observe that this distribution means that RecommenderAg is not aware of the rules in group B. it is not sufficient to compare pairs of arguments.. Let O be an OWL ontology representing a domain model.e. Therefore.and • ∃V ∈ Ω such as V undercuts U. A formal definition of the concept of defeat is given as follows. RecommenderAg will control all rules in group A. now it is necessary to evaluate them in order to decide which argument is more preferable. Thus. The type of the rule distribution has several inherit implications. represented as U V . it is said that U defeats V . ΣV ) be two arguments in Ω. we propose here two different types of implementations of this decision mechanism. As previously commented in Section 12. ΣU ). when the construction of arguments is distributed as in the approach of Figure 12. whereas rules in group B are separately maintained by DiscarderAg. an argument can be attacked by other arguments. it will contain all rules in the domain (A and B rule groups in the SmartTV system). U is not undercut by any other argument. iff • U rebuts V and U is preferable to V according to a domain criteria in O. in the SmartTV system. acceptable. For example.4 only states a defeat relationship between two arguments.2. One of these implications is related to the placement of the domain criteria used to evaluate conflicting arguments. V = (γ .Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 341 Once attacks between arguments are detected. it is important to notice how domain rules are distributed depending on the decision mechanism selected. Then.

Bot´a. the UE premise set contains an assertion belonging to ExplicitProg (ExplicitProg(Lost)). the following arguments about the program Lost have been built: UF = (RecommendedProg(Lost). RuleExp}) According to the SmartTV ontology. suppose that the relationships (ExplicitProg > FavouriteProg) and (ExplicitProg > InterestingProg) are defined on the SmartTV ontology. That is to say. RuleInt}). meaning that the type of information about the explicit content of a program is more relevant that its preference or interest. Moreover. The feasibility of this solution is based on two features of the knowledge model representation. which is more relevant than the type of information of the premises FavouriteProg(Lost) and InterestingProg(Lost) in arguments UF and UI . For example. Augusto n ı the whole set of domain rules. respectively. FavouriteProg and InterestingProg represent the set of programs that have been inferred as explicit content. the argumentative agents are totally aware of the 3 The operator > represents the relationship “moreRelevant” . DiscarderAg cannot set its own rule RuleExp as more relevant than RuleFav or RuleInt (both rules belong to group A). Firstly. J. UE UF and UE UI . UI = (RecommendedProg(Lost). Therefore. Now. respectively. {FavouriteProg(Lost).C.4. {InterestingProg(Lost). meaning that an argument discarding a program p because it based on the rule RuleExp prevails over arguments recommending p when they are based on the rules RuleFav or RuleInt (see Example 12. this representation is shared by all the entities involved in the AmI system. As a result. UE = (DiscardedProg(Lost). to the “type of information” of the assertions in the premises. it is impossible for it to set a prevailing relation between its own rules and those ones. {ExplicitProg(Lost). the concepts ExplicitProg. Mu˜ oz. This means that the concept or relationship in the knowledge model to which the assertions belong will be used to evaluate the prevalence of any argument. this type of rule relationship cannot be held when rules are distributed. favourite and interesting. However. Let us see an example. If an agent does not know the rules maintained by other agents.A. As shown in the groups A and B of domain rules. RuleFav}). the relationships3 (RuleExp > RuleFav) and (RuleExp > RuleInt) could be stated in the central entity.2. However. as commented in Section 12. For instance.342 A. and J. argument UE is inconsistent with arguments UF and UI since their conclusions are in conflict. or more specifically. The solution proposed here is to shift the placement of the domain criteria from the rules to the assertions in the premises.1. since the agent does not even know the existence of such rules.4). the domain criteria takes the information in the assertions used as premises to evaluate conflicting arguments. Example 12.

the modeling language OWL-DL allows for semantic annotations on the different concepts and relationships that constitute the domain.4 Domain Criteria in the SmartTV System (a) Several arguments (based on different type of information) supporting the same conclusion (b) Two argument supporting conflicting conclusions. Secondly. For example. Examples of these criteria and how they are introduced in the model will be given in Section 12.4(a)).Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 343 knowledge model used in the system. . based on the same type of information Figure 12. Due to these annotations. based on different type of information (c) Two argument supporting conflicting conclusions. the conclusion RecommendedProg(BBCNews) could be inferred due to BBCNews is classified as an important. 12. and the domain criteria is included in it. three different situations could happen when building arguments: (1) The same conclusion may be justified by different types of information.4 Three different situations when building arguments in the SmartTV system According to the domain rules listed in Section 12.3. the domain criteria are easily added to the model.4. favourite and interesting program (see Figure 12.

• Based on the viewers’ preferences on p . two criteria are defined which include this new association: A criterion C+ for the arguments’ premises supporting the conclusion ϕ is proposed..e. J. This best option is understood as the argument whose assertions in the support set belong to the most relevant type of information according to the knowledge model. the conclusion InterestingProg(ScaryMovie4) could be inferred due to its similarity with another preferred program (e.g. ¬InterestingProg(ScaryMovie4) could be inferred because there is a review about the program assigning it a low rating (see Figure 12. and another criterion C− is given for arguments’ premises supporting the conclusion ¬ϕ .2. .4. For example.. Contrarily.. When dealing with these situations.C. Note that the situations 2 and 3 are equivalent to the concept of rebutting given in Definition 12. Case 1: Recommended Criteria Three types of arguments can be built for the conclusion RecommendedProg(p): • Based on the importance of p.344 A. and it is included in the SmartTV ontology as an OWL-DL relationship. For example.1 Same Conclusion Supported by Different Type of Information This subsection deals with criteria for deciding the “best option” between arguments supporting the same conclusion.. 12.4(c)). ScaryMovie). The next subsections give some ideas about how to define a criteria in each situation.4(b)). Bot´a. Let us see several examples of these criteria in the SmartTV system.A. we include here a new association between the elements of the knowledge model (i. This association will be referred as preferred in the examples of this section. Mu˜ oz. (3) Inconsistent conclusions may be justified by the same type of information.g.g. e. concepts and their relationships in the SmartTV ontology) to express when an element is more relevant than others. the conclusion FavouriteProg(CSI) could be inferred due to one viewer’s liking. because someone likes the genre of the program (FavouriteProg(p)). e. and J. Contrarily. Thus. Augusto n ı (2) Inconsistent conclusions may be justified by different types of information. As a result. the SmartTV system needs some criteria in order to decide which argument is more acceptable in each case. because it is tagged as urgent (ImportantProg(p)). ¬FavouriteProg(CSI) could be inferred because there is another viewer who dislikes the program (see Figure 12.

(ii) An argument discarding p because ¬FavouriteProg(p) contains a more relevant type of information than those based on ¬InterestingProg(p)..g. due to a bad review in a magazine (¬InterestingProg(p))..g.5): (i) An argument recommending p because ImportantProg(p) contains a more relevant type of information than those based on FavouriteProg(p) or InterestingProg(p). An important program could represent breaking news or a live event. the following criteria RecCriteria+ may be established (see Figure 12. Likewise.5 Diagram representation of RecCriteria+ and RecCriteria− . The viewer’s personal preferences on a program are superior to reviews or similar features to other programs. e. Then. the following criteria RecCriteria− may be established (see Figure 12. because someone dislikes the actors playing in it (¬FavouriteProg(p)). therefore it is a powerful reason to discard it more than not being favourite or interesting.g.g. e. (ii) An argument recommending p because FavouriteProg(p) contains a more relevant type of information than those based on InterestingProg(p). An explicit program could contain violence or sex. e. • Based on the viewers’ preferences on p .. Figure 12.. • Based on some interesting features about p. due to a similarity found with another favorite program p (InterestingProg(p)). . e. because it is not suitable for some viewers according to their ages (ExplicitProg(p)).5): (i) An argument discarding p because ExplicitProg(p) contains a more relevant type of information than those based on ¬FavouriteProg(p) or ¬InterestingProg(p). Then.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 345 • Based on some interesting features about p. three types of arguments can be built for the conclusion DiscardedProg(p): • Based on the PG of p.

Mu˜ oz. Figure 12. p)) contains a more relevant type of information than those based on the program’s genre or cast.C. Bot´a. the following criteria IntCriteria+ /IntCriteria− may be established (see Figure 12. . Case 3: Interesting Criteria Two types of arguments can be built for InterestingProg(p) / ¬InterestingProg(p): • Based on the review of a program p (hasReview(p. s)). p)). J. Then. p)/dislikeProg(x.346 A. • Based on the similarity between p and other programs (hasSimilar(p. Augusto n ı Case 2: Favourite Criteria Three types of arguments can be built for FavouriteProg(p) / ¬FavouriteProg(p): • Based on the direct viewer’s like/dislike of program p (likesProg(x. g)) contains a more relevant type of information than those based on the program’s cast. • Based on the viewer’s like/dislike of p’s cast (likesActor(x. a)/dislikesActor(x. and J. g)/dislikesGenre(x.6): (i) An argument classifying p as FavouriteProg(p)/¬FavouriteProg(p) by means of direct information (likesProg(x. • Based on the viewer’s like/dislike of p’s genre (likesGenre(x. r)). s)) contains a more relevant type of information than those based on the program’s review.7): An argument classifying p as InterestingProg(p)/¬InterestingProg(p) by means of similar information (hasSimilar(p.6 Diagram representation of FavCriteria+ and FavCriteria− . p)/dislikesProg(x. (ii) An argument classifying p as FavouriteProg(p)/¬FavouriteProg(p) by means of p’s genre information (likesGenre(x.A. Then. g)/dislikesGenre(x. the following criteria FavCriteria+ /FavCriteria− may be established (see Figure 12. g)). a)).

However.4.5.1.5. in RecCriteria+ and RecCriteria− . {¬InterestingProg(Lost). UD = (DiscardedProg(Lost). 12.6 and 12. These assertions belonging to the same group have an equal strength. RuleNonInt}) Since UF is based on the FavouriteIn f oGroup type of information.7). it is possible to define an evaluation mechanism for this type of inconsistency. and UD is based on the InterestingIn f oGroup one. suppose the two following conflicting arguments: UF = (RecommendedProg(Lost). For example.4. UF group. As a result. Then. the arguments are based on different types of information. Let us see an example of a comparison between two groups.5 does not mean that UF will finally be acUD according to the type of relevant information of each Figure 12. RuleFav}).2 Inconsistent Conclusions Supported by Different Type of Information This subsection deals with criteria for evaluating arguments supporting inconsistent conclusions.7 Diagram representation of IntCriteria+ and IntCriteria− . careless of the type of information being positive or negative. {FavouriteProg(Lost). the assertions based on the type of information FavouriteProg are more relevant than those based on InterestingProg. Suppose that FavouriteIn f oGroup is the group of type of information about the favorite status of a program (positive or negative). . Note that the decision reached in Example 12.4. we can now define prevailing relationships between groups of types of information. In this case. it is possible to find groups of types of information composed of positive and negative assertions about the same information.1. the relationship between these two groups could be (FavouriteIn f oGroup > InterestingIn f oGroup). and InterestingIn f oGroup is the type of information about its degree of interest (again.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 347 12. Looking deeply in the examples of criteria C+ and C− given in the three cases of Section 12. By combining the criteria C+ and C− defined in Section 12. there are similarities between the associations preferred defined over the types of information in each case (see Figures 12. Example 12. positive or negative). Now.

whereas the association preferred defined in Section 12. new relations of preference between types of information must be defined in C+/− . {¬FavouriteProg(Lost).A. careless of this information being positive or negative. Case 1: RecCriteria+/− Figure 12. J. suppose the following arguments are built in the SmartTV system: UImp = (RecommendedProg(BBCNews). For instance.348 A. Mu˜ oz. As a result. Note that there is a question that must be resolved in this criteria: What happens among ImportantProg and ExplicitProg?. Here it is important to remember that UImp can still be attacked by other arguments whose conclusions oppose to the premise ImportantProg(BBCNews) in UImp . since new attacks on the argument employed to infer FavouriteProg(Lost) are still possible. RecCriteria+ and RecCriteria− can be merged by grouping the same type of information as depicted in Figure 12. In these situations. RuleImp}).8. Thus.C. U¬F = (DiscardedProg(BBCNews).4. we can agree that an argument recommending p because it is classified as . They cannot be grouped since they represent different types of information. and J. {ImportantProg(BBCNews). RuleNonInt}) Then. Obviously. RuleNonFav}). an argument based on ImportantProg(p) is preferred to those based on FavouriteIn f oGroup or InterestingIn f oGroup type. The groups of type of information are modeled as OWL-DL concepts in the SmartTV ontology. UImp U¬F and UImp U¬Int . U¬Int = (DiscardedProg(BBCNews). a new relation of prevalence between them is needed. there will be cases where groups of type of information cannot be found.1 is used again to represent the prevailing relationships between these groups. {¬InterestingProg(Lost). The following examples show how a criterion C+/− could be defined by creating a new XXXIn f oGroup concept and merging the criteria C+ and C− previously given.8 RecCriteria+/− as a result of combining RecCriteria+ and RecCriteria− . For example. Bot´a. Augusto n ı cepted in the AmI system.

a new preference relation from ImportantProg to ExplicitProg is depicted in Figure 12. Program(XFiles). hasGenre(XFiles.. likesProg(Peter. deaths. RuleDislikeG}). D. Program(XFiles).8. As an example of this criteria. RuleDislikeA}) Then. etc. Program(XFiles).Duchovny). Sci − Fi).g. Now. which have been modeled as OWL concepts. {Viewer(Peter). Thus. FavCriteria+ and FavCriteria− can be merged by grouping the same type of information as depicted in Figure 12.9. RuleLikeP}). {Viewer(Lois). suppose the following arguments in the SmartTV system: UP = (FavouriteProg(XFiles). U¬A = (¬FavouriteProg(XFiles). Case 2: FavCriteria+/− Figure 12.Duchovny).9 FavCriteria+/− as a result of combining FavCriteria+ and FavCriteria− . dislikesGenre(Lois. where USim U¬Rev : . the prevailing relationships among these arguments is inferred as UP UP U¬A . U¬G = (¬FavouriteProg(XFiles). Actor(D. XFiles). U¬G and Case 3: IntCriteria+/− IntCriteria+ and IntCriteria− can be merged by grouping the same type of information as depicted in Figure 12. Sci − Fi). Genre(Sci − Fi). D. {Viewer(Chris). Observe the three groups of type of info in the figure.Duchovny). suppose the following arguments in the SmartTV system. p could be breaking news or a live event containing high relevant information which must be broadcast although it contains violence.).Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 349 important is preferred to an argument discarding p because it includes explicit content (e. cast(XFiles. dislikesActor(Chris.10.

hasReview(ScaryMovie4. Mu˜ oz. the argument URA1 concludes Lost as a recommended program based on the FavouriteProg type of information. hasSimilar(ScaryMovie4.C. and J. The question here is how to decide which argument is preferable when their premises have the same type of information value. U¬Rev = (¬InterestingProg(ScaryMovie4). MagaZine 001). As we have seen in Section 12. Program(ScaryMovie4). Bot´a. rating(MagaZine 001. USim = (¬InterestingProg(ScaryMovie4). Figure 12. In this case.11.11 Two conflicting arguments based on the same type of information.10 IntCriteria+/− as a result of combining IntCriteria+ and IntCriteria− .4. J. RuleSimilar}) 12. an attack to the support set (the undecutting concept given in Definition 12. the arguments are based on the same type of information. {Review(MagaZine 001). whereas UDA1 discards the same program based on ¬FavouriteProg.4. Firstly. Two situations can be developed in this scenario.4. In this figure. {FavouriteProg(ScaryMovie). then it is possible to compare these subarguments using the criteria described in Section 12.A. Augusto n ı Figure 12.2. both types of information are equally strong according to RecCriteria+/− (see Figure 12. An example of these type of inconsistency is given in Figure 12. Program(ScaryMovie4).350 A.2.3) may exist with respect to the two conflicting arguments if the subarguments can be evaluated according to that crite- . ScaryMovie). if the premises with the same type of information are derived assertions from other arguments (let us call them subarguments). RuleNegRev}). In this case.8). 3).3 Inconsistent Conclusions Supported by the Same Type of Information This subsection deals with criteria for evaluating arguments supporting inconsistent conclusions.

the previous solution is not possible. In this case. and eventually URA1 defeats UDA1 according to Definition 12. both subarguments URA2 and UDA2 are equally strong too. which is based on the direct viewer’s preference on a program. could be the key to set a prevalence between the two arguments. URA2 undercuts argument UDA1 . In this example. Then.4. Secondly. they do not come from a subargument). as shown in Figure 12. the premise FavouriteProg(Lost) in URA1 is supported by the argument URA2 . URA2 is preferred to UDA2 according to the criteria FavCriteria+/−. ¬FavouriteProg(Lost) in UDA1 is supported by the argument UDA2 . Then. according to criteria FavCriteria+/−. Thus. could be use to evaluate the conflicting arguments.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 351 ria. This “implicit” sense means that some other information about the assertions in the premises. When comparing these two subarguments. when all premises are grounded facts (i.13. As a result.12 Subarguments for the derived assertions FavouriteProg(Lost) and ¬FavouritePorg(Lost) based on different type of information. For instance.e. Peter) could be included in the knowledge model. since these subarguments may be supported by other subarguments. Then. Figure 12. the criteria that an argument with a viewer who .12. Observe that comparing subarguments could be a recursive process. which is based on the viewer’s genre preference. and both are composed of grounded facts. respectively. as its support has been defeated. UDA1 is not a valid argument now. An example of this situation is given in Figure 12. which is not explicitly present therein.. a relationship superiorViewer(Lois. implicit information about the premises Viewer(Peter) and Viewer(Lois) in arguments URA2 and UDA2 . A possibility to solve this unresolved situation could be to use implicit semantic information in the premises. Contrarily. the inconsistent situation between arguments could be left as unresolved.

and J.e.352 A. Augusto n ı is superior is preferred may be established in the knowledge model (i. As a result. and the subarguments related to these.. The next two subsections explain these alternatives. This new information is not explicitly included in the premises of the arguments.A. Two main approaches have been devised to implement the decision mechanism in the SmartTV system.4 have shown that different arguments may support inconsistent information about the same domain element (e.5. a central entity builds all the possible arguments for and against recommending a program. To this end. premises in these arguments could be supported by other arguments. J.. Mu˜ oz.3 and 12. This evaluation process is implemented here as a decision mechanism.5 Decision Mechanism: Two Approaches Sections 12. UDA2 URA2 ). To illustrate both alternatives. Figure 12. Hence.C.g. This solution is beyond the scope of this chapter and it is being studied as a future line of work. the evaluation process of inconsistent arguments must also consider all the subarguments related to them. a the argumentation process may be extended to take into account implicit information. and so on. Moreover. which in turn may be involved in sequent conflicts. 12. Bot´a. a TV program). some SmartTV scenarios are then evaluated. However.1 A Centralized Approach The first idea is based on a centralized design. 12.13 Subarguments for the derived assertions FavouriteProg(Lost) and ¬FavouritePorg(Lost) based on the same type of information. the entity receives all the .

3. Contrarily. For example. along with the whole set of domain rules integrated in it. thus sharing the same information. The argumentation process starts with an agent claiming a specific proposal. argument UF will pass to the second level. The specific domain criteria scale in the SmartTV system is composed of three levels. The bottom (or first) level contains the criteria IntCriteria+/−. a TV program should be recommended). As this decision does not affect the final result of the argumentation process. the agents begin an argumentation process. The arguments that fulfill all the levels in the hierarchical scale are then considered as acceptable. In this case. we have chosen RecommenderAg claiming a specific program p as recommended (i. Once they have built their respective arguments.. let us take the arguments UF and UD in Example 12. The domain criteria for comparing arguments (including their subarguments) are captured in a hierarchical scale. whereas rule set B is maintained by DiscarderAg. whereas another agent finds justifications which are inconsistent to that claim. 12. a DiscarderAg agent builds arguments for discarding programs.e. Moreover.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 353 domain knowledge from the Information module. RecommenderAg owns the rule set A given in Section 12. it seems to be more natural to begin with a claim recommending 4 Note . conflictive arguments which are defeated by a prevailing argument according to the specific criteria are not considered in the next level. and finally the top level is controlled by RecCriteria+/−. Then. After the initial claim. an agent builds arguments trying to support a particular claim (e. Here. Hence. when their premises are equally strong according to the criteria. Centering on the SmartTV system. the intersection set of rules of both agents is disjoint (i.5. α ≡ “p is recommended”)4. each agent has now an independent set of domain rules.. they do not share rules).2 A Distributed Approach The second idea is based on a distributed construction of arguments.e. DiscardedAg asks the that a start with DiscarderAg claiming p as not recommended is also possible.5. Examples of the employment of this scale are given in Section 12.5. denoted by α . the decision to take could be left to the viewers. while UD is eliminated from the evaluation process since it is defeated by UF . Each level in the scale contains a criterion C+/− .g. The middle (second) level incorporates the criteria FavCriteria+/−. suppose that the first level in the hierarchical scale is RecCriteria+/−. Note that inconsistent arguments could reach the status of acceptable after covering all levels in the scale.3. In that cases. Now. a RecommenderAg agent is in charge of building arguments for recommending programs. In each level. However.. The domain knowledge is delivered to both agents.

. Augusto n ı justification for α . DiscarderAg utters the argument VDA1 to attack URA1 . and J. (b) If all the premises in URA1 are categorical facts. To compare its own arguments against URA1 .UDA1 . if URA1 has been attacked by a counterargument (S2. Now. . Let URA2 denote this argument. then it is RecommenderAg who compares UDA1 or VDA1 with its own arguments as a program and then exchange arguments supporting/defeating that claim. which is then communicated to DiscarderAg. denoted by the finite set Ωeq = {UDA1 . Therefore. Suppose that RecommenderAg has the finite set of arguments Ωic = {URA1 .1. . DiscarderAg must accept the argument. Bot´a.}. . Let UDA1 be the best argument in Ωeq according to the correspondent C− criteria defined in Section 12. (S3) DiscarderAg has some arguments stronger than URA1 . Let VDA1 be the best argument in Ωst according to the correspondent C− criteria defined in Section 12. . DiscarderAg evaluates URA1 to find a better or at least equally-strong argument against it.a and S2.C.VDA1 . J. Two sub-cases are detected: (a) If any premise in URA1 is a derived assertion. Let us denote this argument by URA1 .}. Then.UDA1 . Then.2.4. (b) If all the premises in URA1 are categorical facts. Next. DiscarderAg utters the argument UDA1 to attack URA1 . . Mu˜ oz.4.1.A. DiscarderAg employs the criteria defined in Section 12. three situations can occur when DiscarderAg receives URA1 : (S1) DiscarderAg has no stronger arguments than URA1 . Two sub-cases are detected: (a) If any premise in URA1 is a derived assertion. DiscarderAg must then ask for the justifications of it. in order to try to undercut URA1 . .URA1 . . RecommenderAg selects the best argument in Ωic according to the correspondent C+ criteria (see Section 12. if URA1 has been challenged by asking the justification of any derived assertion in its premises (S1. DiscarderAg compares URA2 with its own arguments as shown above.a cases). RecommenderAg must give the best argument for it.} supporting α .354 A. denoted by the finite set Ωst = {VDA1 .4. Then. (S2) DiscarderAg has some arguments just as stronger as URA1 .b and S3 cases).1).VDA1 . in order to try to undercut URA1 . . . DiscarderAg must then ask for the justifications of it. On the other hand.4.URA1 .

which is not asserted in the shared knowledge model). RecommendedAg has to utter a new argument VRA1 for supporting the proposal whose previous supporting argument URA1 led to the blocked path. and the final decision about the program is again left to the viewers. the argument WRA could not be accepted by DiscardedAg as stated in S1. a dialog is developed where arguments of both agents are exchanged. the categorical/grounded assertions found in the premises (viewers’ preferences. the situation would be different. the argumentation process is left unresolved. As a result. If a new path can not be started. Observe that during the argumentation process. Then.3. However.5. For a more detailed explanation on the argument process. V could be defeated later if any of its subarguments is defeated. Since RecommendedAg cannot send a new attack to UDA1 . if any argument in the dialog supporting α is accepted because it has not been attacked (case S1. In this manner.b) or all its attackers have been 5 This fact could even be the conclusion of a DiscarderAg’s argument. Thus. the assertion “Peter likes musicals”. and a new evaluation method will be needed here. if possible. a new argumentation path must be explored.g. if argument URA1 and its counterargument UDA1 uttered in S2.b is possible due to the agents totally share the domain knowledge. Therefore. Suppose that DiscardedAg receives an argument WRA containing a RecommendedAg’s local categorical premise (e.. The argumentation process ends when an agent has run out of new arguments for attacking the ones received previously. the result of the evaluation is unresolved. non-shared facts only asserted by one agent). Likewise. Example 12. However. the main goal of the argumentation process is to determine whether the initial claim α is accepted or not by taking into account all the related arguments. if the agents could have their own local categorical facts (i. . Firstly.6.4.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 355 above. the formalization of the argumentative dialog is given elsewhere [33]. To this end. Hence. since V contains prevailing information.b could lead to the second situation of inconsistency detected in Section 12. Two special situations must be noticed in this process. Let us see an example.3. DiscardedAg could have another local grounded fact5 asserting “Peter does not like musicals”. the case S1.b. the interested reader is referred to [32]. As a result. the case S2. where two arguments based on the same type of information support conflicting conclusions.. programs’ features) are also accepted by both agents and cannot be attacked. which is in conflict with WRA local assertion. An example of this situation is given in Scenario 3 of Section 12.b are both based on premises which are all grounded assertions.e. V U when argument V attacks argument U according to the situation S3. Secondly.

RuleDislikeP}) . the following arguments are built: URec1 = (RecommendedProg(BBCNews). {ImportantProg(BBCNews).3 SmartTV Scenarios The following scenarios evaluate the use of the two decision mechanisms previously introduced. {Prog(BBCNews). and J. BBCNews) Review(TheTimes 001).TheTimes 001). Centralized entity: From the information listed above and the domain rules given in Section 12.3. {Viewer(Peter). Tag(TagBBC1). RuleRev}). Contrarily. several arguments are given in favor and against the inclusion of a program in the final list of recommendations. two situations are developed: 12.5. 12. TagBBC1). BBCNews). J. U¬Fav = (¬FavoriteProg(BBCNews). Let us see some examples in the next subsection. hasReview(BBCNews. dislikesProg(Peter. Notice that the state of affairs in each scenario is represented by means of OWL-DL assertions. RuleInt}).8) Tag(TagBBC1). α is accepted by the system. rating(T heTimes 001. RuleUrg}). The following information is collected by the Information module: • • • • Program(BBCNews) Viewer(Peter). RuleImp}). Augusto n ı defeated by other accepted arguments. if all the arguments for α have been defeated. it is rejected. Mu˜ oz. hasTag(BBCNews. hasReview(BBCNews. T heTimes 001).A. In each of them. hasTag(BBCNews. which are contained in the SmartTV ontology. URec2 = (RecommendedProg(BBCNews). Urgent). Bot´a. {InterestingProg(BBCNews). Scenario 1: BBC News In this scenario. RuleNoFav}).3. Program(BBCNews). Program(BBCNews). tagValue(TagBBC1. 8). tagValue(TagBBC1. dislikesProg(Peter. This information is then delivered to the Decision module to set BBC News as recommended or not. UDsc = (DiscardedProg(BBCNews).1 Case 1. the SmartTV system discovers the viewer Peter in front of the TV.C.356 A. {¬FavouriteProg(BBCNews).5. According to the decision mechanism employed. UImp = (ImportantProg(BBCNews). {Review(T heTimes 001). TagBBC1) The above OWL-DL assertions represent the ongoing situation in the domain.Urgent). rating(TheTimes 001. URev = (InterestingProg(BBCNews).

to recommend a program because has an urgent annotation is more convincing than to use newspaper information. The persuasion dialog developed from this claim is depicted in Figure 12.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 357 Observe that there are two arguments supporting BBC News as recommended program. Likewise.5. argument U¬Fav is the only one affected by the criteria FavCriteria+/− in the second level of the scale. and moreover there are no other acceptable arguments which are inconsistent with it.e. UDsc type of information ImportantProg. DiscarderAg (Dag) builds the above arguments UDsc and U¬Fav from the knowledge received from the Information module and the rule set B. it is given as justification.g. ImportantProg(BBCNews) is preferred to InterestingProg(BBCNews)). the arguments passes to the next level with the rest of arguments which can not be evaluated yet.e. so URec1 URec2 . where the criteria RecCriteria+/− is applied. supported by the Peter’s dislike for it (U¬Fav).5. The first level of the scale contains the criteria IntCriteria+/−. UImp ) and a positive review of a newspaper (URev ). URec2 and UDsc . Therefore. However. Therefore. Argument URev is the only one affected by this criteria. URec1 is based on the UDsc ..14.. the entire set of arguments reaches the top level. according to RecCriteria+/− (i. URev and UImp shown in the Case 1 from the knowledge received from the Information module and the rule set A. However. Contrarily. Argumentation process: RecommenderAg (Rag henceforth) creates the arguments URec1 . ImportantProg(BBCNews) is preferred to ¬FavouriteProg(BBCNews)). Finally. breaking news. Since URec1 is based on more relevant information according to RecCriteria+ (i. it is based on a type of information less prevailing than URec1 is.2 Case 2. Namely. based on an urgency annotation (e. URec2 . On the other hand. UDsc can not be used as counterargument. Thus. The argumentation process starts with Rag claiming BBC News as a recommended program. BBC News is accepted as recommended program because URec1 is an accepted argument after applying the scale. The type of information related to FavouriteIn f oGroup is more relevant than the InterestingIn f oGroup one. an argument for discarding the program is found.. it must select between arguments URec1 and URec2 . Three arguments are evaluated with this criteria: URec1 .3. respectively. Dag’s argument UDsc is inconsistent with URec1 . As it is not defeated by another argument according to the criteria. Dag tries an attack to URec1 . When Rag is asked for justifications of this claim. 12.1. This set of arguments is now evaluated according to the domain criteria scale defined in Section 12. since the viewers’ . Now.

and the program is added to the list of recommendations. Terror) Viewer(Chris). likesGenre(Peter. . Augusto n ı preferences should not block important news (URec1 assertion.14 Argumentation dialog to decide on recommending BBC News Scenario 2: Friday the 13th In this scenario. 13) This information is delivered to the Decision module to set Friday the 13th as recommended or not.358 A. Figure 12. The following information is collected by the Information module: • • • • Genre(Terror) Program(Friday the 13th ). Dag concedes the Rag’s claim about recommending BBC News. UDsc ).A. hasAge(Chris. According to the decision mechanism employed. Dag must accept it. Moreover. Since this argument is totally composed of categorical facts. Dag also accepts URec1 since the agent does not have stronger counterarguments nor URec1 has any other derived assertions. as it is a derived Rag responds the argument UImp as justification for ImportantProg(BBCNews). Finally. The only possibility for Dag to attack URec1 is to ask for its justification of ImportantProg(BBCNews). 18) Viewer(Peter). hasGenre(Friday the 13th . and J. Mu˜ oz. PG(Friday the 13th .C. J. the SmartTV system discovers viewers Peter and Chris in front of the TV. Bot´a. Terror). two situations are developed.

the following arguments are built: VRec = (RecommendedProg(Friday the 13th ). The argumentation process starts with Rag claiming Friday the 13th as recommended program.5. Here. and only the argument VFav is evaluated (as accepted) by the second level (FavCriteria+/−). 13). 18). hasGenre(Friday the 13th . On the other hand. RuleFav}). argument VDsc discards the program due to the information about the Chris’s age and the parental guidance of the movie (VExp ). RuleLikeG}). Dag searches an attack to that argument: the counterargument VDsc rebuts VRec . Dag builds the above arguments VDsc and VExp from the knowledge received from the Information module and the rule set B. Finally. Terror). hasAge(Chris. Contrarily.5. RuleAge}) Observe that argument VRec supports Friday the 13th as recommended since Peter likes terror movies (VFav).e. {ExplicitProg(Friday the 13th ). Friday the 13th is accepted as discarded program because VDsc is an accepted argument after applying the scale. none of the arguments are affected by the first level (IntCriteria+/−). Then. Rag answers it when the claim is challenged. Therefore. likesGenre(Peter.3 Case 1.3. As a result. VDsc is based on more relevant information according to RecCriteria+/− (i. VExp = (ExplicitProg(Friday the 13th ). the whole set of arguments reaches the top level.15. Argumentation process: Rag creates the arguments VRec and VFav shown above from the knowledge received from the Information module and the rule set A. VDsc VRec and Dag can attack VRec directly with its argument VDsc . arguments VRec and VDsc are detected as rebutting each other. According to the scale criteria. Program(Friday the 13th ). {Viewer(Peter). The persuasion dialog developed from this claim is depicted in Figure 12. By employing RecCriteria+/−.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 359 12. 12. {FavouritetProg(Friday the 13th ). VDsc VRec since the type of information ExplicitProg is more relevant than FavouriteIn f oGroup.3. Centralized entity: From the information listed above and the domain rules given in Section 12. Program(Friday the 13th ). Namely. Genre(Terror). {Viewer(Chris). VFav = (FavoriteProg(Friday the 13th ). ExplicitProg(Friday the 13th ) is more relevant than FavouriteProg(Friday the 13th )). VDsc = (DiscardedProg(Friday the 13th ).3. and moreover there are no other acceptable arguments which are inconsistent with it. PG(Friday the 13th . As the only argument supporting this claim is VRec .. and moreover. an argument saying that the program has explicit content must .4 Case 2. Terror). RuleExp}).

Mu˜ oz. trying to undercut this argument. Augusto n ı be considered as more important than those recommending the program due to viewers’ preferences. only an age rating is allowed. Moreover. 18) in VExp is a local categorical fact only maintained by Dag. Notice that the acceptance of the argument VExp by Rag is possible due to all the premises in it are categorical facts shared by both agents.2. and J. J. Its only possibility is to challenge the derived assertion ExplicitProg(Friday the 13th ) in VDsc . Rag has to accept VExp . Dag utters VExp as the justification for the derived premise. and eventually the program is not included in the list of recommendations. Rag can not accept the argument VExp . Since all the premises in this new argument are categorical facts. Suppose also that Rag holds the local categorical fact PG(Friday the 13th . 12).15 Argumentation dialog to decide on recommending Friday the 13textth .A. since the premise PG(Friday the 13th . Figure 12. However.5. suppose that the premise PG(Friday the 13th . 12). Rag detects that its best argument for supporting the initial claim has been attacked. 18) is in conflict with its own local fact PG(Friday the 13th . In this new situation. but it is out of the scope of this chapter. Afterward. Now. meaning that for the same program. and therefore VDsc must also be accepted by this agent. the relationship PG is defined as functional in the SmartTV ontology. Bot´a.360 A. a new method to resolve this “local-local” conflict is needed. As the only argument supporting the initial claim has now been defeated. Rag must retract the recommendation for Friday the 13th .C. As commented in Section 12.

hasGenre(Fringe. hasReview(Lost. likesProg(Peter. Program(Fringe). rating(MZine 1. {Viewer(Lois). the latter subargument is in turn supported by the Lois’s dislike for the genre of the similar program Fringe (subargument T¬Fav2 ). Fringe). respectively. According to the decision mechanism employed.3. Lost). (DiscardedProg(Lost). RuleDislikeP}). MZine 1). two situations are developed. TDsc1 and TDsc2 discard Lost. T¬Fav1 TFav2 T¬Fav2 {Viewer(Peter). . {¬FavouriteProg(Fringe). Program(Lost). the following arguments are built: TRec1 TRec2 TDsc1 TDsc2 = = = = (RecommendedProg(Lost). Fringe) Viewer(Lois). 8). Program(Lost). RuleLikeP}). hasGenre(Fringe.5. {Viewer(Lois). likesProg(Peter. rating(MZine 1. TInt = (InterestingProg(Lost). hasReview(Lost. RuleNonInt}). dislikesProg(Lois. Action) Viewer(Peter). RuleFav}). Action). likesProg(Peter. MZine 1). RuleDislikeG}). {¬FavouriteProg(Lost). Moreover. Program(Lost). Lost). {Viewer(Peter).5 Case 1. T¬Int {Review(MZine 1). = (FavoriteProg(Fringe). Fringe). TFav1 = (FavoriteProg(Lost). {FavouriteProg(Lost). = (¬FavoriteProg(Lost). based on the Peter’s preference (subargument TFav1 ) and the positive review of a magazine (subargument TInt ). Program(Fringe). 8). 12. {InterestingProg(Lost). Program(Lost). Genre(Action). dislikesProg(Lois. = (¬InterestingProg(Lost). according to the Lois’s preference (subargument T¬Fav1 ) and the similarity of the program with another one classified as non-favourite (program Fringe in subargument T¬Int ). dislikesGenre(Lois. On the other hand. = (¬FavoriteProg(Fringe). hasSimilar(Lost. the SmartTV system discovers viewers Peter and Lois in front of the TV. (RecommendedProg(Lost). (DiscardedProg(Lost). {¬InterestingProg(Lost). Lost). RuleNonFav}). RuleInt}). Action) This information is delivered to the Decision module to set Lost as recommended or not. arguments TRec1 and TRec2 support the recommendation of Lost. RuleRev)}). Centralized entity: From the information listed above and the domain rules given in Section 12. Fringe) Program(Fringe). RuleLikeP}).3. hasSimilar(Lost. dislikesGenre(Lois. RuleNegSimilar}) In this scenario. The following information is collected by the Information module: • • • • • Genre(Action) Program(Lost). Action).Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 361 Scenario 3: Lost In this scenario. Lost). likesProgram(Peter.

TFav1 . T¬Fav2 ). since T¬Int is supported by T¬Fav2 . Dag builds the above arguments TDsc1 . and in turn. T¬Fav1 .5. if its supporting subargument is defeated. Again. {TRec1 . The reason for this elimination is founded on the dependence of TRec2 upon TInt . Thus. On the other hand. DirectIn f oGroup. Augusto n ı Let us apply the scale of domain criteria to this set of arguments. Here we find the pairs of inconsistent arguments (TFav1 . T¬Fav2 is removed from the set of arguments.362 A. It means that both arguments are equally acceptable. and the final decision about recommending/discarding Lost can not be set by the SmartTV system. Therefore. Consequently. and they are evaluated according RecCriteria+/−. Argumentation process: Rag creates the arguments TRec1 . T¬Int } are the remaining arguments in the set when the top level of the scale is reached. On the other hand. TRec2 . Hence. A relationship of defeat between the arguments in the first pair can not be established. and J. TDsc2 . T¬Fav2 and T¬Int from the knowledge received from the Information module and the rule set B. there is not a defeat relationship between TRec1 and TDsc1 . the two arguments are based on the same type of information. The argumentation process starts with Rag claiming Lost as recommended program. TFav2 and TInt shown above from the knowledge received from the Information module and the rule set A. TFav2 . 12.. Now. Mu˜ oz.A. T¬Fav1 ) and (TFav2 .e. TDsc2 is supported by T¬Int . and besides the argument TRec2 is also removed from the set. TFav1 . T¬Int TInt as the former is based on the type of information SimilarGroupIn f o.C. arguments TInt and T¬Int rebut each other about the degree of interest on Lost and both can be evaluated according the criteria in Level 1. TDsc1 . As a result. J. arguments TRec1 and TDsc1 rebut each other in the conclusion of recommending/discarding Lost. TInt does not pass to the next level. which is more relevant than the type ReviewGroupIn f o used in TInt . In this . TRec2 is then defeated too. both arguments are held in the set. Eventually. Firstly. IntCriteria+/−. Consequently. the system asks the viewers to decide about that program. since TFav1 and T¬Fav1 are both based on the same type of information. Next. with an unresolved relationship between them (i. such as TFav2 T¬Fav2 since the former is based on a direct preference on the program Fringe. Bot´a. Lost cannot be set as favorite neither as non-favorite). together with T¬Int and TDsc2 .6 Case 2. T¬Fav1 . the second level of the scale evaluates inconsistent arguments based on the criteria FavCriteria+/−. As a result. and T¬Fav2 only relies on the Lois’s dislike for the genre of that program.3. FavouriteIn f oGroup. a prevailing relationship between arguments in the second pair exists.

Then. since they are based on the same type of information. which is its only justification for InterestingProg(Lost). Moreover. so Dag can not use it here. Now. Then. However. the agents build two argumentation paths in the dialog. Rag utters the message “Undecided(FavouriteProg(Lost))” to notify to Dag that the classification of that program as favorite/non-favorite can not be established. The first branch of the dialog is depicted in Figure 12. and both are based on the same type of information (DirectIn f oGroup according to FavCriteria+/−). Dag detects its argument T¬Int supporting ¬InterestingProg(Lost) as a possible attack to TInt .17. Furthermore.2. Namely. In that moment. Neither can Rag find an new attack to T¬Fav1 . Dag challenges first this derived assertion. namely FavouriteGroupIn f o.16). Rag finds a new argumentation path after going back in the dialog and evaluating the question “Why(RecommendedProg(Lost))” again. the viewer’s preference is a better type information than that based on reviews. However. According to RecCriteria+ . The Rag’s response for this new challenge is the argument TInt .. instead of uttering argument TDsc2 to attack TRec2 . both arguments TRec1 and TRec2 can be used by Rag. TRec1 TDsc2 . which is its only justification for FavouriteProg(Lost).16. Dag detects its argument TDsc2 to be in conflict with TRec2 . as argument T¬Fav1 is also composed of categorical premises uniquely.5. as shown in Figure 12. Thus. as they support Lost as discarded program. As a result. The previous argument TRec2 can be employed as a new justification for the initial claim. TRec1 is used as justification for the initial claim in this branch. As a result. the premise InterestingProg(Lost) in TRec2 is a derived assertion. After receiving this argument. nor does Dag have a better argument to attack TFav1 . Therefore. i. Moreover. and equally strong to this according to RecCriteria+/−.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 363 case. Dag can only utter T¬Fav1 to attack it. Dag challenges first this derived assertion (see Figure 12. the case S2. TRec1 is preferable to TRec2 since the type of information in the premise FavouriteProg(Lost) is more relevant to the information in InterestingProg(Lost). The Rag’s response for this new challenge is the argument TFav1 . It starts with Rag giving a justification for the initial claim after the challenge posed by Dag to that claim. Dag detects its argument T¬Fav1 supporting ¬FavouriteProg(Lost) as a possible attack to TFav1 . Dag detects that its arguments TDsc1 and TDsc2 are in conflict with TRec1 .e. According to RecCriteria+/− . instead of uttering argument TDsc1 to attack TRec1 . Now. the agents have now arrived to a blocking situation. T¬Int is supported by . the premise FavouriteProg(Lost) in TRec1 is a derived assertion. Since argument TFav1 does not have any derived premise. TDsc1 is as stronger as TRec1 .b commented in Section 12.

which classifies Fringe as favourite based on the Peter’s direct preferences.364 A. Fringe)). Dag concedes the program Lost as recommended. TFav2 directly (see Figure 12. Moreover. neither can the agent challenge it again. this agent does not have more counterarguments. whereas TInt is supported by a premise in the ReviewGroupIn f o (hasReview(Lost. as an attack to T¬Fav2 . Bot´a. Afterward. the information based on programs similarities is more important than that based on reviews. which relies on Lois’s dislike for the genre “Action” to classify Fringe as non-favourite. Eventually.C. all the premises in TFav2 are categorical facts. Dag utters T¬Int to attack TInt . Rag finds the argument TFav2 . T¬Int directly. Now it is the turn for Dag to find an attack to TFav2 . Action). J. The first argumentation path is left unre- a premise in the SimilarGroupIn f o group (hasSimilar(Lost. Following the same steps. By accepting it. Due to TFav2 uses the premise likesProg(Peter.16 solved Argumentation dialog to decide on recommending Lost. Mu˜ oz. since Dag does not have new counterarguments for TInt . Dag also accepts the argument TRec2 . Augusto n ı Figure 12. The premise ¬FavouriteProg(Fringe) is detected as a derived assertion in T¬Int . Fringe) and TFav2 employs dislikesGenre(Lois. Dag answers the argument T¬Fav2 as a justification. so Dag must accept this argument. Subsequently. T¬Fav2 . and therefore Rag utters it TInt . Namely.A. After receiving the counterargument T¬Int . However. Rag needs to attack it to reinstate its argument TInt . MZine 1)). According to IntCriteria+/−. Consequently.17). Dag must then retract its claim ¬FavouriteProg(Fringe)). so Rag challenges it. and J. TInt is also accepted.

but two main problems are detected when using the scale. Firstly.5. but this task is not scalable.6 Comparing Decision Mechanisms in the SmartTV System The scenarios in Section 12. Contrarily. The first and most important contrast among the two approaches resides in the application of the domain criteria.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 365 Figure 12. all the domain criteria must be pre-arranged before employing the scale. the argumentation . The hierarchical scale in the centralized entity avoids the exchange of arguments and thus the necessity of defining a dialog as in the argumentation process.17 Argumentation dialog to decide on recommending Lost. The second argumentation path ends conceding the program as recommended 12. It could be easy to put in order a low number of criteria. to find the optimal order among all the criteria could be a rather handmade and fallible task. Moreover.3 have revealed several differences when using a centralized entity or an argumentation process as the decision mechanism in the SmartTV system.

Bot´a. J. A second difference between both approaches is the possibility of including local grounded facts. as demonstrated in Scenario 3-Case 2. i 1. argument TInt is ruled out by argument T¬Int in Level 1.A. and then they are eliminated according to each level in the hierarchical scale. the outcome in Scenario 3-Case 1 is an unresolved decision due to the pre-arranged order of domain criteria and the failure of reinstating arguments during the evaluation process. and J. and then TInt is restored again when TFav2 undercuts T¬Int . since they are re-evaluated in the dialog as shown in Scenario 3-Case 2 of Section 12. the centralized entity can not represent this kind of facts. Indeed. the argumentation process is the only suitable approach. A step-by-step process showing the creation of the dialog tree in the argumentation process could be implemented as a graphical tool. Therefore. Finally. where each agent could keep its own separate facts about elements in the domain. TInt (and the arguments that are supported by it) should be reinstated in the set of argument. Augusto n ı process does not enforce a pre-defined order in the use of domain criteria. as previously evaluated arguments are not taken into consideration again. Contrarily. since all the information managed by this entity comes as if provided by a sole source. where all the arguments are shown at the same time from the beginning of the decision process. Contrarily. this situation is shown in Scenario 3-Case 1 of Section 12. this is not possible in the hierarchical scale.C. the viewer has to cope with more information when asking a explanation to the central entity. the correct solution is to add the program as recommended. Secondly.3. the hierarchical scale presents a static and fixed evaluation of arguments. However. This option is feasible when employing a distributed architecture as in the argumentation process. this argument TInt is later defeated in Level 2. the argumentation process approach is preferable when a distributed archi- .3. in scenarios where asserting different local grounded facts is enabled. Here. In fact. Instead of this. thus providing the minimal information to understand the interaction between arguments and the final outcome. As a result. However.366 A. an argument ruled out in the level n can not be reinstated even though the argument(s) defeating it is(are) ruled out in subsequent upper levels n + i. the argument TInt is first ruled out by T¬Int . Mu˜ oz. the use of a dialog for exchanging arguments is more intuitive and easy to show in case an explanation about the decision of recommending or not a program is requested by the viewers. Nevertheless. Hence. Therefore. and the agents select the most suitable one depending on the argument that is currently being processed.5.5. the argumentation process does allow for the reinstatement of arguments. In conclusion.

and moreover. To this end. Decision-making is hard for any artificial system. as it is focused on a well-defined and restricted set of choices. may not be that disastrous after all. This type of applications will become more common as the area of Intelligent Environments advances and transitions from a one-user focus to a many-users focus. arguments could be supported by other sub-arguments. Still the scenario we have considered offers a good starting point to explore the improvement of artificial systems for taking decisions. several domain criteria are defined. and a justification for that conclusion. These criteria establish relationships of relevance among the type of information employed in the arguments’ premises.7 Conclusions and Further Research Directions The problem we addressed in this chapter offers plenty of challenges for the scientific community. Now. In turn. This system is integrated with argumentation techniques as a means of detecting and solving the conflicts derived from different viewers’ preferences and TV programs data. We believe argumentation systems offer a valuable tool in these cases as a way to naturally encode different views and preference mechanisms. Whenever there is a group of humans there will be a difference between their needs and preferences. we propose two decision . Consequently. We humans find it hard too. Hence. In order to measure the information contained in the arguments. and the consequences. the same conclusion could be supported by different arguments. In order to illustrate this problem. and a more compelling explanation.Intelligent Decision-Making for a Smart Home Environment with Multiple Occupants 367 tecture can be employed in the decision mechanism. it is not sufficient to just evaluate conflictive arguments. They consist of a conclusion. when conflictive conclusions are detected. the possibility of using different local facts. the AmI system uses the information given by their supporting arguments to take a decision on which conclusion is preferable. and therefore a need for mediation. the new information inferred in the AmI system is presented as arguments. which comprises a set of premises and a domain rule entailing the conclusion by means of a deductive reasoning process. In this manner. but also all arguments related to them. representing the new inferred piece of information. an argument based on a more relevant type of information has a more preferable conclusion that another argument with a conflictive conclusion using less relevant information. in case of misjudgment on behalf of the artificial system. 12. we have introduced here an Ambient Intelligence (AmI) system controlling a smart TV recommending service. since it offers a flexible and dynamic implementation of the domain criteria. There are plenty of disasters reported in human history due to errors in human judgment.

it is necessary to provide ways for the argumentation system to: (a) gather ID information of the different users the system should be aware of (i. the final goal of the AmI system presented here is to recommend/discard a set of TV programs depending on the current viewers detected in front of the TV. although solutions for them have been explored before and they seem viable. Two questions are relevant here: How can this preferences be fed to the system or learn by it in a natural way (minimizing work for the users)? What mechanisms to chose the best options are adequate in each case (each group of people in each house)? All these questions remain challenging. There are at least two bottlenecks the community should address. Different type of information about the viewers’ preference. This process is developed by means of a persuasion dialog between two agents which exchange arguments to defend their claims about conflictive conclusions.. As a result. Bot´a. The first approach consists in a centralized entity gathering all the arguments built in the system and evaluating them through a hierarchical scale of domain criteria. Finally. We are now in the historical position to put these systems to work and help humans in daily life matters.A. etc. knowing Andr´ s. a comparison between both approaches is given. Augusto n ı mechanisms. Then. J.e. (d) detect an implicit demand for the system to act (do not miss an opportunity to mediate positively).. is a user).368 A. the two decision mechanism approaches are instantiated in the smart TV service by means of specific use cases.g. One is the connection between the argumentation system and the environment. A set of domain rules is defined to reason about this information and entail program recommendations/discards.C. (c) update the information easily as there are new users or they change their preferences with time. TV programs containing sex or violence may be discarded by the system). and offer the opportunity to bring science to our living rooms. cast. Mu˜ oz. together with several domain criteria to evaluate the arguments built from the reasoning process. The second approach uses a distributed architecture to implement an argumentation process. (b) gather the different views e of the users (who likes what). if an under-18 viewer is detected. information about the viewers’ age is also considered in order to take more intelligent decisions (e. The other major challenge for these systems is to acquire good preference mechanisms in a natural way. Still there is a long way to go until these systems become trustworthy and accepted. along with TV program information (parental guidance.) collected from the Internet constitutes the domain model in this scenario. Moreover. genre. reviews. aged 26. The importance and relevance of this step should not be underesti- . Thus. and J. (e) know when it is pertinent and adequate for the system to act (do not interrupt too much).

Secaucus. The computer for the 21st century. Still this is a necessary step for AmI in its road to widespread acceptance and sooner or later the scientific community will have to make that committed and focused investment in this problem. In The invisible future: the seamless integration of technology into everyday life. real-traffic conditions. Rotterdam. (1999). 94–104.. (1991). (Springer-Verlag New York. Schuurmans. This chapter represents a natural first answer to one of the challenges highlighted recently in [38]. Ambient Intelligence. Bibliography [1] M. In the first scenario. and M. R.g. 35. and E. and also thank the Fundaci´ n S´ neca o e for the grant “Programa de Ayuda a los grupos de excelencia 04552/GERM/06”. M. 7–11. Aarts. 235–250. 11(1). McGraw-Hill.. the argumentation techniques could help to select an action when a possible alarming event (e. a person falls) is detected. the argument process may advise the driver about which route is preferable according to her profile and preferences. IEEE Multimedia.g. Acknowledgements The authors would like to thank the Spanish Ministerio de Ciencia e Innovaci´ n for spono soring this work under the FPU grant AP2006-4154. Weiser. The Netherlands. IT Monitor. Inc. Rabaey. 12–19. Harwig. 9. (010 Publishing. supported by weather information. The New Everyday: Visions of Ambient Intelligence.. Ambient Intelligence: a multimedia perspective. It may not represent a final and conclusive solution to the challenge but it is a first step in that direction the overall complexity involved in the challenge is huge and will take probably many years for a community as a team to be able to deliver systems which are aware of a group of users and capable to optimize the way they are satisfied as a group. as for example Assistance Living [34. J. Several arguments will be built for and against different routes. [6] W. 2003). (2002). Marzano. etc. USA. Aarts. . Aarts. NJ. Weber. 36] or Smart Cars [37]. [4] E. 2005). Moreover. Scientific American. pp.. In the Smart Car scenario. (2004). [3] E. it could be seen as a reference toward the application of argumentation techniques to other AmI scenarios. number of times that a route has been selected.Computational Intelligence in Complex Decision Systems 369 mated. We hope this chapter will inspire other colleagues to develop AmI in that direction. Ambient Intelligence. Ambient Intelligence. 265(3). [2] E. Eds. This action will be reached as a consensus between the agent on behalf of the person living in a house (e. Appelo. Aarts and S. Aarts and L. [5] E. a patient or an elderly person) and the agent on behalf of the nurse caring for that person.

In Communication in Multiagent Systems. Agent dialogues with conflicting preferences. 2006). Rahwan and P. Das. (State University of New York Press. (1993). London. Parsia. Web Ontology Language (OWL). (2008). [26] D. 164–178. Augusto. LNCS. [23] F. L. 3(3). J. F. In ATAL ’01: Revised Papers from the 8th International Workshop on Intelligent Agents VIII. W3C. Augusto n ı [7] C. and applications. J. Elements of Argumentation. PatelSchneider. D. 190–205. N. The Semantic Web. Eds. Berners-Lee. Springer-Verlag. Ambient Intelligence: The Confluence of Ubiquitous/Pervasive Computing and Artificial Intelligence. D. 287–318. implementation. (2001). [15] J. Augusto and G. Ed. and Y. 26–42. and O. A. (2003). Augusto and C. Protocols and Applications.C. Lassila. Xu. Li. K. Amgoud and S. M. R. The Description Logic handbook: theory. Vrije Universiteit Amsterdam. 15–18. UK. (2009). and L. vol. C. 284(5). IEEE Explore. ISBN 978-0-471-54448-7. McGuinness. Mu˜ oz. In eds. Argumentation-based communication between agents. Integration of Knowledge Intensive Multi-Agent Systems. and K. 3rd International Conference on Pervasive Computing Technologies for Healthcare 2009. [16] S. J.. Zhang. W. Connoll. Krabbe. Mcguinness. Mcdermott. Ambient Intelligence: applications in society and opportunities for Artificial Intelligence. Commitment in Dialogue: Basic Concepts of Interpersonal Reasoning. Solanki. Stein. D. Burstein. Hexmoor. Smart Environments: Technology. Technical report. 1995). and J. Cardoso and A. Calvanese. Walton and E. J. S. (2004). F. 2003). Portugal. Shapiro. [20] L. D. Springer London. D. (2001). D. Sheth.370 A. 2005). The development and prospect of personalized TV program recommendation systems. McBurney. and V. [8] D. Horrocks. Paolucci. Baader.-J. London. Hendler. In MSE ’02: Proceedings of the Fourth IEEE International Symposium on Multimedia Software Engineering. 3387. [12] (to appear). (2002). [13] H. Parsons. J. (WileyInterscience. Cook. pp. Fifth International Workshop on Argumentation in MultiAgent Systems (ArgMAS’08). 23(2). and J. (to appear). Cook and S.A. (MIT Press. [18] I. Srinivasan. An adaptive preference learning method for future personalized TV. 213–234. Lee. pp. (2009). pp. IEEE Computer Society. IEEE Computer Society. Augusto. Bringing Semantics to Web Services: The OWL-S Approach. Designing Smart Homes. M. Bot´a. 119–149. D. Argument-based applications to knowledge engineering. Simari. 34–43. [11] J. 5384. [9] D. (Cambridge University Press. C. R. Moraitis. Sabou. Estoril. (2005). 82–95. 2008). [21] M. In eds. NY. Temporal defeasible reasoning. Knowledge Engineering Review. Robertson. vol. M. [17] P. J. M. ISSN 1541-1672. and P. Sycara. (2009). J. van Harmelen. [10] J. Besnard and A. C. (Springer-Verlag. V. F. (2002). SWSWPC 2004. Patel-Schneider. Parsons and P. Jakkula. J. UK. Lu. [22] T. Hunter. . McGuinness. [14] D. 260– 264. [25] J. Martin. L. C. Eds. Lecture Notes in Artificial Intelligence. B. Payne. Springer. Knowledge Information Systems. Scientific American. I. UK. Springer. H. Pervasive and Mobile Computing. pp. 15(2). pp. L. T. (2000). Hendler. IEEE Computer Society. Dean. Hua. Intelligent Computing Everywhere. Nugent. Albany. D. Thompson and H. Carbogim. Ramos. [19] D. (2004). C. Shi and J. Ambient Intelligence: The Next Step for Artificial Intelligence. and D. D. Logical tools for modelling legal argument. The Role of Artificial Intelligence. (2007). PhD thesis. Augusto. N. C. [24] X. P. IEEE Intelligent Systems. Nardi. C. London. ISBN 3 540 35994 X. Prakken. pp. Mcilraith.

(2008). McCullagh. A. T. (MIT Press. Natale. A. M. H. Jena: implementing the Semantic Web recommendations. Cognitive Carpentry: A Blueprint for how to Build a Person. In ECAI ’92: Proceedings of the 10th European conference on Artificial intelligence. 9–24. o 7th International Conference on Practical Applications of Agents and Multi-Agent Systems (PAAMS’09). and M. ACM Press. NY. 1995). (2009). and J. 74–83. Journal of Web Semantics. USA. A. vol. (to appear). Katz. Dickinson. J.-C. (2009). J. Yamazaki. C. Grau. and S. A. [38] J. pp. Krause. Intelligent Technology for an Aging Population: The Use of AI to Assist Elders with Cognitive Impairment. T. Wu. Bot´a. P. [30] I. Mu´ oz. and L. and J. In eds. A Software Architecture for an Argumentationn a ı Oriented Multi-Agent System. SWRL: A semantic web rule language combining OWL and RULE ML. Grosof. [33] A. (2007). Prisciandaro. (1992). Okadome. Tealdi. Augusto. . Present and Future of Ambient Intelligence and Smart Environments. J. Plomp and P. Eindhoven. 5180. Horrocks. B. Springer. Fox. Parsia. Y. Springer Verlag. and J. Polleres. P. Bajo. 51–53. (2001). Tabet. S´ nchez. New York. Springer. ICOST. E. Springer. [32] A. 26(2).Computational Intelligence in Complex Decision Systems 371 [27] J. 197–207. 623–627. [29] J. C. B. J. Scapolo. Lecture Notes in Computer Science. D. and K. In Selected Papers from the 1st International Conference on Agents and Artificial Intelligence (ICAART). IST Advisory Group. Aghajan. AI Magazine. pp. Corchado. In eds. K. L. [28] E. Seaborne. [35] F. Reynolds. Wilkinson. MA. and J.-A. Dollin. Pechoucek. P. F. Klusch. USA. F. 5(2). Pellet: A practical OWL-DL reasoner. In eds. Interactive home assitant for supporting elderly citizens in emergency situations. ASBO: Argumentation System Based on Ontologies. Bot´a. contradictions and practical reasoning. [34] M. 2004). J. 4541. and A. Past. EUASI2004. Arguments. Boley. 191–205. Pollack. Demazeau. C. Ambler. Cooperative Information Agents XII. Netherlands. W3C (May. Pollock. Leijten. Ducatel. Carroll. Pav´ n. pp. Patel-Schneider. Walkden. vol. Mu´ oz and J. Inc. C. pp. S. M. Proceedings of Workshop on AmI Technologies for WellBeing at Home. Dean. Cambridge. In eds. C. In Proceedings of the 13th international World Wide Web conference. (2004). Scenarios for Ambient Intelligence in 2010. I. vol. J. B. O. Burgelman. [31] J. [37] K. and Y. (2004). Kalyanpur. Advances in Intelligent and Soft Computing. 196–205. Distributed VisionBased Accident Management for Assisted Living. Augusto. A. (2007). Technical report. Sirin. pp. Lecture Notes in Artificial Intelligence. Makhtari. J. Bogdanovicz. (2005). Technical report. 55. n ı M. Mayora-Ibarra. and M. John Wiley & Sons. [36] H.

A case study is given to demonstrate the potential of the methodology. Successful implementation involves responding to rapidly changing and continually fragmenting global markets by being dynamic. it needs to be able quickly to adjust its outputs to match market demand and to switch rapidly from one variant to another. it would be more suitable to apply a fuzzy integral model. The agile supply chain paradigm relates to the interface between companies and markets. 34357 Ortak¨ y. to approximate the human subjective evaluation process. and driven by customer [2. © 2010 Atlantis Press/World Scientific 373 . context-specific. In other words. an external perspective on flexibility. flexible across the organization. Computational Intelligence in Complex Decision Systems. o ˙ Turkey E-mail: gulcin.Chapter 13 Applying a Choquet Integral Based Decision Making Approach to Evaluate Agile Supply Chain Strategies G¨ lcin B¨ y¨ k¨ zkan u¸ u u o Galatasaray University. most evaluation criteria have inter-dependent or interactive characteristics. Istanbul. For this reason. An agile supply chain focuses on responding to unpredictable market changes D. Ruan. growth-oriented. 13. To meet this challenge. an organization needs to focus its efforts on achieving greater agility such that it can respond in shorter time-frames. hence. Thus.1 Introduction For a variety of reasons product and technology lifecycles are shortening. Atlantis Computational Intelligence Systems 2. in this chapter the conventional multi-criteria approach based on weighted mean has been replaced by the 2-additive Choquet integral. both in terms of volume change and variety change [9].com This chapter proposes a multi-criteria decision-making approach based on the Choquet integral for assessing and selecting a suitable agile supply chain strategy. In the real world.1007/978-94-91216-29-9_13.buyukozkan@gmail. competitive pressures force more frequent product changes and consumers demand greater variety than ever before. Industrial Engineering Department. DOI 10. where it is not necessary to assume additivity and independence. 10]. they cannot be evaluated by conventional additive measures. 22.

A further distinction is that since conventional supply chains are forecast-driven. Simi- . To attain the expected benefits of the agile supply chain. have been forecast-driven. which is relatively simple with only quadratic complexity and enables the modeling of interaction only between criteria pairs.374 G. Selecting a proper method requires an insight analysis among available multi-criteria decision making (MCDM) techniques [16]. and techniques to solve unexpected problems. and forms virtual organizations and production entities based on customer needs. 8]. 6. of necessity. methods. Conventional supply chains have been lengthy with long lead-times and hence. we use 2-additive Choquet integral described next. the agile supply chain strategies evaluation process requires an appropriate multi criteria analysis and solution approach. Recently. It utilizes information systems and technologies as well as electronic data interchange capabilities to move information faster and make better decisions. It places more emphasis on organizational issues and people (knowledge systems and empowered employees). agile supply chains are shorter and seek to be demand-driven. By contrast. there has been a growing interest in the design and implementation of agile supply chain strategies [2. tools. In recent years. For this reason. they are inventory-based. that decision-making can be pushed down the organization. network-based and process aligned. The purpose of this chapter is to increase the understanding of agile supply chain management and its strategies in supply chain networks. The idea of agility in the context of supply chain management focuses on ‘responsiveness’ [10]. The main feature of such aggregation is that it can take into account interaction among decision criteria. Specifically the agile supply chain is market (customer) sensitive. [17] used Choquet integral to aggregate several attributes in the objective function of a flexible job-shop scheduling problem. It is a systemic approach that integrates the business. 25]. It deploys new technologies. B¨ y¨ k¨ zkan u u o and capitalizing on them through fast delivery and lead-time flexibility. virtual. this chapter proposes to use an analytical model in order to support the effective evaluation of agile supply chain strategies. It has been suggested that an agile supply chain has a number of strategies [11. companies have to identify the suitable strategies for them. which is generally ignored in other MCDM methods [7. Tsai and Lu [20] have applied this approach to evaluate the service quality of e-stores while Saad et al. enhances innovations across the company. 3]. More precisely. this study presents a decision framework based on the Choquet integral aggregation for evaluating and selecting agile supply chain strategies. Among numerous methods of MCDM. Due to its multi criteria nature. Agile supply chains are more likely to be information-based [12].

evaluation of agile supply chain strategies as a case study and the results of analysis are discussed.2 Multi Criteria Decision Making with Choquet Integral Until recently. and interactions among them. ordered weighted average) and can thus express a variety of decision maker’s behaviors (severity. Moreover. The next section presents the details of Choquet integral method and the related elucidation procedure to identify strong and weak dimensions of alternatives. we briefly present the Choquet fuzzy integral and its principal properties as an operator of multi-criteria aggregation. min. this aggregation function is too poor to model the three main attitudes of a decision maker: conjunction (tnorm. However. The proposed evaluation process is composed of four main steps: (1) Identify the evaluation main and sub factors of agile supply chain strategies and describe the relationships between them. compromise. it is important that an aggregation operator can satisfy other behavioral properties: possibility of expressing weights of importance on criteria. like the max). In the next paragraph. (3) Evaluate each agile supply chain strategy based on the identified criteria.Applying a Choquet Integral Based Decision to Evaluate Agile Supply Chain Strategies 375 larly. this problem has been partially solved since the introduction of fuzzy integrals [7]. Indeed. max. median. But. like the min). All operators are not able to model all these properties. In the section that follows. Finally. some concluding remarks are given in last section. tolerance) and various effects of importance and interaction among criteria [6]. aggregation operations based on the family of fuzzy integrals include many operators (weighted mean. all the intermediate attitudes can also be imagined [19]. (4) Elucidation and make a decision with proposed approach and its verification. 13. (2) Quantify the importance degrees of each evaluation factor by the Choquet based aggregation technique. Feyzio˘ lu and B¨ y¨ k¨ zkan [5] used this method in a benchmarking procedure to rate g u u o competing systems and find which customer requirements should need improvement for effective product development. compromise (like the weighted mean) or disjunction (t-co-norm. The remainder of this chapter is organized as follows. . the most common tool which is used in multi criteria decision making is the weighted arithmetic mean.

representing the importance of each criterion relative to all the others. . x j ) Ii j <0 Ii j + ∑ xi νi − i=1 n 1 Ii j 2∑ j=i (2) with the fact that vi − (1/2) ∑ j=i |Ii j | 0 and where vi are the Shapley indices. . .2. cn } of n criteria. c2 . xn ) = ∑ ··· x(n) 1.r. . . Also. the 2-additive Choquet integral which takes only a pair of criteria interactions into account [7. . . 13]. s2 . this will further affect the product presentation positively. .376 G. . expressed on a [0. In the next paragraphs. it is assumed that we deal with scores on criteria. . . . x j )Ii j + ∑ max(xi . such that xi is the satisfaction degree of criterion ci . . For example. . For the sake of simplicity. so that. . 0 i=1 According to the application context. xn } represents the numerical satisfaction values of the criteria. s p }. . was the weighted arithmetic mean. if the key can be easily inserted to a lock and if it can also be easily turned. . xk . 100]n → [0. 100] scale. . x2 . 100]. . Each solution sk ∈ S is associated with a profile sk = (xk . Being sufficient in only one criterion can have much less meaning for the decision maker. we recall briefly the notations of the Choquet fuzzy integral and its principal properties as an operator of multi-criteria aggregation In the remaining of the chapter. 7]. x = {x1 . . 100 corresponds to the complete satisfaction of a criterion. . v1 + v2 + · · · + vn = 1. . Let us still consider a set of solutions (candidates) S = {s1 . B¨ y¨ k¨ zkan u u o 13. where xk represents the partial evaluation n i 1 2 (score) of sk related to the criterion ci . Especially. P(C) is the power set of C. The (discrete) Choquet integral of x = (xk . and 0 implies no satisfaction under a criterion. xk ). and . . for each solution sk : Cμ (sk ) = Cμ (xk . xk . xk ) w. .1 Background and Notation Simultaneous satisfaction or dissatisfaction of some criteria can be significant. . The problem of aggregation is the construction of a function Cμ : [0. c(n) }. which is used in MCDM. as it is general the case in the supply chain evaluations. . . Fuzzy integral based operators have been introduced notably in order to consider interactions between criteria to be aggregated [19. . . and a(i) = {c(i) . Until recently.t. . . n (x(i) − x(i−1)) μ (a(i) ) x(1) (1) x(2) where ( · ) in the subscript indicates a permutation such that: x(0) = 0. we will consider only a particular case of Choquet fuzzy integrals known as the 2-additive measure [6]. the most common aggregation tool. . xk ). xk . we will work in a finite universe C = {c1 . x2 . Let μ be a fuzzy measure on C. 6. This type of Choquet integral can be expressed in the interpretable form as follows [6]: Cμ (x) = Ii j >0 ∑ min(xi . n 1 2 μ is defined by: Cμ (x1 . where Cμ (sk ) represents the overall evaluation of the n 1 2 solution sk according to the all n criteria.

Our approach to determine the contribution of partial evaluations is to decompose the aggregated value as a sum of terms where partial evaluations are involved [1. can be expressed qualitatively or/and quantitatively at several levels [4]. Two kinds of elucidations are considered as summarized in the following. we have Cμ (sk ) Cμ (sl ). . and a null value means the criteria are not interacting or are independent. c j ) with values contained in the interval [−1. This understanding of the “why” (and how) of the decisions derived by the fusion system. we are interested in the partial contribution of a specific criterion for a considered fused result. a value of −1 indicates redundancy. Let us consider that the solution that has been elected is sk : it has been adopted following an aggregation strategy (relative importance weights and interactions between criteria) identified with the aggregation operator Cμ . . 13. The coefficients of importance vi and Ii j are more natural to decision maker and can be determined by surveying the decision maker. This elucidation is explained in the “how” (and the “why”) such a decision is taken. • a negative Ii j implies that the satisfaction on ci and c j is sufficient to obtain a significant effect on the aggregated performance. We may thus decompose the Choquet integral into three parts: one conjunctive. 15]. . 1].Applying a Choquet Integral Based Decision to Evaluate Agile Supply Chain Strategies 377 Ii j represents the interactions between pairs of the criteria (ci . Note that. . • zero Ii j implies that there is no interaction between the two considered criteria: the vi indices correspond to classic weighted means. p (where p is the candidatessolutions number).2 Elucidation of the Aggregation Result The need for elucidation aims at providing better quality information to decision maker during phases of process decision. but are not discussed here. Thus for all l = 1. another disjunctive and the third additive: • a positive Ii j implies that the simultaneous satisfaction of criteria ci and c j will have a significant effect on the aggregated performance. .2. and that satisfying only one criterion will have no effect. The why can be quantitatively expressed in terms of the relative influence or dominance of particular criterion satisfactions on the ultimate decision? Indeed. Other methods based on the identification of these coefficients from experimental data exists. A value 1 means a complementarity between the two criteria. it must be verified that the monotonicity conditions are satisfied [6] in order to use the transformation relation with the conventional fuzzy measure representation.

Another aspect of elucidation consists in providing evidence concerning the dimensions according to which ak is preferred over al . The closer is this ratio to 1. the axes along which ak makes the difference over al . it may be pertinent to identify the elements related to the value of sk : it is an absolute elucidation.. This is a relative argument in which the quantities anai .l . We might wish to include several levels of detail in the argumentation: a “one-word” justification. i k k l l where Rk. (3) so that k Δμ( j) xk j) ( k Δμ( j+1) xk j+1) . . B¨ y¨ k¨ zkan u u o First. and Δμ(i) = μ(i) − μ(i+1) .l . It should then be possible to parameterize the level of detail required for the explanation. then: (i) (i−1) (i) (i) k n i=1 k k k k k Cμ (s ) = (μ(1) − μ(2))xk + · · · + (μ(i) − μ(i+1))xk + · · · + μ(n)xk (1) (i) (n) n k Cμ (sk ) = Δμ(i) xk (i) i=1 ∑ (3) k k k k where μ(n+1) = 0.e. al ∈ A. (5) lyzed are the sums of the individual relative potentials Rk. k = l. n − 1. or an exhaustive report. xai ) = Cμ (xak ) − Cμ (xal ) = ∑ Rk. In the case of k 2-additive fuzzy measure. (1) we have: k Cμ (sk ) = ∑ (xk − xk )μ (Ak ). by Eq.378 G. (3) becomes [15]: k Δμ(i) = v(i) + 1 1 ∑ I(i)( j) − 2 ∑ I( j)(i) 2 j>i j<i (4) where. the ( (1) greater is the contribution of the score of the criterion c( j) . detailed reasons. k Δμ( j) xk j) (this term is defined as the absolute potential of criterion c( j) ) into classes relative ( k k to the orders of magnitude of the Δμ( j) xk j) /Δμ(1) xk ratio. . the expression Δμ(i) in Eq. i. .l = Δμ(i) x(i) − Δμ(i) x(i) . For this. i i=1 a a a a n ∀ ak . The latter consists of decomposing the aggregated score in a ranked sum of criteria score contribution. We can initially simply re-rank the terms of the sum Eq. ( We can then partition the absolute contributions of the scores related to the criterion c( j) . ∀ j = 1. . and the more c( j) represents an essential dimension in the decision (local interpretation of elucidation). We use the following analysis criterion: ΔCμ R (xak . v(i) is the relative importance of criterion c(i) and I(i)( j) are the interaction between criteria c(i) and c( j) . let us note: μ(i) = μ (Ak ). the main reason for this preference.

IT Competency IT Capacity IT Compatibility 0.3 Evaluation of Agile Supply Chain Strategies Evaluation criteria: A set of evaluation criteria has to be defined prior to evaluate the identified four main agile supply chain strategies. 2. IT Compet.05 IT Capacity 0.05 Weight 0.40 .30 0. Based on our literature survey [see for instance [9.1 Interactions among and weights of criteria belonging to IT performance.1-13. 13. 22. 23. Figure 13.Applying a Choquet Integral Based Decision to Evaluate Agile Supply Chain Strategies 379 13. 12.1. 13.4. 18. we identify the agile supply chain evaluation criteria as given in Fig. 25.05 0. Table 13.1.05 IT Compat. We assume that in the criteria hierarchy of Fig. only level two criteria belonging to the same first level criterion and all first level criteria are interacting with each others.30 0. 0. Hence possible interactions among all second level criteria are considered in the first level. 10]] and on our expertise in the domain.1 Decision criteria hierarchy for supply chain strategy selection in agile environment Criteria interaction: The interactions among criteria have been identified as shown in Tables 13.

10 Collab.05 0. Collab. Cost 0. Mngnt Quality Mngnt Quality Mngnt Flexibility Mngnt Sustainability 0.380 G.01 Weight 0. IT Perfor. the agile supply chain may be customer (market) sensitive (CS).10 0.05 0. Perfor.05 0.20 Table 13.01 Weight 0. 0.50 0.05 0.05 0. The breakthroughs of the last decade in the form of Efficient Consumer Response and the use of information technology to capture data on demand direct from the point-of-sale or point-of-use are now transforming the organization’s ability to hear the voice of the market and to respond directly to it. 0.05 0. Virtual supply chains are information based rather than .02 Weight 0.30 0.02 Mngnt Flexibility 0. in effect. Collab. 0. The use of information technology to share data between buyers and suppliers is.01 Val.2 Interactions among and weights of criteria belonging to collaborative working performance.05 0. 0.02 Mngnt Sustain. By customer sensitive it is meant that the supply chain is capable of observing the real demand responding to it.3 Interactions among and weights of criteria belonging to management performance. since they have little direct feed-forward from the marketplace by way of data on actual customer requirements.40 0.40 0. B¨ y¨ k¨ zkan u u o Table 13.20 Evaluation of alternatives: An agile supply chain possesses different strategies [11.4 Interactions among and weights of criteria belonging to agile supply chain strategy selection. 3.02 0. IT Performance Collaborative Performance Mngnt Performance 0. Add.05 Collab. Firstly. creating a virtual (V) supply chain. Quality Collaboration Quality Collaboration Cost Value Added Collaboration 0. Most organizations are forecast-driven rather than demand-driven.01 Mngnt Perfor.20 0.50 0. they are forced to make forecasts based upon past sales or shipments and convert these forecasts into inventory. In other words.30 Table 13. 10].

as it is often called. Along with process integration comes joint strategy determination.Applying a Choquet Integral Based Decision to Evaluate Agile Supply Chain Strategies 381 inventory based. if we look into overall aggregate scores. Electronic Data Interchange and now the Internet have enabled partners in the supply chain to act upon the same data i. Finally. These four agile supply chain strategies are accepted as alternatives in this study. common systems and shared information. joint product development. a new style of relationship is essential. 81. 77. results given in Table 13. Conventional logistics systems are based upon a paradigm that seeks to identify the optimal quantities of inventory ands its spatial location. buyer-supplier teams. Based on our expertise in the domain. Paradoxically. In this new world a greater reliance on suppliers and alliance partners becomes inevitable and hence. the route to sustainable advantage lies in being able to leverage the respective strengths and competencies of network partners to achieve greater responsiveness to market needs. (2) is used to compute the aggregate score of each strategy considering only second level criteria.81 for the fourth strategy.88 for the second. rather than be dependent upon the distorted and noisy picture that emerges when orders are transmitted from one step to another in an extended chain. what we are now learning is that once we have visibility of demand through shared information. In the extended enterprise. coordinate and manage the relationships with their partners in a network committed to better. the premise upon which these formulae are based no longer holds. This form of co-operation in the supply chain is becoming ever more prevalent as companies focus on managing their core competencies and outsource all other activities. It can be argued that in today’s challenging global markets. This idea of the supply chain as a confederation of partners linked together as a network based (NB) structure provides the fourth ingredient of agility. We are now entering the era of network competition where the prizes will go to those organizations who can better structure.5. The results are given in Table 13. closer and more agile relationships with their final customers. real demand. we have evaluated each strategy under the identified criteria. transparency of information and even open-book accounting. This implies . we have 72.99 for the third and finally 76. Complex formulae and algorithms exist to support this inventory-based business model.88 for the first strategy. there can be no boundaries and an ethos of trust and commitment must prevail.e. Shared information between supply chain partners can only be fully leveraged through process integration (PI). By process integration is meant collaborative working between buyers and suppliers.6 are obtained. There is a growing recognition that individual businesses no longer compete as stand-alone entities but rather as supply chains. If Eq.

5 Evaluation of supply chain strategies.75 NB (1) 28.75 75.40 67.00 (2) 22.00 (3) 30.7-13.00 (3) 33. For each agile supply chain strategy alternative.50 (2) 19.25 79.50 (2) 22.40 75.10 that the network based strategy should be selected as the best agile supply chain strategy since it has the highest score.00 69.6 Evaluation of supply chain strategies. For others columns.00 87.30 86. Table 13.25 (3) 30. (4) and are given below in Tables 13.00 (1) 27.75 4 86. B¨ y¨ k¨ zkan u u o Table 13.75 1 80.00 (3) 33. Elucidation of the aggregation: According to the proposed elucidation procedure.00 V (2) 26. only the results are given with the criterion rank in the superscript on the left of the value. k Δ μ(i) xk (i) CS (1) 24. the criterion ranking is modified according to the partial score ranking.30 IT Performance Collaborative working performance Mngnt Performance 73.00 3 80.10. 2-additive Choquet CS V Integral NB 80.75 75.70 73.25 2 . We could also rank strategies based on their desirability as NB > V > PI > CS. the quantities are computed from Eq.65 PI 80.382 G. Alternatives V NB PI 90 80 90 80 60 80 90 70 60 80 80 80 90 80 90 80 70 70 70 80 90 90 70 70 80 60 70 Criteria IT Competency IT Capacity IT Compatibility Collaboration Quality Collaboration Cost Value Added Collaboration Mngnt Quality Mngnt Flexibility Mngnt Sustainability CS 70 70 80 80 60 80 80 60 60 Table 13.7 Absolute potentials and aggregation results for IT performance.00 PI (1) 24. The real criterion is indicated in superscript before the detail of the computation for the first column.75 IT Competency IT Capacity IT Compatibility Aggregation Rank 73.

. There is a need to clarify now the rationale of such a ranking. More valuable information is the relative elucidation between the four strategies.20 PI (2) 26.50 Collaboration Quality Collaboration Cost Value Added Collaboration Aggregation Rank 75. By absolute elucidation.e.80 (1) 40.60 NB (2) 18.40 4 87. and thus collaborative performance accounts to 50 % of the overall scores. For almost all strategies.10 (3) 14. to give more explications about “why” these results are obtained.00 (2) PI (3) 17.00 (3) 21.06 76. CS (3) 17. the overall aggregated results indicate that the Network Based strategy has to be selected as the most appropriate strategy.40 (1) 40.30 PI (2) 16.78 (1) 22.20 V (3) 13. and the Virtual strategy takes the second place.20 Mngnt Quality Mngnt Flexibility Mngnt Sustainability Aggregation Rank 67.88 4 19.67 (2) 41. Table 13.65 2 69.80 (1) 40. scores for IT and management performances are approximately equal to the 25 % of the overall score.08 (1) IT Performance Collaborative Performance Mngnt Performance Aggregation Rank 35.30 4 75.16 (1) 20. k Δ μ(i) xk (i) CS (2) 26. (5).20 (2) 29.00 (1) 29.00 (1) 29.22 81. k Δ μ(i) xk (i) CS (2) 13.21 (1) NB (3) 18.10 k Δ μ(i) xk (i) Overall absolute potentials and aggregation results.. the more significant scores come first from collaborative performance.85 NB (2) 30.70 1 79.15 (2) 39.00 (3) 21.9 Absolute potentials and aggregation results for management performance.00 (1) 29.81 3 As previously.44 72.30 3 Table 13.40 3 75.88 2 41.05 (1) 32.45 (3) 14.11 presents the corresponding individual relative potentials according to Eq. then from IT performance and finally from management performance.40 (1) 45. we observe that for all strategies.10 (3) 22.8 Absolute potentials and aggregation results for collaborative working performance.00 2 Table 13.10 1 73.Applying a Choquet Integral Based Decision to Evaluate Agile Supply Chain Strategies 383 Table 13. i.10 (3) 12.52 77.60 V (2) 13.00 (3) 24.99 1 18.28 (2) V (3) 19.

Meanwhile. When comparing NB – CS we observe that collaborative performance is most important and when comparing NB – PI.1 (i) 2. companies have to identify the suitable strategy. Therefore.62 13.4 Concluding Remarks To attain the expected benefits of the agile supply chain approach. This will clearly enrich the assessment capability of decision-makers. Perfor. it is the IT performance that makes the most difference.94 1.78 Mngnt Perfor. An important aspect of the proposed model is its ability to consider decision criteria dependently. 2003. Table 13.14 5. While NB has better scores for IT and collaborative performance. NB does not surpass V on all criteria.01 0. no previous work investigated such a problem and we believe that this research will lead to a greater body of literature on the topic. the conventional multicriteria approach based on the weighted mean has been replaced by the 2-additive Choquet integral.. G. we want to concentrate on the case where evaluation could be made with verbal terms [24.48 3. Bibliography [1] B¨ y¨ k¨ zkan. O. As a second perspective. R3. Berrah. In this chapter. In the near future. Collab. G. it has a slightly lower score on management performance. In traditional multi-criteria evaluation approaches. -0.. For this reason. To our knowledge. 2.37 1.2 (i) R3. and Feyzio˘ lu. in this study. each criterion is assumed to be independent of the others.4 (i) R3. L. we aim to develop a user interface that will facilitate assessments and automate calculations of the proposed methodology.. the characteristics that have interactions and mutual influence among criteria in a real system cannot be handled by the concept of traditional additive measures alone. “Providing Elucidations of u u o g Web Site Evaluation based on a Multi-Criteria Aggregation”. Proceedings of the Interna- . 14]..72 Individual relative potentials.384 G. Mauris. we have proposed a multi criteria decision making framework to identify the most suitable agile supply chain strategy among some discussed in the literature.11 IT Perfor. B¨ y¨ k¨ zkan u u o The NB strategy outperforms the CS and PI strategies on all performance criteria.11 1.

In: M. 28-29 March.D. 366–375. Christopher. J. 1999. 2008. Sugeno. 927–938. O..E. “Multicriterion Decision in Management: Principles and Practice”.. 89–102. Fuzzy Sets and Systems. Meyer. A..-H. B. G. Pomerol. M. Huang... 2000... M. Gunasekaran. Vonderembse. T. 2. A. and Roubens M. 7. 36 (4). “Hierarchical MADM with fuzzy integral for evaluating enterprise intranet web sites”.. 89. 169. 2007. Grabisch. 2006.. European Journal of Operational Research. and Chu. Information Sciences. 21. Dasarathy. and Barba Romero. Hammadi.-H. M. J. “Agility index in the supply chain”. “Elucidation and decisional risk in a multicriteria decision based on a Choquet integral aggregation . Sugeno. and van Hoek.-L. 2008. 409–426.. and Akharraz. 87–106. 5–15. M.-C. Grabisch. “Agile manufacturing A framework for research and development”..). 85... M. M. Information Sciences.Computational Intelligence in Complex Decision Systems 385 [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] tional Fuzzy Systems Association World Congress (IFSA).. S. C. 37–44. “Responsive supply chain: A competitive strategy in a networked economy”.. Gaines (Eds. In M. and Roubens. M. G. Christopher.. M. and Lu. Ou Yang Y. 1. and Dismukes. 239–258. P. M. “Choquet integral for criteria aggregation in the flexible job-shop scheduling problems”. . Chiu. 223–238. “Creating the Agile Supply Chain”. R. Istanbul. 13 (5-6).. 445–456.. Lin. 2000. “An integrated group decision-making approach for g u u o new product development”.A. Wurzburg. 2003. Editors. 285–299. P. M. 2006..-T.a cybernetic framework”. Chartered Institute of Logistics and Transport.. Physica-Verlag. “On the use of the Choquet integral with fuzzy numbers in multiple criteria decision support”. and Roubens. 124. 2000.-Y. Omega. “K-order additive discrete fuzzy measures and their representation”.. Turkey. 176. R. S. International Journal of Multi-Criteria Decision Analysis. J. G. Mauris. I. 3-4. “Application of the Choquet integral in multicriteria decision making”.. 2.C. Marichal. T. 2008. International Journal of Production Economics. 2005.-T. and Cheng. Fuzzy Sets & Systems. M. Industrial Marketing Management. 1996. Lin C. 2006.N. Christopher. Norwell: Kluwer Academic Publishers. ˙ Istanbul. A. 100. P. 2000. Gupta. Stratton. H. Grabisch. 100. North-Holland.. 1997. and Roubens. Grabisch.V... 2006. and Borne. “The strategic integration of agile and lean supply”. S.R. and Chen C.H.. Montmain.. Saad. International Journal of Production Economics. 157. 92. Amsterdam. Gunasekaran. Logistics Management Summit. Feyzio˘ lu. M. 62. International Journal of Production Economics. 1999..... Murofushi and M. Mathematics and Computers in Simulation. A. “Designing supply chains: Towards theory development”. Uppal..M. Benrejeb. 29. 549–564. 76. K. “The Agile Supply Chain: Competing in Volatile Markets”. “The evaluation of service quality using generalized Choquet integral”. 2000. 2. H. “Fuzzy measures and fuzzy integrals: a survey”. Presentation of 9th. Tsai.H. Fuzzy Measures and Integrals: Theory and Applications. 447–462. 2005. 167–189.. 415–434. 183–198. 641–650. Harrison. International Journal of Computer Integrated Manufacturing. and B¨ y¨ k¨ zkan... and Warburton. International Journal of Production Economics.P. Lai. Fuzzy Automata and Decision Processes. “The application of fuzzy integrals in multi criteria decision making”. I. J. R. 1977.. Information Fusion. 131–134. 1 (1). “Determination of weights of interacting criteria from a reference set”.-B. Tzeng G. 640–663. M.-Y. Intenational Journal. European Journal of Operational Research. Saridis and B.. “Elucidative fusion systems – an exposition”. 1/2.-P. 1999.

159. [24] Yang. 153 (1). Wang. International Journal of Information Management. M..-A. Gunasekaran. B¨ y¨ k¨ zkan u u o [23] White. Daniel. “Agile supply chain capabilities: Determinants of competitive objectives”. A. E.O. Y. 95–113. A. Heng. “The role of emergent information technologies and systems in enabling supply chain agility”. and Leung K.... and Mohdzain.-S. 2. 396–410.Y. Adeleye. 5. E. (2005). 2005... P. and Sivayoganathan. [25] Yusuf. Fuzzy Sets and Systems... 1. R.386 G. European Journal of Operational Research.M. 25. 2004. “Fuzzy numbers and fuzzification of the Choquet integral”. K. Z. . 379–392.

208 analytic hierarchy process. 20 Combination rule of rank vectors. 375 evidence theory. DOI 10. 373 ABM. 126 DiNI. 301 Elucidation. Computational Intelligence in Complex Decision Systems. 209 Choquet Integral. 378 knowledge management. 1. 3 fuzzy logic. 68 Fuzzy logic. 18 Haptic devices. 275 Ant colony optimization. 138 fusion system. 12 fuzzy-number-ranking. 301 Evolutionary programming. 19 expert systems. 17 decision making. 274 correlations between criteria. 206 intelligence. 95 Hybrid Systems. 138 Bayesian models. 205 belief decision matrices. 14 evolutionary computation. 301 genetic. 206 Genetic algorithms. 205 electric power market. 374 agile supply chain strategy. 373 fuzzy linear programming. 95 behavior. 205 backward reasoning. 185 Decision Support Systems. 67. 200 D. 125 decision matrix. 153 complex decision systems. 85 interaction between criteria. 148 aggregation operator.1007/978-94-91216-29-9. 19 Evolutionary Computation.Subject Index 2-additive Choquet integral. 67 forward reasoning. 377 fuzzy integral. 183 Evolution strategies. 206 absolute elucidation. 273 fuzzy multiple criteria decision making. Atlantis Computational Intelligence Systems 2. 205 intelligent decision making. 378 agent-based modelling. 373 Classifier systems. 275 Fuzzy Set Theory. 89 hidden Markov models. 128 Fuzzy Multiple Criteria Decision Making. 209 Crossover. 375 agile supply chain. 184 India. 21 Artificial intelligence. 5 Atoms for Peace. 183 benefit. 29 computational intelligence. 3 Artificial neural networks. © 2010 Atlantis Press/World Scientific 387 . 24 impact assessment. Ruan. 373 Altruism. 95 exponential possibility distributions. 205 Discerning Nuclear Intentions. 147 cost. 14 Genetic programming. 185 Evidential Reasoning.

167 sensitivity analysis. 23 Pearson’s correlation. 155 MCDM Procedure with cardinal or mixed data. network-based. 277 Multiobjective Methods. 374 Prolog. 374 multi-objective programming. 205 Neural Networks. 186 Simple Additive Weighting Method. 374 Multi-Attribute Decision Making. 88 Supervised learning. 149 universe of discourse. 160 performance assessment. 11 relative elucidation. 95 Ordinal MCDM procedure. 150 rank-reversal immunity. 68 process aligned. 9 symmetrical triangular possibility distributions. 166 Unsupervised learning. 205 stereoscopic visualization. 129 many-valued temporal logic. 95 multi criteria analysis. 277 simulation. 138 market (customer) sensitive. 278 Multiple Criteria Decision Making. 276 Pareto-optimal overall ranking. 277 training. 92 uncertainty. 205 online evaluation. 184 performance distributions. 277 multi-criteria decision analysis. 9 linear programming. 129 many-valued logic. 278 Virtual reality. 147. 273 possibilistic linear programming. 125 TOPSIS. 125 uncertainty modelling. 130 Technique for Order Performance by Similarity to Ideal Solution. 85 smart home. 200. 11 upper and lower possibility distributions. 153 Outranking Methods. 183 Weighted Product Method. 85 VIKOR. 301 Łukasiewicz logic. 275 multiple-criteria ranking. 169 mixture models. 273 policy simulator. 147 Mutation. 278 . 149 Rank vector. 67 USA. 144 Rank Matrix. 193 Unit n-vector. 301 Multiattribute Utility Theory. 125 Temporality. 200 Policy. 145 Learning processes. 147 Reinforcement learning. 126 Many-valued Temporal Reasoning. 67 load dispatch. 148 maximum number of criteria.388 Computational Intelligence in Complex Decision Systems large samples. 147 lattice valued logic. 5 nuclear proliferation. 205 policy-making process. 277 temporal logic. 85 training systems. 85 virtual. 151 Particle swarm optimization. 374 Marketing Research. 67 Syntax and Semantics. 205 policy making. 206 user evaluation. 128 temporal reasoning. 374 web based assessment. 383 scoring scale. 183 Multi-criteria decision-making (MCDM). 17 NetLogo. 125 socio-biological.