tls ; — 8 s <= = Randy Goebe bot es Ee eeComputational Intelligence A Logical Approach David Poole Alan Mackworth Randy Goebel New York Oxford Oxford University Press 1998Oxford University Press Oxford New York Athens Auckland Bangkok Bogota Bombay Buenos Aires Calcutta Cape Town Dares Salaam Delhi Florence Hong Kong Istanbul Karachi KualaLumpur Madras Madrid Melbourne Mexico City Nairobi Paris Singapore Taipei Tokyo Toronto Warsaw and associated companies in Berlin Ibadan Copyright © 1998 by Oxford University Press, Inc. Published by Oxford University Press, Inc., 198 Madison Avenue, New York, New York, 10016 http://www oup-usa.org Oxford is a registered trademark of Oxford University Press All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, without the prior permission of Oxford University Press. Library of Congress Cataloging-in-Publication Data Poole, David (David Lynton), 1958- Computational intelligence: a logical approach / David Poole, Alan Mackworth, Randy Goebel. p. cm, Includes bibliographical references and index. ISBN 0-19-510270-3 (cloth) 1. Artificial Intelligence, 1. Mackworth, Alan K. II. Goebel, Randy G. II, Title Q335.P657 1997 006.3—de21 97-9075 clip The authors and publisher of this book have used their best efforts in preparing this book. These efforts include the development, research, and testing of the theories and programs to determine their effectiveness. The authors and publisher make no warranty of any kind, expressed or implied, with regard to these programs or the documentation contained in this book. The author and publisher shall not be liable in any event for incidental or consequential damages in connection with, or arising out of, the furnishing, performance, or use of these programs. 987654321 Printed in the United States of America on acid-free paperTo our families for their love, support, and patience Jennifer and Alexandra Marian and Bryn Lanie, Kari, Jodi, and KellyContents Preface xv 1 Computational Intelligence and Knowledge 1 Ll What Is Computational Intelligence? ................ 1 1.2 Agents inthe World... . 2.2... 2.2.20. 0 00s eee 7 1.3 Representation and Reasoning ...........--...0-05 9 14 Applications. 2... ee 11 15 Overview 2... ee ee ee ee 19 1.6 References and Further Reading ................... 20 17 ISI Gol see eee ees ee ooseo 000000000500 21 2 A Representation and Reasoning System 23 2.1 Introduction. ©... eee eee ee 2B 2.2 Representation and Reasoning Systems ............... 23 2.3 Simplifying Assumptions of the Initial RRS .. . . 27 24 Datalog 29 2.5 Semantics... .... 31 2.6 Questions and Answers 40 2.7 Proofs 46 2.8 Extending the Language with Function Symbols. .......... 58 2.9 References and Further Reading ...............00.4. 63 2.10 Exercises 2... ee ee 63 3 Using Definite Knowledge 69 3.1 Introduction... 6... ee eee 69 3.2 Case Study: House Wiring... 2.2... 2 ee eee 70 3.3 Databases and Recursion... . 2.2... 20 eee eee eee 75 3.4 Verification and Limitations... 2... ......-0-2---- 79 3.5 Case Study: Representing Abstract Concepts ............- 81 3.6 Case Study: Representing Regulatory Knowledge. ......... 86 3.7 Applications in Natural Language Processing ............ 91 3.8 References and Further Reading ...............-... 104 viiviii CONTENTS 3.9 Exercises 2... ee ee ee ees 104 Searching 113 4.1 Why Search? 0... ee 113 4.2 Graph Searching .... . ce ee ee ee ee 114 43 A Generic Searching Algorithm Se eee ee eee ee TID 44 Blind Search Strategies ....... . 125 4.5 HeuristicSearch .......... 132 46 Refinements to Search Strategies . . - 138 47 Constraint Satisfaction Problems... ................ 147 4.8 References and Further Reading ...............-.-. 163 49 EXetciscs ei i ie i i 163 Representing Knowledge 169 5.1 Introduction... 6. eee 169 5.2 DefiningaSolution.. 2... 2.2... 000.000.0000 0000e 170 5.3 Choosing a Representation Language .............-.. 174 5.4 Mapping from Problem to Representation .............. 180 5.5 Choosing an Inference Procedure... 2... 2... 2. ee ae 192 5.6 References and Further Reading ................... 195 5.7 EXerciscs Brie i ee 196 Knowledge Engineering 199 6.1 Introduction... ee 199 6.2 Knowledge-Based System Architecture ...........-... 200 6.3 Meta-Interpreters 2... eee 202 6.4 Querying the User 2... 2... ee ee ee 212 6.5 Explanation... . . be ee ee ee 217 6.6 Debugging Knowledge Bases . 221 6.7 A Meta-Interpreter with Search. . . . 226 6.8 Unification Dood 230 6.9 References and Further Reading ................-4. 233 6.10 Exercises 2.2... ee 233 Beyond Definite Knowledge 7A Introduction... 6. ee 7.2 Equality... 2... ee 73 Integrity Constraints 2... 2. ee ee 74 Complete Knowledge Assumption . 75 Disjunctive Knowledge . 76 Explicit Quantification ....... 77 First-Order Predicate Calculus... 1... 2... ee eee eeeCONTENTS 10 ll ix 7.8 Modal Logic .... 2... ..0..200 02000000020 000. 275 71.9 References and Further Reading ................... 277 [/>] OMe EXelCiScs ae 278 Actions and Planning 281 8.1 Introduction. 6.6... ee 281 8.2 Representations of Actions and Change .............. . 287 8.3 Reasoning with World Representations 298 8.4 References and Further Reading ................0.0. 315 8.5 [Exeiciscs me i i ee eee | © Assumption-Based Reasoning 319 9.1 Introduction... 2... ee 319 9.2 An Assumption-Based Reasoning Framework ............- 321 9.3 Default Reasoning ... 2.2... ee 323 9.4 Abduction... 2... ee ee 332 9.5 Evidential and Causal Reasoning. ...............0.0.-. 335 9.6 Algorithms for Assumption-Based Reasoning ........... . 339 97 References and Further Reading ................... 342 9.8 IEXercises aie ei 343 Using Uncertain Knowledge 345 10.1. Introduction... 1... eee 345 10.2. Probability ©... ee 346 10.3. Independence Assumptions... ..............-0005 361 10.4 Making Decisions Under Uncertainty ................ 381 10.5 References and Further Reading ................... 394 10.6 Exercises 2... 2. ee 395 Learning 397 11.1 Introduction... 2... eee 397 11.2 Learning as Choosing the Best Representation. ........... 403 11.3. Case-Based Reasoning .................0.20000. 414 11.4 Learning as Refining the Hypothesis Space ............. 416 11.5 Learning Under Uncertainty ..................0.4. 424 11.6 Explanation-Based Leaning ..................04. 433 11.7 References and Further Reading ................00. 437 11.8 Exercises 2... ee ee 43812 Building Situated Robots CONTENTS 12.1 Introduction. ©... ee 12.2 Robotic Systems 12.3. The Agent Function. . . 12.4 Designing Robots... ......... 0.0.0... 0 2000004 12.5. Usesof Agent Models .......... 0.0.0.0 00000 449 12.6 Robot Architectures... 2.2.2... 2 ee ee ee 450 12.7. Implementing aController .. 2... 2.0... 000.002000.5 451 12.8 Robots Modeling the World ..................0.. 12.9 Reasoning in Situated Robots... ...............00. 12.10 References and Further Reading . . . 2a Exctciscs ara arena ae ee Appendix A Glossary 461 Appendix B The Prolog nn naan 477 B.1 Introduction. ..... a 2) B.2 Interacting with Prolog... ............ 478 B.3 Syntax 479 B4 Arithmetic. © 0. ee ee ee es 481 B.S Database Relations .... 2... 2... ee eee 483 B.6 Returning All Answers . 2... 2... ee ee 485 B.7 Inputand Output... 2.2... ee eee 487 B.8& Controlling Search... 2... ee en 488 Appendix C Some More Implemented Systems 491 C.l Bottom-Up Interpreters... 2.2... 2 ee eee 491 C.2 Top-Down Interpreters . 498 C3 Constraint Satisfaction Problem ‘Solver 507 C4 Neural Network Learner . 2... 2. 2 ee eee Si C5 Partial-Order Planner... 2.1... ee ee ee ee SIS) C6 Implementing Belief Networks .................... 521 (ox) Robot Controller . 2... 0. ee ee 529 Bibliography 533 Index 549List of Figures Ll 1.2 13 2.1 2.2 2.3 2.4 2.5 2.6 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 4.1 4.2 43 4.4 45 4.6 47 48 49 4.10 An agent asa black box... 2.6... ee eee eee 8 Anenvironment for the delivery robot ...............-. 14 An electrical environment for the diagnostic assistant... . 2... 16 The role of semantics in a representation and reasoning system... 26 Truth table defining Aand<— ... 2.2... 2 ee ee een 35 Clauses provided by the user... 2.2... 2. eee ee 44 Bottom-up proof procedure for computing consequences of KB. . . 47 Top-down definite clause interpreter, without variables ....... Sl Top-down definite clause interpreter, with variables ......... 56 An electrical environment with individuals named... ... 2... 71 Sample database... 2... ee 76 Example database of student record information. .......... 87 Example database relations about the university ........... 88 A context-free grammar for a very restricted subset of English . . . 96 A simple dictionary... 2... ee ee 97 Grammar for output of canned English... ............-. 99 A grammar that enforces number agreement and builds a parse tree . 100 A grammar that constructs a query ...............00- 102 A dictionary for constructing aquery................. 103 The delivery robot domain with interesting locations labeled . . . . 115 A graph to search for the delivery robotdomain ........... 117 A search graph for an SLD derivation ...............-. 118 Problem solving by graph searching ...............-. 120 A graph, with cycles, for the delivery robot domain ......... 127 Summary of search strategies... 2... 1 ee ee eee 138 Iterative deepening searcher... 2... 2.2.0.0... 0.000. 142 Arc consistency algorithm AC-3 ................000. 153 Domain-consistent constraint network for the scheduling problem . 154 A foothill, plateau, and ridge on acontourmap ........... 159 xixii 4.11 4.12 4.13 5.1 5.2 5.3 5.4 5.5 5.6 6.1 6.2 6.3 6.4 6.5 6.6 6.7 6.8 6.9 6.10 6.11 6.12 6.13 6.14 7A 7.2 7.3 74 15 7.6 77 78 719 7.10 TAA 7A2 8.1 8.2 8.3 FIGURES Simulated annealing algorithm .................... 160 A grid searching problem... 2.0.2... .0. 00000000005 165 A crossword puzzle to be solved with six words ........... 167 The knowledge representation framework ...... . ... 170 Solution quality as a function of time for an anytime algorithm ... 174 Lattice of sublogics .. 2. 2... ee ee ee ee ee. 176 A hierarchy of representations .. . . .. 178 A flat semantic network - 186 A structured semantic network .. 2... 0.2... eee 189 Knowledge-based system architecture . . . . ee eo00) The non-ground representation for the base language ee 204 The vanilla definite clause meta-interpreter ©... .....0000. 205 A base-level knowledge base for house wiring. ........... 206 A meta-interpreter that uses built-in calls and disjunction... ... 208 A meta-interpreter for depth-bounded search A meta-interpreter that collects delayed goals . . . An ask-the-user interpreter in pseudo-code. . . . . A meta-interpreter that builds a proof tree A meta-interpreter that collects ancestor rules for WHY questions. . 221 An algorithm for debugging incorrect answers. ........... 223 An algorithm for debugging missing answers ............ 225 Meta-interpreter with search .. 2.2... .......... 000. 228 Pseudocode for unification using the ground representation . ... . 232 eT WOlChail Shae 236 Bottom-up proof procedure for computing conflictsof T ...... 246 A meta-interpreter to find conflicts... .. 0.0.0.2... 00048 248 Bottom-up negation as failure proof procedure. . . . 254 Truth table for the clausal operators . 259 Bottom-up proof procedure for computing prime 1e implicates of KB . 261 Meta-interpreter for general clauses ................. 263 A proof tree for Example 7.32... ................. 265 Theorem prover with answer extraction ............... 267 Truth table for predicate calculus operators .............272 Modal logic, their constraints on R, and their axioms ....... . 276 AXiOmS}(Oly aie i ee ee 277 The delivery robot domain with objects... . ee. 285 A simple STRIPS planner that uses STRIPS representation eee 302 Finding weakest preconditions using the STRIPS representation . . 306FIGURES 8.4 8.5 8.6 8.7 9.1 9.2 9.3 10.1 10.2 10.3 10.4 10.5 10.6 10.7 10.8 10.9 11.1 11.2 11.3 11.4 11.5 11.6 11.7 11.8 11.9 11.10 11.11 11.12 12.1 12.2 12.3 12.4 Bl Cl xiii A regression planner 307 Partial-order planner . . 311 Threat handler for the partial-order planner... 2... 2... 313 Partial elaboration of the partial-ordered planning example .. . . . 314 Diagrammatic representation of the defaults from Example 9.3 . . . 325 A diagrammatic representation of Example9.5 ........... 328 Causal network for Example 9.8 2.2.2... 0. ee ee 336 Belief network for the electrical domain of Figure 3.1........ 365 Belief network for report of leaving of Example 10.16 ....... 370 Belief network for aching limbs of Example 10.17 ......... 374 Decision tree for delivery robot decision of Example 10.20... . . 383 Decision network for delivery robot decision 387 Decision network for the alarm problem... .. . 388 Value for alarm decision network... 2.2... 2.2.2... + 388 Possible money-utility tradeoff from Example 10.29 ...... . . 394 Belief network for overhead projector ................ 396 The learning architecture... 2.2... . eee ee ee 398 Some classification data for learning user preferences... ... . . 400 Simple decision tree... 2... 2... ee ee ee 404 Simple decision tree learning program for binary attributes... . . 406 The sigmoid or logistic function ........0......2.20. 410 Simple neural network with one hidden layer 411 Simulation of the neural net learning algorithm ........... 413 Candidate elimination algorithm ................... 419 Posterior distributions of probA based on different samples . . . . . 425 Belief network corresponding to a naive Bayesian classifier... . . 430 A meta-interpreter for explanation-based leaning .......... 434 Example background KB for explanation-based learning .... . . 436 A robotic system and its components................. 445 A hierarchical robotic system architecture... ........0.. 451 A hierarchical decomposition of the delivery robot ......... 454 A simulation of the robot carrying out the plan of Example 12.4. . 457 Syntactic translation between this book and standard Prolog . . . . 479 Forward-chaining assumption-based reasoner ........... . 495Preface Computational Intelligence: A Logical Approach is a book about a new science— computational intelligence. More commonly referred to as artificial intelligence, computational intelligence is the study of the design of intelligent agents. The book is structured as a textbook, but it is accessible to a wide audience. We wrote this book because we are excited about the emergence of computational intelligence as an integrated science. As with any science worth its salt, computational intelligence has a coherent, formal theory and a rambunctious experimental wing. We balance theory and experiment, and show how to link them intimately together. We develop the science of CI together with its engineering applications. We believe the adage, “There is nothing so practical as a good theory.” The spirit of our approach is captured by the dictum, “Everything should be made as simple as possible, but not simpler.” The book works as an introductory text on artificial intelligence for advanced undergraduate or graduate students in computer science, or related disciplines such as computer engineering, philosophy, cognitive science, or psychology. It will ap- peal more to the technically-minded; parts are technically challenging, focusing on learning by doing: designing, building, and implementing systems. Any curious scientifically-oriented reader will profit from studying the book. Previous experience with computational systems is desirable, but prior study of logic is not necessary, as we develop the concepts as required. The serious student will gain valuable skills at several levels ranging from expertise in the specification and design of intelligent agents to skills for implementing, testing, and improving real software systems for several challenging application domains. The thrill of participating in the emergence of a new science of intelligent agents is one of the attractions of this approach. The practical skills of dealing with a world of ubiquitous, intelligent, embedded agents are now in great demand in the marketplace. The focus is on an intelligent agent acting in an environment. We start with simple agents acting in simple, static environments and gradually increase the power of the agents to cope with more challenging worlds. We make this concrete by repeatedly illustrating the ideas with three different agent tasks: a delivery robot, a diagnostic assistant, and an information slave (the infobot). Our theory is based on logic. Logic has been developed over the centuries as a for- xvXvi PREFACE mal (that is, precise not obtuse) way of representing assumptions about a world and the process of deriving the consequences of those assumptions. For simple agents in simple worlds we start with a highly restricted simple logic. Then as our agent/environment requires, we increase the logical power of the formalism. Since a computer is simply a symbol-manipulation engine, we can easily map our formal theories into computer programs that can control an agent or be used to reason about an agent. Everything we describe is implemented that way. Any high-level symbol-manipulation language, such as LISP or Prolog, can be used to support these implementations. We also provide complete implementations in Prolog, a language that is designed around the idea of computations as proof in a simple logic. But this connection is not essential to an understanding or use of the ideas in this book. Our approach, through the development of the power of the agent’s representation language, is both simpler and more powerful than the usual approach of surveying and cataloging various applications of artificial intelligence. However, as a consequence, some applications such as the details of computational vision or the pragmatics of low-level robot control are not covered in this book. We do show how our theory allows us to design situated robots performing in dynamically changing worlds. We invite you to join us in an intellectual adventure: building a science of intelligent agents. David Poole Alan Mackworth Randy Goebel Acknowledgments We are very grateful to Valerie McRae for contributing her proofreading and for- matting skills and for helping in many other ways. We also gratefully acknowledge the helpful comments on earlier drafts of this book received from Carl Alphonce, Brock Atherton, Veronica Becher, Wolfgang Bibel, Craig Boutilier, Jason Burrell, Hugh Clapin, Keith Clark, Jim Delgrande, Koichi Furukawa, Chris Geib, Holger Hoos, Michael Horsch, Ying Li, Greg McClement, Claude Sammut, Julie Schlifske, Daniele Theseider Dupré, Paul van Arragon, Bonnie Webber, Bob Woodham, and Ying Zhang. Thanks to our editors at OUP, Bill Zobrist, Karen Shapiro, and Krysia Bebick. The four-limbed robot shown on the cover is a Platonic Beast, developed by Dinesh Pai, Rod Barman, and Scott Ralph at the University of British Columbia’s Laboratory for Computational Intelligence. Based on the spherical symmetry of the tetrahedron, one of the five symmetrical Platonic polyhedra, the Beast can walk, roll over, and climb over obstacles. Aidan Rowe and John Chong provided the technical graphical expertise to complete the cover design.Chapter 1 Computational Intelligence and Knowledge 11 What Is Computational Intelligence? Computational intelligence is the study of the design of intelligent agents. An agent is something that acts in an environment—it does something. Agents include worms, dogs, thermostats, airplanes, humans, organizations, and society. An intel- ligent agent is a system that acts intelligently: What it does is appropriate for its circumstances and its goal, it is flexible to changing environments and changing goals, it learns from experience, and it makes appropriate choices given perceptual limitations and finite computation. The central scientific goal of computational intelligence is to understand the prin- ciples that make intelligent behavior possible, in natural or artificial systems. The main hypothesis is that reasoning is computation. The central engineering goal is to specify methods for the design of useful, intelligent artifacts. Artificial or Computational Intelligence? Artificial intelligence (AI) is the established name for the field we have defined as computational intelligence (CI), but the term “artificial intelligence” is a source of much confusion. Is artificial intelligence real intelligence? Perhaps not, just as an artificial pearl is a fake pearl, not a real pearl. “Synthetic intelligence” might be a better name, since, after all, a synthetic pearl may not be a natural pear! but it is a real pearl. However, since we claimed that the central scientific goal is to understand both naturalCHAPTER |. COMPUTATIONAL INTELLIGENCE AND KNOWLEDGE and artificial (or synthetic) systems, we prefer the name “computational intelligence.” It also has the advantage of making the computational hypothesis explicit in the name. The confusion about the field’s name can, in part, be attributed to a confounding of the field’s purpose with its methodology. The purpose is to understand how intelli- gent behavior is possible. The methodology is to design, build, and experiment with computational systems that perform tasks commonly viewed as intelligent. Building these artifacts is an essential activity since computational intelligence is, after all, an empirical science; but it shouldn’t be confused with the scientific purpose. Another reason for eschewing the adjective “artificial” is that it connotes simulated intelligence. Contrary to another common misunderstanding, the goal is not to simulate intelligence. The goal is to understand real (natural or synthetic) intelligent systems by synthesizing them. A simulation of an earthquake isn’t an earthquake; however, we want to actually create intelligence, as you could imagine creating an earthquake. The misunderstanding comes about because most simulations are now carried out on computers. However, you shall see that the digital computer, the archetype of an interpreted automatic, formal, symbol-manipulation system, is a tool unlike any other: It can produce the real thing. The obvious intelligent agent is the human being. Many of us feel that dogs are intelligent, but we wouldn’t say that worms, insects, or bacteria are intelligent (Exercise 1.1). There is a class of intelligent agents that may be more intelligent than humans, and that is the class of organizations. Ant colonies are the prototypical example of organizations. Each individual ant may not be very intelligent, but an ant colony can act more intelligently than any individual ant. The colony can discover food and exploit it very effectively as well as adapt to changing circumstances. Similarly, companies can develop, manufacture, and distribute products where the sum of the skills required is much more than any individual could understand. Modern computers, from the low- level hardware to high-level software, are more complicated than can be understood by any human, yet they are manufactured daily by organizations of humans. Human society viewed as an agent is probably the most intelligent agent known. We take inspiration from both biological and organizational examples of intelligence. Flying Machines and Thinking Machines It is instructive to consider an analogy between the development of flying machines over the last few centuries and the development of thinking machines over the last few decades. First note that there are several ways to understand flying. One is to dissect known flying animals and hypothesize their common structural features as necessary fundamental characteristics of any flying agent. With this method an examination of birds, bats, and insects would suggest that flying involves the flapping of wings made of some structure covered with feathers or a membrane. Furthermore, the hypothesis1.1. WHAT IS COMPUTATIONAL INTELLIGENCE? 3 could be verified by strapping feathers to one’s arms, flapping, and jumping into the air, as Icarus did. You might even imagine that some enterprising researchers would claim that one need only add enough appropriately layered feather structure to achieve the desired flying competence, or that improved performance required more detailed modeling of birds such as adding a cloaca. An alternate methodology is to try to understand the principles of flying without restricting ourselves to the natural occurrences of flying. This typically involves the construction of artifacts that embody the hypothesized principles, even if they do not behave like flying animals in any way except flying. This second method has provided both useful tools, airplanes, and a better understanding of the principles underlying flying, namely aerodynamics. It is this difference which distinguishes computational intelligence from other cog- nitive science disciplines. CI researchers are interested in testing general hypotheses about the nature of intelligence by building machines which are intelligent and which don’t simply mimic humans or organizations. This also offers an approach to the question “Can computers really think?” by considering the analogous question “Can airplanes really fly?” Technological Models of Mind Throughout human history, people have used technology to model themselves. Consider this Taoist parable taken from the book Lieh Tzu, attributed to Lieh Yu- Khou: “Who is that man accompanying you?” asked the king. “That, Sir,” replied Yen Shih, “is my own handiwork. He can sing and he can act.” The king stared at the figure in astonishment. It walked with rapid strides, moving its head up and down, so that anyone would have taken it for a live human being. The artificer touched its chin, and it began singing, perfectly in tune. He touched its hand and it began posturing, keeping perfect time .... The king, looking on with his favorite concubine and other beauties, could hardly persuade himself that it was not real. As the performance was drawing to an end, the robot winked its eye and made advances to the ladies in attendance, whereupon the king became incensed and would have had Yen Shih executed on the spot had not the latter, in mortal fear, instantly taken the robot to pieces to let him see what it really was. And, indeed, it turned out to be only a construction of leather, wood, glue and lacquer, variously colored white, black, red and blue. Examining it closely, the king found all the internal organs complete— liver, gall, heart, lungs, spleen, kidneys, stomach and intestines; and over these again, muscles, bones and limbs with their joints, skin, teeth andCHAPTER |. COMPUTATIONAL INTELLIGENCE AND KNOWLEDGE hair, all of them artificial. Not a part but was fashioned with the utmost nicety and skill; and when it was put together again, the figure presented the same appearance as when first brought in. The king tried the effect of taking away the heart, and found that the mouth could no longer speak; he took away the liver and the eyes could no longer see; he took away the kidneys and the legs lost their power of locomotion. The king was delighted. This story, dating from about the third century B.C., is one of the earliest written accounts of building intelligent agents, but the temples of early Egypt and Greece also bear witness to the universality of this activity. Each new technology has been exploited to build intelligent agents or models of mind. Clockwork, hydraulics, tele- phone switching systems, holograms, analog computers, and digital computers have all been proposed both as technological metaphors for intelligence and as mechanisms for modeling mind. Parenthetically, we speculate that one reason for the king’s delight was that he re- alized that functional equivalence doesn’t necessarily entail structural equivalence. In order to produce the functionality of intelligent behavior it isn’t necessary to reproduce the structural connections of the human body. This raises the obvious question of whether the digital computer is just another technological metaphor, perhaps a fad soon to be superseded by yet another mecha- nism. In part, the answer must be empirical. We need to wait to see if we can get substantial results from this approach, but also to pursue alternate models to determine if they are more successful. We have reason to believe the answer to that question is “no.” Some reasons are empirical: The results to date are impressive but not, of course, conclusive. There are other reasons. Consider the following two hypotheses. The first is called the symbol-system hypothesis: Reasoning is symbol manipulation. The second hypothesis is called the Church-Turing thesis: Any symbol manipulation can be carried out on a Turing machine. A Turing machine is an idealization of a digital computer with an unbounded amount of memory. These hypotheses imply that any symbol manipulation, and so any reasoning, can be carried out on a large enough deterministic computer. There is no way you can prove these two hypothesis mathematically. All you can do is empirically test then by building reasoning systems. Why should you believe that they are true or even reasonable? The reason is that language, which provides one of the few windows to the mind, is inherently about transmission of symbols. Reasoning in terms of language has symbols as inputs and outputs, and so the function from inputs to outputs can be described symbolically, and presumably can be implemented in terms1.1. WHAT IS COMPUTATIONAL INTELLIGENCE? 5 of symbol manipulation. Also the intelligence that is manifest in an organization or in society is transmitted by language and other signals. Once you have expressed something in a language, reasoning about it is symbol manipulation. These hypotheses don’t tell us how to implement arbitrary reasoning on a computer—this is CI’s task. What it does tell us is that computation is an appropriate metaphor for reasoning. This hypothesis doesn’t imply that every detail of computation can be interpreted symbolically. Nor does it imply that every machine instruction in a computer or the function of every neuron in a brain can be interpreted symbolically. What it does mean is that there is a level of abstraction in which you can interpret reasoning as symbol manipulation, and that this level can explain an agent’s actions in terms of its inputs. Before you accept this hypothesis, it is important to consider how it may be wrong. An alternative is that action is some continuous function of the inputs to an agent such that the intermediate values don’t necessarily correspond to anything meaningful. It is even possible that the functionality can’t be interpreted symbolically, without resorting to using meaningless numbers. Alternative approaches are being pursued in both neural networks (page 408) and in building reactive robots (page 443) inspired by artificial insects, Science and Engineering As suggested by the flying analogy, there is tension between the science of CI, trying to understand the principles behind reasoning, and the engineering of CI, build- ing programs to solve particular problems. This tension is an essential part of the discipline. As Cl is a science, its literature should manifest the scientific method, especially the creation and testing of refutable theories. Obvious questions are, “What are CI theories about?” and “How would I test one if I had one?” CI theories are about how interesting problems can be represented and solved by machine. Theories are sup- ported empirically by constructing implementations, part of whose quality is judged by traditional computer science principles. You can’t accomplish CI without specify- ing theories and building implementations; they are inextricably connected. Of course, not every researcher needs to do both, but both must be done. An experiment means nothing without a theory against which to evaluate it, and a theory without potentially confirming or refuting evidence is of little use. Ockham’s Razor is our guide: Always prefer simple theories and implementations over the more complex. With these thoughts in mind, you can quickly consider one of the most often considered questions that arises in the context of CI: “Is human behavior algorithmic?” You can dispense with this question and get on with your task by acknowledging that the answer to this question is unknown, it is part of cognitive science and CI’s goal to find out.CHAPTER |. COMPUTATIONAL INTELLIGENCE AND KNOWLEDGE Relationship to Other Disciplines Cl is a very young discipline. Other disciplines as diverse as philosophy, neu- robiology, evolutionary biology, psychology, economics, political science, sociology, anthropology, control engineering, and many more have been studying intelligence much longer. We first discuss the relationship with philosophy, psychology, and other disciplines which study intelligence; then we discuss the relationship with computer science, which studies how to compute. The science of of CI could be described as “synthetic psychology,” “experimental philosophy,” or “‘computational epistemology”—Epistemology is the study of knowl- edge. It can be seen as a way to study the old problem of the nature of knowledge and intelligence, but with a more powerful experimental tool than was previously available. Instead of being able to observe only the external behavior of intelligent systems, as philosophy, psychology, economics, and sociology have traditionally been able to do, we are able to experiment with executable models of intelligent behavior. Most im- portantly, such models are open to inspection, redesign, and experiment in a complete and rigorous way. In other words, you now have a way to construct the models that philosophers could only theorize about. You can experiment with these models, as opposed to just discussing their abstract properties. Our theories can be empirically grounded in implementation. Just as the goal of aerodynamics isn’t to synthesize birds, but to understand the phenomenon of flying by building flying machines, CI’s ultimate goal isn’t necessarily the full-scale simulation of human intelligence. The notion of psychological validity separates CI work into two categories: that which is concerned with mimicking human intelligence—often called cognitive modeling—and that which isn’t. To emphasize the development of Cl as a science of intelligence, we are concerned, in this book at least, not with psychological validity but with the more practical desire to create programs that solve real problems. Sometimes it will be important to have the computer to reason through a problem in a human-like fashion. This is especially important when a human requires an explanation of how the computer generated an answer. Some aspects of human cognition you usually do not want to duplicate, such as the human’s poor arithmetic skills and propensity for error. Computational intelligence is intimately linked with the discipline of computer science. While there are many non-computer scientists who are researching CI, much, if not most, CI (or AI) research is done within computer science departments. We believe this is appropriate, as the study of computation is central to CI. It is essential to understand algorithms, data structures, and combinatorial complexity in order to build intelligent machines. It is also surprising how much of computer science started as a spin off from AI, from timesharing to computer algebra systems. There are other fields whose goal is to build machines that act intelligently. Two of these fields are control engineering and operations research. These start from different1.2. AGENTS IN THE WORLD 7 points than CI, namely in the use of continuous mathematics. As building real agents involves both continuous control and CI-type reasoning, these disciplines should be seen as symbiotic with CI. A student of either discipline should understand the other. Moreover, the distinction between them is becoming less clear with many new theories combining different areas. Unfortunately there is too much material for this book to cover control engineering and operations research, even though many of the results, such as in search, have been studied in both the operations research and CI areas. Finally, CI can be seen under the umbrella of cognitive science. Cognitive sci- ence links various disciplines that study cognition and reasoning, from psychology to linguistics to anthropology to neuroscience. CI distinguishes itself within cognitive science because it provides tools to build intelligence rather than just studying the external behavior of intelligent agents or dissecting the inner workings of intelligent systems. 1.2 Agents in the World There are many interesting philosophical questions about the nature and substance of CI, but the bottom line is that, in order to understand how intelligent behavior might be algorithmic, you must attempt to program a computer to solve actual problems. It isn’t enough to merely speculate that some particularly interesting behavior is algo- rithmic. You must develop a theory that explains how that behavior can be manifest in a machine, and then you must show the feasibility of that theory by constructing an implementation. We are interested in practical reasoning: reasoning in order to do something. Such a coupling of perception, reasoning, and acting comprises an agent. An agent could be, for example, a coupling of a computational engine with physical actuators and sensors, called a robot. It could be the coupling of an advice-giving computer—an expert system—with a human who provides the perceptual informa- tion and who carries out the task. An agent could be a program that acts in a purely computational environment—an infobot. Figure 1.1 shows the inputs and outputs of an agent. At any time the agent has: e Prior knowledge about the world e Past experience that it can learn from e Goals that it must try to achieve or values about what is important ¢ Observations about the current environment and itself and it does some action. For each agent considered, we specify the forms of the inputs and the actions. The goal of this book is to consider what is in the black box so that the action is reasonable given the inputs.CHAPTER |. COMPUTATIONAL INTELLIGENCE AND KNOWLEDGE prior knowledge past experiences actions goals/values observations Figure 1.1: An agent as a black box For our purpose, the world consists of an agent in an environment. The agent’s environment may well include other agents. Each agent has some internal state that can encode beliefs about its environment and itself. It may have goals to achieve, ways to act in the environment to achieve those goals, and various means to modify its beliefs by reasoning, perception. and learning. This is an all-encompassing view of intelligent systems varying in complexity from a simple thermostat to a team of mobile robots to a diagnostic advising system whose perceptions and actions are mediated by human beings. Success in building an intelligent agent naturally depends on the problem that one selects to investigate. Some problems are very well-suited to the use of computers, such as sorting a list of numbers. Others seem not to be, such as changing a baby’s. diaper or devising a good political strategy. We have chosen some problems that are representative of a range of applications of current CI techniques. We seek to demonstrate, by case study, CI’s methodology with the goal that the methodology is transferable to various problems in which you may be interested. We establish a framework that places you, the reader, in a position to evaluate the current CI literature and anticipate the future; and, most importantly, we develop the concepts and tools necessary to allow you to build, test, and modify intelligent agents. Finally we must acknowledge there is still a huge gulf between the dream of computational intelligence and the current technology used in the practice of building what we now call intelligent agents. We believe we have many of the tools necessary to build intelligent agents, but we are certain we don’t have all of them. We could, of course, be on the wrong track; it is this fallibility that makes CI science and makes the challenge of CI exciting.1.3. REPRESENTATION AND REASONING 9 1.3. Representation and Reasoning Experience shows that the performance of tasks that seem to involve intelligence also seem to require a huge store of knowledge. A major thesis of this book is that Cl is the study of knowledge. This raises the question which is part of our subject material, “What is knowledge?” Informally, knowledge is information about some domain or subject area, or about how to do something. Much of our effort will be devoted to formalizing and refining a common-sense notion of knowledge, with the motivation of developing both a theoretical and practical framework for representing and using knowledge. Humans require and use a lot of knowledge to carry out even the most simple common-sense tasks. Computers are very good at tasks which do not require much knowledge, such as simple arithmetic, symbolic differentiation, or sorting. They aren’t, as yet, very good at many knowledge-intensive tasks at which humans excel, such as recognizing faces in a picture, medical diagnosis, understanding natural lan- guage, or legal argumentation. At the heart of this book is the design of computational systems that have knowledge about the world and that can act in the world based on that knowledge. The notion of knowledge is central to this book. The systems we want to develop should be able to acquire and use knowledge to solve the problems at hand. The main issues are how to acquire and represent knowledge about some domain and how to use that knowledge to answer questions and solve problems. You will notice that we make a strong commitment to logic approach in this book. Our commitment is really to a precise specification of meaning rather than to any particular syntax. We have no great commitment to any particular syntax. Many different notations are possible. Sometimes we will write sentences, sometimes we will use diagrams. In order to represent anything, you have to commit to some notation, and the simpler the better. We use Prolog’s syntax, not because we particularly like Prolog or its syntax, but because it is important for scholars of CI to get experience with using logic to solve problems, and Prolog is probably the most accessible system that allows you to do this. Representation and Reasoning System In order to use knowledge and reason with it, you need what we call a representation and reasoning system (RRS). A representation and reasoning system is composed of a language to communicate with a computer, a way to assign meaning to the language, and procedures to compute answers given input in the language. Intuitively, an RRS lets you tell the computer something in a language where you have some meaning associated with the sentences in the language, you can ask the computer questions,10 CHAPTER 1. COMPUTATIONAL INTELLIGENCE AND KNOWLEDGE and the computer will produce answers that you can interpret according to the meaning associated with the language. At one extreme, the language could be a low-level programming language such as Fortran, C++, or Lisp. In these languages the meaning of the sentences, the programs, is purely in terms of the steps the computer will carry out to execute the program. What computation will be carried out given a program and some input, is straightforward to determine. How to map from an informal statement of a problem to a representation of the problem in these RRSs, programming, is a difficult task. At the other extreme, the language could be a natural language, such as English, where the sentences can refer to the problem domain. In this case, the mapping from a problem to a representation is not very difficult: You need to describe the problem in English. However, what computation needs to be carried out in the computer in response to the input is much more difficult to determine. In between these two extremes are the RRSs that we consider in this book. We want RRSs where the distance from a natural specification of the problem to the representation of the problem is not very far. We also want RRSs where the appropriate computation, given some input, can be effectively determined. We consider languages for the specification of problems, the meaning associated with such languages, and what computation is appropriate given input in the languages. One simple example of a representation and reasoning system between these two extremes is a database system. In a database system, you can tell the computer facts about a domain and then ask queries to retrieve these facts. What makes a database system into a representation and reasoning system is the notion of semantics. Seman- tics allows us to debate the truth of information in a knowledge base and makes such information knowledge rather than just data. In most of the RRSs we are interested in, the form of the information is more flexible and the procedures for answering queries are more sophisticated than in a database. A database typically has table lookup; you can ask about what is in the database, not about what else must be true, or is likely to be true, about the domain. Chapter 2 gives a more precise definition an RRS and a particular RRS that is both simple and yet very powerful. It is this RRS that we build upon throughout this book, eventually presenting RRSs that can reason about such things as time, typicality, uncertainty, and action. Ontology and Conceptualization An important and fundamental prerequisite to using an RRS is to decide how a task domain is to be described. This requires us to decide what kinds of things the domain consists of, and how they are to be related in order to express task domain problems. A major impediment to a general theory of Cl is that there is no comprehensive theory of how to appropriately conceive and express task domains. Most of what we know about1.4. APPLICATIONS 11 this is based on experience in developing and refining representations for particular problems. Despite this fundamental problem, we recognize the need for the following com- mitments. e The world can be described in terms of individuals (things) and relationships among individuals. An ontology is acommitment to what exists in any particular task domain. This notion of relationship is meant to include propositions that are true or false independently of any individuals, properties of single individuals, as well as relationships between pairs or more individuals. This assumption that the world can be described in terms of things is the same that is made in logic and natural language. This isn’t a strong assumption, as individuals can be anything nameable, whether concrete or abstract. For example, people, colors, emotions, numbers, and times can all be considered as individuals. What is a “thing” is a property of an observer as much as it is a property of the world. Different observers, or even the same observer with different goals, may divide up the world in different ways. e For each task or domain, you need to identify specific individuals and relations that can be used to express what is true about the world under consideration. How you do so can profoundly affect your ability to solve problems in that domain. For most of this book we assume that the human who is representing a domain decides on the ontology and the relationships. To get human-level computational intelligence it must be the agent itself that decides how to divide up the world, and which relationships to reason about. However, it is important for you to understand what knowledge is required for a task before you can expect to build a computer to learn or introspect about how to solve a problem. For this reason we concentrate on what it takes to solve a problem. It should not be thought that the problem of CI is solved. We have only just begun this endeavor. 1.4 Applications Theories about representation and reasoning are only useful insofar as they provide the tools for the automation of problem solving tasks. CI’s applications are diverse, including medical diagnosis, scheduling factory processes, robots for hazardous envi- ronments, chess playing, autonomous vehicles, natural language translation systems, and cooperative systems. Rather than treating each application separately, we abstract essential features of such applications to allow us to study principles behind intelligent reasoning and action.12 CHAPTER |. COMPUTATIONAL INTELLIGENCE AND KNOWLEDGE This section outlines three application domains that will be developed in examples throughout the book. Although the particular examples presented are simple—for otherwise they wouldn’t fit into the book—the application domains are representative of the sorts of domains in which CI techniques can be, and have been, used. The three application domains are: e An autonomous delivery robot that can roam around a building delivering pack- ages and coffee to people in the building. This delivery agent needs to be able to, for example, find paths, allocate resources, receive requests from people, make decisions about priorities, and deliver packages without injuring people or itself. e A diagnostic assistant that helps a human troubleshoot problems and suggests repairs or treatments to rectify the problems. One example is an electrician’s assistant that can suggest what may be wrong in a house, such as a fuse blown, a light switch broken, or a light burned out given some symptoms of electrical problems. Another example is of a medical diagnostician that finds potential diseases, possible tests, and appropriate treatments based on knowledge of a particular medical domain and a patient’s symptoms and history. This assistant needs to be able to explain its reasoning to the person who is carrying out the tests and repairs, as that person is ultimately responsible for what they do. The diagnostic assistant must add substantial value in order to be worth using. e An “infobot” that can search for information on a computer system for naive users such as company managers or people off the street. In order to do this the infobot must find out, using the user’s natural language, what information is re- quested, determine where to find out the information, and access the information from the appropriate sources. It then must report its findings in an appropriate format so that the human can understand the information found, including what they can infer from the lack of information. These three domains will be used for the motivation for the examples in the book. In the next sections we discuss each application domain in detail. The Autonomous Delivery Robot Imagine a robot that has wheels and can pick up objects and put them down. It has sensing capabilities so that it can recognize the objects it needs to manipulate and so it can avoid obstacles. It can be given orders in natural language and obey them, making common sense assumptions about what to do when its goals conflict. Such a robot could be used in an office environment to deliver packages, mail, or coffee. It needs to be useful as well as safe. In terms of the black box definition of an agent in Figure 1.1, the autonomous delivery robot has as inputs:1.4. APPLICATIONS 13 e Prior knowledge in terms of knowledge about its capabilities, what objects it may encounter and need to differentiate, and perhaps some prior knowledge about its environment, such as maps. e Past experience about, for instance, which actions are useful in which situations, what objects are in the world, how its actions affect its position, and experience about previous requests for it to act. Goals in terms of what it needs to deliver and when, as well as values that specify tradeoffs such as when it must forgo one goal to pursue another or the tradeoff between acting quickly and acting safely. e Observations about its environment from such input devices as cameras, sonar, sound, laser range finders, or keyboards for requests. The robot’s output is motor controls that specify where its wheels should turn, where its limbs should move, and what it should do with it grippers. In order for this robot to be able to function, it has to be able to: e Determine where individuals’ offices are, where to get coffee, how to estimate the length of a trip, and so on. This involves being able to infer information from a database of facts about the domain. How to infer implicit information from a knowledge base is explored in Chapters 2 and 3. e Find a path between different locations. It may want the shortest, the quickest, or the safest path. This involves searching as developed in Chapter 4. e Be able to represent knowledge about the domain so that inference can be quick, so that knowledge can be easily acquired, and so that the appropriate knowledge is represented. Such issues are discussed in Chapter 5. e Plan how to carry out multiple goals, even when they use the same resources, for example, when the robot’s carrying capacity is limited. Planning is discussed in Chapter 8. e@ Make default assumptions—for example, about where people will be or where coffee can be found. See Chapter 9. e Make tradeoffs about plans even though there may be uncertainty about what is in the world and about the outcome of its actions. Such reasoning under uncertainty is discussed in Chapter 10. e Learn about features of its domain, as well as learn about how its actions affect its position and its rewards. See Chapter 11. e Sense the world, know where it is, steer around the corridors (avoiding people and other objects), and pick up and put down objects. See Chapter 12. Figure 1.2 depicts a typical laboratory environment for a delivery robot. This environment consists of four laboratories and many offices arranged in a grid. We assume that the robot can only push doors, and the directions of the doors in theCHAPTER 1. COMPUTATIONAL INTELLIGENCE AND KNOWLEDGE 131 | 129 | 127 | 1125 | 123 | r121 | r119 m7 lab4 lab3 115 lab1 lab2 113 stairs 101 r103 | r105 | r107 | ri09 | 1111 Figure 1.2: An environment for the delivery robot diagram reflect the directions where the robot can travel. We also assume that rooms need keys, and that keys can be obtained from various sources. The robot needs to deliver parcels and letters from room to room. The environment also contains a stairway that can be hazardous to the robot. The Diagnostic Assistant A diagnostic assistant is intended to advise a human about some particular artifact, such as a medical patient, the electrical system in a house, or an automobile, when symptoms are manifest. It should advise about potential underlying faults or diseases, what tests to carry out, and what treatment to prescribe. In order to give such advice the assistant needs to have some model of the system, knowledge of potential causes, available tests, available treatments, and observations about a particular artifact. As-1.4, APPLICATIONS 15 sisting a human involves making sure that the system provides added value, is easy for a human to use, and isn’t more trouble than it is worth. It must be able to justify why the suggested diagnoses or actions are appropriate. Humans are, and should be, suspicious of computer systems that are impenetrable. When humans are responsible for what they do, even if it is based on a computer system’s advice, they need to have reasonable justifications for the suggested actions. In terms of the black box definition of an agent in Figure 1.1, the diagnostic assistant has as inputs: e Prior knowledge such as what’s normal and what’s abnormal about how switches and lights work, how diseases or malfunctions manifest themselves, what infor- mation tests provide, and the side effects of repairs or treatments. e Past experience in terms of data of previous cases that include the effects of repairs or treatments, the prevalence of faults or diseases, the prevalence of symptoms for these faults or diseases, and the accuracy of tests. ¢ Goals of fixing the device and tradeoffs such as between fixing or replacing different components, or whether a patient prefers to live longer if it means they will be less coherent. e Observations of symptoms of a device or patient. The output of the diagnostic assistant is in terms of recommendations of treatments and tests, along with rationales for its recommendations. In order for the diagnostic assistant to be useful it must be able to: e Derive the effects of faults and interventions (Chapter 3). e Search through the space of possible faults or disease complexes (Chapter 4). e Explain its reasoning to the human who is using it (Chapter 6). e Derive possible causes for symptoms; rule out other causes based on the symp- toms (Chapter 7). e Plan courses of tests and treatments to address the problems (Chapter 8). e Hypothesize problems and use default knowledge that may not always be true (Chapter 9). e Reason about the uncertainties about the artifact given only partial information about the state of the artifact, the uncertainty about the effects of the treatments, and the tradeoffs between the alternate courses of action (Chapter 10). e Learn about what symptoms are associated with the faults or diseases, the effects of treatments, and the accuracy of tests (Chapter 11). Figure 1.3 shows a depiction of the electrical distribution in a house. In this house, power comes into the house through circuit breakers, and then it goes to power outlets or to lights through light switches. For example, light /, is on if there is power comingCHAPTER 1. COMPUTATIONAL INTELLIGENCE AND KNOWLEDGE circuit breaker switch re — two-way __A— _ switcn ya light —_| power outlet Figure 1.3: An electrical environment for the diagnostic assistant into the house, if circuit breaker cb; is on, and if switches s; and s2 are either both up or both down. This is the sort of model that a normal householder may have of the electrical power in the house which they could use to determine what is wrong given evidence about the position of the switches and which lights are on and which are off. The diagnostic assistant is there to help the householder or an electrician to troubleshoot electrical problems. The Infobot An infobot is like a robot, but instead of interacting with a physical environment, it interacts with an information environment. Its task is to extract information from a network of diverse information sources such as the Internet or a multimedia ency- clopedia. The infobot must determine what information is needed from a query in a formal language, from a sophisticated user, or from a natural language query from a1.4. APPLICATIONS 17 casual user such as a manager or person off the street. It must determine where the information may be obtained, retrieve the information, and present it in a meaningful way to the user. In terms of the black box definition of an agent in Figure 1.1, the infobot has as inputs: e Prior knowledge about the meaning of words, the types of information sources, and how to access information. e Past experience about where information can be obtained, the relative speed of various servers, and information about the preferences of the user. e Goals in terms of what information it needs to find out and tradeoffs about how much expense should be involved to get the information and the tradeoff between the volume and quality of information. e Observations about what information is at the current sites, what links are avail- able, and the load on various connections. The output of the infobot is information presented so that the user can understand what is there and the significance of missing information. The infobot needs to be able to: e Derive information that is only implicit in the knowledge base(s), as well as interact in natural language (Chapter 3). e Search through a variety of knowledge bases looking for relevant information (Chapter 4). e Find good representations of knowledge so that answers can be computed effi- ciently (Chapter 5). e Explain how an answer was derived or why some information was unavailable (Chapter 6). e Make conclusions about lack of knowledge, determine conflicting knowledge, and be able to conclude disjunctive knowledge (Chapter 7). e Use default reasoning about where to obtain different information (Chapter 9). e@ Make tradeoffs between cheap but unreliable information sources and more expensive but more comprehensive information sources (Chapter 10). e Learn about what knowledge is available where, and what information the user is interested in (Chapter 11). We consider two different infobots: the “unibot” and the “‘webbot.” The unibot in- teracts with a database of information about courses, scheduling, degree requirements, and grades. The webbot interacts with the World Wide Web, finding information that may be of use to a user. One of the most interesting aspects of an infobot is that it ought to be able to volunteer information that users don’t know exists, and so can’t be expected to ask for even though they may be interested.CHAPTER |. COMPUTATIONAL INTELLIGENCE AND KNOWLEDGE Common Features These three examples have common features. At one level of abstraction, they each have four tasks: Modeling the environment The robot needs to be able to model the physical en- vironment, its own capabilities, and the mechanics of delivering parcels. The diagnostic assistant needs to be able to model the general course of diseases or faults, know how normal artifacts work, know how treatments work, and know what information tests provide. The infobot needs to be able to model how information can be obtained, what are legal answers to questions, and what information is actually needed, based on a request. Evidential reasoning or perception This is what control theorists call “system iden- tification” and what doctors call “diagnosis.” Given some observations about the world, the task is to determine what is really in the world. This is most evident in the diagnostic assistant, where the system is given symptoms (observations) and has to determine the underlying faults or diseases. The delivery robot must try to determine where it is and what else is in its environment based on limited sensing information such as touch, sonar, or vision. The infobot has to deter- mine where information is available, given only partial information about the contents of information sources. Action Givena model of the world and a goal, the task is to determine what should be done. For the delivery robot, this means that it must actually do something, such as rove around the corridors and deliver things. For the diagnostic assistant, the actions are treatments and tests. It isn’t enough to theorize about what may be wrong, but a diagnostician must make tests and has to consider what it will do based on the outcome of tests. It isn’t necessary to test if the same treatment will be carried out no matter what the test’s outcome, such as replacing a board on a computer or giving a patient an antibiotic. The actions of the infobot are computational, such as, consulting a knowledge base in order to extract some information. Learning from past experience This includes learning what the particular environ- ment is like, the building the delivery robot is in, the particular patient being diagnosed, or the communication bottlenecks of a computer network; learning general information, how the robot sensors actually work, how well particular diseases respond to particular treatments, or how fast different types of computer connections are; and learning how to solve problems more efficiently. These tasks cut across all application domains. It is essentially the study of these four tasks that we consider in this book. These four tasks interact. It is difficult to study one without the others. We have decided that the most sensible organization is to build1.5. OVERVIEW 19 the tools needed from the bottom up and to show how the tools can be used for each task and, through these tasks, demonstrate the limitations of the tools. We believe this organization will help in understanding the commonalities across different domains and in understanding the interaction among the different tasks. 1.5 Overview Our quest for a unified view of CI is based on the fundamental nature of the concepts of representation and reasoning. We seek to present these techniques as an evolution of ideas used to solve progressively more difficult problems. Chapter 2 starts with a simple representation and reasoning system, where we assume that the agents have complete knowledge about the world and that the world never changes. Subsequent chapters discuss the removal of such constraints in terms of their effect on representation and reasoning. In Chapter 3, we give specific examples of using a definite knowledge encoding for various useful applications. In Chapter 4, we show how many reasoning problems can be understood as search problems. We review some standard approaches to search-based problem solving, including various kinds of informed and uninformed search. Chapter 5 discusses knowledge represen- tation issues and explains how they are manifest in the ideas developed in the first three chapters. Chapter 6 provides further details about knowledge-based systems and presents an overview of a system architecture of a typical expert system, including tools to enable an expert system to explain its reasoning. Chapter 7 removes the as- sumptions about definite knowledge, by allowing disjunctive and negative knowledge, culminating in full first-order predicate calculus and some aspects of modal logic. Chapter 8 removes the assumption that the world is static. To represent a dynamic world requires some notion of time or state change, which, in tum, introduces us to the planning problem. Chapter 9 discusses hypothetical reasoning and its application to default reasoning, diagnostic reasoning, and recognition. Chapter 10 introduces reasoning under uncertainty, representations for uncertain knowledge, and decision making under uncertainty. Chapter 11, which discusses learning, shows how previous experience can be used by an agent. Chapter 12 shows how the reasoning capabilities can be put together to build agents that perceive and interact in an environment. Three appendices provide supplementary material including a glossary of terms used in CI, a Prolog tutorial, and implementations of various system components presented in the main text.20 CHAPTER |. COMPUTATIONAL INTELLIGENCE AND KNOWLEDGE 1.6 References and Further Reading The ideas in this chapter have been derived from many sources. Here we will try to acknowledge those that are explicitly attributable to other authors. Most of the other ideas are part of AI folklore. To try to attribute them to anyone would be impossible. Minsky (1986) presents a theory of intelligence as emergent from a “society” of unintelligent agents. Haugeland (1997) contains a good collection of articles on the philosophy behind computational intelligence. Turing (1950) proposes an objective method for answering the question “Can machines think?” in terms of what is now known as the Turing test. The symbol-system hypothesis is due to Newell & Simon (1976). See also Simon (1996) who discusses the role of symbol systems in a multi-disciplinary context. The distinctions between real, synthetic and artificial intelligence are discussed by Hauge- land (1985), who also provides useful introductory material on interpreted, automatic formal symbol systems, and the Church-Turing thesis. For a critique of the symbol- system hypothesis see Winograd (1990). Wegner (1997) argues that computers that interact with the world may be more powerful than Turing machines, thus that the Church-Turing thesis is in fact false. The Taoist story is from Needham’s classic study of science and technology in China (Ronan, 1986). For discussions on the foundations of AI and the breadth of research in AI see Kirsh (1991a), Bobrow (1993), and the papers in the corresponding volumes, as well as Schank (1990) and Simon (1995). The importance of knowledge in AI is discussed in Lenat & Feigenbaum (1991) and Smith (1991). For overviews of cognitive science and the role that AI and other disciplines play in that field see Gardner (1985), Posner (1989), and Stillings, Feinstein, Garfield, Rissland, Rosenbaum, Weisler & Baker-Ward (1987). A number of AI texts are valuable as reference books complementary to this book, providing a different perspective on AI. See classic books by Nilsson (1971; 1980), Genesereth & Nilsson (1987), Charniak & McDermott (1985) and more recent books including Ginsberg (1993), Russell & Norvig (1995) and Dean, Allen & Aloimonos (1995). The Encyclopedia of Artificial Intelligence (Shapiro, 1992) is an encyclopedic reference on AI written by leaders in the field. There are a number of collections of classic research papers. The general collections of most interest to readers of this book include Webber & Nilsson (1981) and Brachman & Levesque (1985). More specific collections are given in the appropriate chapters. There are many journals that provide in-depth research contributions and confer- ences where the most up-to-date research can be found. These include the journals1.7. EXERCISES 21 Antificial Intelligence, Journal of Artificial Intelligence Research, IEEE Transactions on Pattern Analysis and Machine Intelligence, Computational Intelligence, Interna- tional Journal of Intelligent Systems, and New Generation Computing, as well as more specialized journals such as Neural Computation, Computational Linguistics, Ma- chine Learning, Journal of Automated Reasoning, Journal of Approximate Reasoning, IEEE Transactions on Robotics and Automation, and the Logic Programming Jour- nal. Al Magazine, published by the American Association for Artificial Intelligence (AAAI), often has excellent overview articles and descriptions of particular applica- tions. There are many conferences on Artificial Intelligence. Those of most interest to a general audience are the biennial Jnternational Joint Conference on Artificial Intelli- gence (IJCAI), the European Conference on Al (ECAI), the Pacific Rim International Conference on Al (PRICAI), and various national conferences, especially the Ameri- can Association for Artificial Intelligence National Conference on Al, and innumerable specialized conferences and workshops. 1.7 Exercises Exercise 11 Exercise 1.2 Exercise 1.3 For each of the following, give five reasons why: (a) A dog is more intelligent than a worm. (b) A human is more intelligent that a dog. (c) An organization is more intelligent than an individual human. Based on these, give a definition of what “more intelligent” may mean. Give as many disciplines as you can whose aim is to study intelligent behavior of some sort. For each discipline find out where the behavior is manifest and what tools are used to study it. Be as liberal as you can as to what defines intelligent behavior. Choose a particular world, for example, what is on some part of your desk at the current time. i) Get someone to list all of the things that exist in this world (or try it yourself as a thought experiment). ii) Try to think of twenty things that they missed. Make these as different from each other as possible. For example, the ball at the tip of the right-most ball-point pen on the desk, or the spring in the stapler, or the third word on page 21 of a particular book on the desk. iii) Try to find a thing that can’t be described using natural language. iv) Choose a particular task, such as making the desk tidy, and try to write down all of the things in the world at a level of description that is relevant to this task. Based on this exercise, discuss the following statements:22 CHAPTER |. COMPUTATIONAL INTELLIGENCE AND KNOWLEDGE (a) What exists in a world is a property of the observer. (b) You need general constructors to describe individuals, rather than expecting each individual to have a separate name. (c) What individuals exist is a property of the task as well as the world. (d) To describe the individuals in a domain, you need what is essentially a dictionary of a huge number of words and ways to combine them to describe individuals, and this should be able to be done independently of any particular domain.Chapter 2 A Representation and Reasoning System 2.1 Introduction For a computer to be intelligent it has to be programmed appropriately. Ideally you would like to tell it only as much as it needs to know in a high-level language, and you would like for it to be able to reason and learn how to act intelligently. The language you use to communicate with the computer, the meaning associated with that language, together with a specification of what the computer does in response makes up a representation and reasoning system (RRS) (page 9). This chapter develops the general idea of an RRS. We present the syntax, semantics, and two different proof procedures for one simple representation and reasoning system, namely the definite clause language. This chapter should be read in conjunction with Chapter 3, which discusses how this particular RRS can be used to represent a variety of problems. 2.2 Representation and Reasoning Systems Before providing a more detailed description of a particular representation and reasoning system, we investigate the features we want in an RRS and how to use an RRS when you have one. 2324 CHAPTER 2. A REPRESENTATION AND REASONING SYSTEM e First, begin with a task domain that you want to characterize. This is a domain in which you want the computer to be able to answer questions, solve problems, or interact with directly. For example, the domain may be an office layout, a house’s electrical wiring system, or a university’s course structure. e Second, distinguish things you want to talk about in the domain, namely the ontology (page 10). The different wires and switches may be individuals, but you may not care about the decorative plates for the switches or each individual nail holding the walls together. e Third, use symbols in the computer to represent the objects and relations in the domain. We say the symbols denote the objects in the world. The symbol sz may refer to a particular light switch, or the symbol david may denote a particular person. e Fourth, tell the computer knowledge about that domain. This may include knowledge about what is true in that domain as well as knowledge about how to solve problems in that domain. You may tell the system that switch s2 is connected to wire wo that is in turn connected to light /). e Fifth, ask the RRS a question which prompts the RRS to reason with its knowl- edge to solve problems, produce answers, or sanction an agent’s actions. You may want to know whether light /, is lit based on the knowledge given to the system. In order to tell the computer anything, you have to have a language of communication, you need a way to give meaning to the sentences, and you need a way for the system to compute answers. Together these form a a representation and reasoning system: A Representation and Reasoning System (RRS) is made up of a formal language, a semantics, and a reasoning theory, where: e A formal language specifies the legal sentences you can use to express knowl- edge about a task domain. The language defines the legal symbols and how they can be put together. A language is specified by a grammar. A knowledge base is a set of sentences in the language. e A semantics specifies the meaning of sentences in the language. It specifies a commitment to how symbols in the language relate to the task domain. It permits you to explain the meaning of sentences in the language in terms of the task domain, as well as to understand the meaning of the computer’s an- swers. With this semantic commitment, you can discuss correctness, or truth, of knowledge independently of how it is used. You can define the meaning of a computed answer, independently of how it was computed. It is important that the semantics be natural and intuitive, so that the result of a computation can be easily interpreted.2.2. REPRESENTATION AND REASONING SYSTEMS 25 e A reasoning theory, or proof procedure, is a possibly nondeterministic spec- ification of how an answer can be derived from the knowledge base. This is typically a set of inference rules that sanction consequences left implicit, or is a nondeterministic specification of how an answer is computed. A reasoning theory may be inadequate for the semantics you’ve defined. A reasoning theory is said to be sound if it only generates correct answers with respect to the seman- tics, and it is said to be complete if it generates all correct answers or guarantees to produce one if one exists. (See page 46 for more details on soundness and completeness.) An implementation of an RRS consists of a language parser and a reasoning procedure: e A language parser is a program that distinguishes legal sentences of the formal language and, given legal sentences, produces some internal form to be stored as data structures in the computer. e A reasoning procedure is an implementation of a reasoning theory together with a search strategy. You can’t implement a reasoning theory without com- mitting to some way to resolve the nondeterminism in the reasoning theory; this is the role of a search strategy. A reasoning procedure manipulates the data structures produced by the parser, to produce answers. There can be many different implementations of the same reasoning theory such as where the rea- soning theory is implemented in different languages, using different encoding techniques, or using different search strategies. A reasoning procedure may be incomplete either because there is some conclusion that the semantics specifies should exists but which the reasoning theory cannot derive, or because the search strategy can’t find an answer sanctioned by the reasoning theory. The semantics isn’t reflected in the definition of the implementation above, but there are corresponding elements to the formal language and the reasoning theory. The semantics isn’t something that is actually implemented; it gives a meaning to the symbols for an external viewer. The semantics also specifies what is a correct conclu- sion. Thus you can study the soundness and completeness of a reasoning procedure with respect to the semantics. Figure 2.1 illustrates the relationship between the components. On the left is a computer implementing the reasoning procedure. Sentences in the language are fed into it and answers come out. The computer doesn’t know what the symbols mean. The person who supplied the information has an associated meaning with the symbols. They can check the truth of the inputs according to the semantics and can interpret the output according to that semantics. Associated with an RRS is a representational methodology. This is a specifica- tion of how the RRS can be used. It could be provided as a user manual for the RRS. The methodology is derived through design and through experience. The value of a26 CHAPTER 2. A REPRESENTATION AND REASONING SYSTEM in(alan, r123) part of (r123. cs building). in(X, ¥) part of Z.Y) A in(X,Z). > in(alan,cs_building) The meaning of the symbols are in the user’s head. The computer takes in symbols and outputs symbols. The output can be interpreted by the user according to the meaning they place on the symbols. Figure 2.1: The role of semantics in a representation and reasoning system methodology is best understood if you consider that many RRSs are equivalent to a Turing machine in that they can compute any computable function; any differences between these RRSs arise in what they can do naturally, simply, and efficiently. Of- ten the form and the assumptions behind an RRS impose constraints on what can be represented, and thus may make the RRS well-suited for some domains and inappro- priate for others. For example, some RRSs allow the representation of uncertainty in a domain, whereas others can represent only certain knowledge. Just as the reasoning procedure was divided into a proof procedure and a search strategy, there is a corresponding distinction in the programming methodology be- tween specification of the logic of the domain and the specification of the control to find a solution to the problem. Different semantics may impose different views of the world and, therefore, different ways to conceptualize the problem. Similarly, different search procedures may provide different ways to specify control information to im- prove reasoning efficiency. For example, some RRSs allow the order of the sentences to contain heuristic control information without having any corresponding semantic information. For any RRS, you need programming methodologies that guide both the2.3. SIMPLIFYING ASSUMPTIONS OF THE INITIAL RRS 27 logical specification and the implicit and explicit control knowledge. While these aspects of an RRS might appear in different guises, all are important. You can’t record anything in the computer without syntax, though it may take various forms, such as text, graphics, or even sound. A semantics without a proof procedure is useless since you actually want to compute. A reasoning procedure without a semantics is also useless, since it fails to provide any way to relate computed consequences to the task domain you are trying to characterize. A reasoning procedure which isn’t sound can’t be relied on to maintain fidelity with the task domain. A reasoning procedure which is sound but not complete may produce useful consequences, but no meaning can be assigned to its failure to establish the truth of an answer; it may be that an answer exists, but the reasoning procedure can’t find it. This notion of an RRS should be compared with more familiar programming language definitions, which typically consist of a formal language, a parser, and a reasoning procedure but no semantics. Often there is a reasoning theory in terms of a procedural semantics, but usually there is a less formal specification of program execution. There is usually no way to interpret the result of a program independently of how it is computed. The above definitions of semantics and reasoning theory correspond to the notions of a Tarskian semantics and proof in mathematical logic. With logic, you can define knowledge independently of how it is used. You can verify the correctness of knowl- edge since you know its meaning. Under a particular domain and interpretation, you can debate the truth of sentences in the language. You can use the same semantics to establish the correctness of an implementation. The separation of a program’s meaning from how it’s computed lets you exploit performance ideas such as parallelism or domain-specific heuristics with the confi- dence of correct results. A program’s correctness is defined by the semantics, not by a particular control strategy. As long as you are faithful to the semantics, the control can be optimized for the machine used. Logic provides the specification of meaning in terms of the relationship between the symbols in a computer and the task domain. We make no assumption that the reasoning procedures are just doing deduction. It is the commitment to a semantics that is inherited from logic, not any particular class of proof procedures. 2.3 Simplifying Assumptions of the Initial RRS Scientific progress on difficult problems often results from solving simpler, more abstract but related problems. For example, classical physics developed under the unrealistic but extremely useful assumption of a friction-less world.28 Assumption IR Assumption DK Assumption SE CHAPTER 2. A REPRESENTATION AND REASONING SYSTEM Similarly, to begin the logical analysis of computational intelligence we shall make various explicit assumptions. Given the view of an intelligent agent in an environment, there are three kinds of assumptions you can make. First, you could assume certain properties are true of the agent; these are type A assumptions. Second, you could assume certain properties are true of the environment; these are type E assumptions. Third, you could make assumptions, of type AE, about the relationship between an agent and its environment. First we restate the ontological assumption (page 10): (Individuals and Relations) An agent’s knowledge can be usefully described in terms of individuals and relations among individuals. The notion of relations includes nullary relations that are true or false, unary relations that are properties of single individuals such as their color or height, binary relations between pairs of individuals, ternary relations among triples of individuals, and rela- tions amongst more individuals. Assumption IR underlies the approach described in this book. This is an AE assumption. It says essentially that linguistic or symbolic knowledge is all that an agent needs to act. Of course, to build a real robot, you need some nonlinguistic knowledge such as how to interpret sounds as speech or how to interpret two-dimensional stimuli on the retina or a camera into features. This assumption implies that results of these perceptive tasks should be symbolic knowledge. We don’t focus on these perceptual tasks because they are well covered elsewhere. This assumption isn’t saying that the world is made up of things and relations among things, but rather that the world can be described in these terms. To the assumption IR we now add two quite restrictive assumptions, both of which are waived or relaxed as we develop our approach. (Definite Knowledge) An agent's knowledge base consists of definite and positive statements. We assume that the sentences in the knowledge base aren’t vague or negative. Knowing that either “Craig is in room r1 13” or “Craig is in the lab,” but not which, is vague. Knowing that “Craig isn’t in room rl17” is negative. This chapter defines definite knowledge precisely. Since the assumption applies only to the form of the agent’s knowledge base, it is of type A. This assumption is relaxed in Chapter 7. One assumption that we’re not making at this point is that the agent’s knowledge is complete. We don’t assume the agent knows everything. Section 7.4 (page 248) con- siders agents with complete knowledge that conclude that a sentence is false because they have no knowledge to the contrary. (Static Environment) The environment is static. This assumption allows you to ignore change. We assume that no actions occur which change the environment while the agent is constructing its representation or reasoning