Artificial intelligence george f luger 6th edition pdf download

SIXTH EDITION

This page intentionally left blank

Boston San Francisco New York London Toronto Sydney Tokyo Singapore Madrid

Mexico City Munich Paris Cape Town Hong Kong Montreal

Structures and Strategies for Complex Problem Solving

George F Luger University of New Mexico

SIXTH EDITION

Executive Editor Michael Hirsch Acquisitions Editor Matt Goldstein Editorial Assistant Sarah Milmore Associate Managing Editor Jeffrey Holcomb Digital Assets Manager Marianne Groth Senior Media Producer Bethany Tidd Marketing Manager Erin Davis Senior Author Support/

Technology Specialist Joe Vetere Senior Manufacturing Buyer Carol Melville Text Design, Composition, and Illustrations George F Luger Cover Design Barbara Atkinson Cover Image © Tom Barrow

For permission to use copyrighted material, grateful acknowledgment is made to the copyright holders listed on page xv, which is hereby made part of this copyright page.

Many of the designations used by manufacturers and sellers to distinguish their products are claimed as trademarks. Where those designations appear in this book, and Addison-Wesley was aware of a trademark claim, the designations have been printed in initial caps or all caps.

Library of Congress Cataloging-in-Publication Data

Luger, George F. Artificial intelligence : structures and strategies for complex problem solving / George F. Luger.-- 6th ed. p. cm. Includes bibliographical references and index. ISBN-13: 978-0-321-54589-3 (alk. paper) 1. Artificial intelligence. 2. Knowledge representation (Information theory) 3. Problem solving. 4. PROLOG (Computer program language) 5. LISP (Computer program language) I. Title. Q335.L84 2008 006.3--dc22 2007050376

Copyright © 2009 Pearson Education, Inc. 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 written permission of the publisher. Printed in the United States of America. For information on obtaining permission for use of material in this work, please submit a written request to Pearson Education, Inc., Rights and Contracts Department, 501 Boylston Street, Suite 900, Boston, MA 02116, fax (617) 671-3447, or online at http://www.pearsoned.com/legal/permissions.htm.

ISBN-13: 978-0-321-54589-3 ISBN-10: 0-321-54589-3

1 2 3 4 5 6 7 8 9 10—CW—12 11 10 09 08

http://www.pearsoned.com/legal/permissions.htm
For my wife, Kathleen, and our children Sarah, David, and Peter.

Si quid est in me ingenii, judices . . .

Cicero, Pro Archia Poeta

GFL

This page intentionally left blank

PREFACE vii

PREFACE

What we have to learn to do we learn by doing. . .

—ARISTOTLE, Ethics

Welcome to the Sixth Edition!

I was very pleased to be asked to produce the sixth edition of my artificial intelligence book. It is a compliment to the earlier editions, started over twenty years ago, that our approach to AI has been so highly valued. It is also exciting that, as new development in the field emerges, we are able to present much of it in each new edition. We thank our many readers, colleagues, and students for keeping our topics relevant and our presenta- tion up to date.

Many sections of the earlier editions have endured remarkably well, including the presentation of logic, search algorithms, knowledge representation, production systems, machine learning, and, in the supplementary materials, the programming techniques developed in Lisp, Prolog, and with this edition, Java. These remain central to the practice of artificial intelligence, and a constant in this new edition.

This book remains accessible. We introduce key representation techniques including logic, semantic and connectionist networks, graphical models, and many more. Our search algorithms are presented clearly, first in pseudocode, and then in the supplementary mate- rials, many of them are implemented in Prolog, Lisp, and/or Java. It is expected that the motivated students can take our core implementations and extend them to new exciting applications.

We created, for the sixth edition, a new machine learning chapter based on stochastic methods (Chapter 13). We feel that the stochastic technology is having an increasingly larger impact on AI, especially in areas such as diagnostic and prognostic reasoning, natu- ral language analysis, robotics, and machine learning. To support these emerging technol- ogies we have expanded the presentation of Bayes' theorem, Markov models, Bayesian

viii PREFACE

belief networks, and related graphical models. Our expansion includes greater use of prob- abilistic finite state machines, hidden Markov models, and dynamic programming with the Earley parser and implementing the Viterbi algorithm. Other topics, such as emergent computation, ontologies, stochastic parsing algorithms, that were treated cursorily in ear- lier editions, have grown sufficiently in importance to merit a more complete discussion. The changes for the sixth edition reflect emerging artificial intelligence research questions and are evidence of the continued vitality of our field.

As the scope of our AI project grew, we have been sustained by the support of our publisher, editors, friends, colleagues, and, most of all, by our readers, who have given our work such a long and productive life. We remain excited at the writing opportunity we are afforded: Scientists are rarely encouraged to look up from their own, narrow research interests and chart the larger trajectories of their chosen field. Our readers have asked us to do just that. We are grateful to them for this opportunity. We are also encouraged that our earlier editions have been used in AI communities worldwide and translated into a number of languages including German, Polish, Portuguese, Russian, and two dialects of Chinese!

Although artificial intelligence, like most engineering disciplines, must justify itself to the world of commerce by providing solutions to practical problems, we entered the field of AI for the same reasons as many of our colleagues and students: we want to under- stand and explore the mechanisms of mind that enable intelligent thought and action. We reject the rather provincial notion that intelligence is an exclusive ability of humans, and believe that we can effectively investigate the space of possible intelligences by designing and evaluating intelligent artifacts. Although the course of our careers has given us no cause to change these commitments, we have arrived at a greater appreciation for the scope, complexity, and audacity of this undertaking. In the preface to our earlier editions, we outlined three assertions that we believed distinguished our approach to teaching artifi- cial intelligence. It is reasonable, in writing a preface to the present edition, to return to these themes and see how they have endured as our field has grown.

The first of these goals was to unify the diverse branches of AI through a detailed dis- cussion of its theoretical foundations. At the time we first adopted that goal, it seemed that the main problem was in reconciling researchers who emphasized the careful statement and analysis of formal theories of intelligence (the neats) with those who believed that intelligence itself was some sort of grand hack that could be best approached in an appli- cation-driven, ad hoc manner (the scruffies). That dichotomy has proven far too simple.

In contemporary AI, debates between neats and scruffies have given way to dozens of other debates between proponents of physical symbol systems and students of neural net- works, between logicians and designers of artificial life forms that evolve in a most illogi- cal manner, between architects of expert systems and case-based reasoners, and finally, between those who believe artificial intelligence has already been achieved and those who believe it will never happen. Our original image of AI as frontier science where outlaws, prospectors, wild-eyed prairie prophets and other dreamers were being slowly tamed by the disciplines of formalism and empiricism has given way to a different metaphor: that of a large, chaotic but mostly peaceful city, where orderly bourgeois neighborhoods draw their vitality from diverse, chaotic, bohemian districts. Over the years that we have devoted to the different editions of this book, a compelling picture of the architecture of intelligence has started to emerge from this city's structure, art, and industry.

PREFACE ix

Intelligence is too complex to be described by any single theory; instead, researchers are constructing a hierarchy of theories that characterize it at multiple levels of abstrac- tion. At the lowest levels of this hierarchy, neural networks, genetic algorithms and other forms of emergent computation have enabled us to understand the processes of adaptation, perception, embodiment, and interaction with the physical world that must underlie any form of intelligent activity. Through some still partially understood resolution, this chaotic population of blind and primitive actors gives rise to the cooler patterns of logical infer- ence. Working at this higher level, logicians have built on Aristotle's gift, tracing the out- lines of deduction, abduction, induction, truth-maintenance, and countless other modes and manners of reason. At even higher levels of abstraction, designers of diagnostic sys- tems, intelligent agents, and natural language understanding programs have come to rec- ognize the role of social processes in creating, transmitting, and sustaining knowledge.

At this point in the AI enterprise it looks as though the extremes of rationalism and empiricism have only led to limited results. Both extremes suffer from limited applicabil- ity and generalization. The author takes a third view, that the empiricist's conditioning: semantic nets, scripts, subsumption architectures and the rationalist's clear and distinct ideas: predicate calculus, non-monotonic logics, automated reasoning - suggest a third viewpoint, the Bayesian. The experience of relational invariances conditions intelligent agents's expectations, and learning these invariances, in turn, bias future expectations. As philosophers we are charged to critique the epistemological validity of the AI enterprise. For this task, in Chapter 16 we discuss the rationalist project, the empiricists dilemma, and propose a Bayesian based constructivist rapprochement. In this sixth edition, we touch on all these levels in the presenting the AI enterprise.

The second commitment we made in earlier editions was to the central position of advanced representational formalisms and search techniques in AI methodology. This is, perhaps, the most controversial aspect of our previous editions and of much early work in AI, with many researchers in emergent computation questioning whether symbolic rea- soning and referential semantics have any role at all in intelligence. Although the idea of representation as giving names to things has been challenged by the implicit representa- tion provided by the emerging patterns of a neural network or an artificial life, we believe that an understanding of representation and search remains essential to any serious practi- tioner of artificial intelligence. We also feel that our Chapter 1 overview of the historical traditions and precursors of AI are critical components of AI education. Furthermore, these are invaluable tools for analyzing such aspects of non-symbolic AI as the expressive power of a neural network or the progression of candidate problem solutions through the fitness landscape of a genetic algorithm. Comparisons, contrasts, and a critique of modern AI are offered in Chapter 16.

Our third commitment was made at the beginning of this book's life cycle: to place artificial intelligence within the context of empirical science. In the spirit of the Newell and Simon (1976) Turing award lecture we quote from an earlier edition:

... AI is not some strange aberration from the scientific tradition, but . . . part of a general quest for knowledge about, and the understanding of, intelligence itself. Furthermore, our AI programming tools, along with the exploratory programming methodology . . . are ideal for exploring an environment. Our tools give us a medium for both understanding

x PREFACE

and questions. We come to appreciate and know phenomena constructively, that is, by pro- gressive approximation.

Thus we see each design and program as an experiment with nature: we propose a representation, we generate a search algorithm, and we question the adequacy of our char- acterization to account for part of the phenomenon of intelligence. And the natural world gives a response to our query. Our experiment can be deconstructed, revised, extended, and run again. Our model can be refined, our understanding extended.

New with The Sixth Edition

The biggest change for the sixth edition is the extension of the stochastic approaches to AI. To accomplish this we revised Section 9.3 and added a new chapter (13) introducing probability-based machine learning. Our presentation of stochastic AI tools and their application to learning and natural language is now more comprehensive.

From probability theory's foundations in set theory and counting we develop the notions of probabilities, random variables, and independence. We present and use Bayes' theorem first with one symptom and one disease and then in its full general form. We examine the hypotheses that underlie the use of Bayes and then present the argmax and naive Bayes approaches. We present examples of stochastic reasoning, including the anal- ysis of language phenomena and the Vierbi algorithm. We also introduce the idea of condi- tional independence that leads to Bayesian belief networks, the BBN, in Chapter 9.

In Chapter 13 we introduce hidden Markov models, HMMs, and show their use in several examples. We also present several HMM variants, including the auto-regressive and hierarchical HMMs. We present dynamic Bayesian networks, DBNs, and demonstrate their use. We discuss parameter and structure learning and present the expectation maxi- mization algorithm and demonstrate its use with loopy belief propagation. Finally, we present Markov decision processes, the MDP, and partially observable Markov decision process, the POMDP, in the context of an extension to the earlier presentation of reinforce- ment learning.

We include several more examples of probabilistic finite state machines and probabi- listic acceptors, as well as the use of dynamic programming, especially with stochastic measures (the Viterbi algorithm). We added a stochastic English language parser (based on the work of Mark Steedman at the University of Edinburgh) as well as the use of dynamic programming with the Earley parser.

We made a major decision to remove the Prolog and Lisp chapters from the book. Part of the reason for this is that these were getting too large. We have also accumulated a number of AI algorithms written in Java. When we added the new Chapter 13 on stochas- tic approaches to machine learning, we determined that the book was getting too large/ cumbersome. Thus the sixth edition is more than 150 pages smaller than the fifth and the AI algorithms in Prolog, Lisp, and Java are being released as supplementary materials. From our earliest days in AI we have always felt that the way to understand the power (and limitations) of AI algorithms is constructively - that is, by building them! We encour- age our present generation of readers to do exactly this: to visit the supplementary materi- als: to build and experiment directly with the algorithms we present.

PREFACE xi

Finally, we have done the usual updating of references and materials that a new edi- tion warrants. In a revised Chapter 16, we return to the deeper questions on the nature of intelligence and the possibility of creating intelligent machines.

Sixth Edition: The Contents

Chapter 1 introduces artificial intelligence, beginning with a brief history of attempts to understand mind and intelligence in philosophy, psychology, and other areas of research. In an important sense, AI is an old science, tracing its roots back at least to Aristotle. An appreciation of this background is essential for an understanding of the issues addressed in modern research. We also present an overview of some of the important application areas in AI. Our goal in Chapter 1 is to provide both background and a motivation for the theory and applications that follow.

Chapters 2, 3, 4, 5, and 6 (Part II) introduce the research tools for AI problem solving. These include, in Chapter 2, the predicate calculus presented both as a mathemat- ical system as well as a representation language to describe the essential features of a problem. Search, and the algorithms and data structures used to implement search, are introduced in Chapter 3, to organize the exploration of problem situations. In Chapter 4, we discuss the essential role of heuristics in focusing and constraining search-based prob- lem solving. In Chapter 5, we introduce the stochastic methodology, important technology for reasoning in situations of uncertainty. In Chapter 6, we present a number of software architectures, including the blackboard and production system, for implementing these search algorithms.

Chapters 7, 8, and 9 make up Part III: representations for AI, knowledge-inten- sive problem solving, and reasoning in changing and ambiguous situations. In Chap- ter 7 we present the evolving story of AI representational schemes. We begin with a discussion of association-based networks and extend this model to include conceptual dependency theory, frames, and scripts. We then present an in-depth examination of a par- ticular formalism, conceptual graphs, emphasizing the epistemological issues involved in representing knowledge and showing how these issues are addressed in a modern repre- sentation language. Expanding on this formalism in Chapter 14, we show how conceptual graphs can be used to implement a natural language database front end. We conclude Chapter 7 with more modern approaches to representation, including Copycat and agent- oriented architectures.

Chapter 8 presents the rule-based expert system along with case-based and model- based reasoning, including examples from the NASA space program. These approaches to problem solving are presented as a natural evolution of the material in Part II: using a pro- duction system of predicate calculus expressions to orchestrate a graph search. We end with an analysis of the strengths and weaknesses of each of these approaches to knowl- edge-intensive problem solving.

Chapter 9 presents models for reasoning with uncertainty as well as the use of unreli- able information. We introduce Bayesian models, belief networks, Dempster-Shafer, causal models, and the Stanford certainty algebra for reasoning in uncertain situations.

xii PREFACE

Chapter 9 also contains algorithms for truth maintenance, reasoning with minimum mod- els, logic-based abduction, and the clique-tree algorithm for Bayesian belief networks.

Part IV, Chapters 10 through 13, is an extensive presentation of issues in machine learning. In Chapter 10 we offer a detailed look at algorithms for symbol-based learning, a fruitful area of research spawning a number of different problems and solution approaches. These learning algorithms vary in their goals, the training data considered, their learning strategies, and the knowledge representations they employ. Symbol-based learning includes induction, concept learning, version-space search, and ID3. The role of inductive bias is considered, generalizations from patterns of data, as well as the effective use of knowledge to learn from a single example in explanation-based learning. Category learning, or conceptual clustering, is presented with unsupervised learning. Reinforcement learning, or the ability to integrate feedback from the environment into a policy for mak- ing new decisions concludes the chapter.

In Chapter 11 we present neural networks, often referred to as sub-symbolic or con- nectionist models of learning. In a neural net, information is implicit in the organization and weights on a set of connected processors, and learning involves a re-arrangement and modification of the overall weighting of nodes and structure of the system. We present a number of connectionist architectures, including perceptron learning, backpropagation, and counterpropagation. We demonstrate Kohonen, Grossberg, and Hebbian models. We present associative learning as well as attractor models, including examples of Hopfield networks.

Genetic algorithms and evolutionary approaches to learning are introduced in Chapter 12. On this viewpoint, learning is cast as an emerging and adaptive process. After several examples of problem solutions based on genetic algorithms, we introduce the application of genetic techniques to more general problem solvers. These include classifier systems and genetic programming. We then describe society-based learning with examples from artificial life, called a-life, research. We conclude the chapter with an example of emergent computation from research at the Santa Fe Institute.

Chapter 13 presents stochastic approaches to machine learning. We begin with a defi- nition of hidden markov models and then present several important variations including the auto-regressive and hierarchical HMM. We then present dynamic Bayesian networks, a generalization of the HMM, and also able to track systems across periods of time. These techniques are useful for modeling the changes in complex environments as is required for diagnostic and prognostic reasoning. Finally, we add a probabilistic component to rein- forcement learning first introduced in Chapter 10. This includes presentation of the Markov decision process (or MDP) and the partially observed Markov decision process (or POMDP).

Part V, Chapters 14 and 15, presents automated reasoning and natural language understanding. Theorem proving, often referred to as automated reasoning, is one of the oldest areas of AI research. In Chapter 14, we discuss the first programs in this area, including the Logic Theorist and the General Problem Solver. The primary focus of the chapter is binary resolution proof procedures, especially resolution refutations. More advanced inferencing with hyper-resolution and paramodulation is also presented. Finally, we describe the Prolog interpreter as a Horn clause and resolution-based inferencing sys- tem, and see Prolog computing as an instance of the logic programming paradigm.

PREFACE xiii

Chapter 15 presents natural language understanding. Our traditional approach to lan- guage understanding, exemplified by many of the semantic structures presented in Chap- ter 7, is complemented with the stochastic approach. These include using Markov models, CART trees, CHART parsing (the Earley algorithm), mutual information clustering, and statistics-based parsing. The chapter concludes with examples applying natural language techniques to database query generation, a text summarization systems well as.the use of machine learning to generalize extracted results from the WWW.

Finally, Chapter 16 serves as an epilogue for the book. It addresses the issue of the possibility of a science of intelligent systems, and considers contemporary challenges to AI; it discusses AI's current limitations, and projects its exciting future.

Using This Book

Artificial intelligence is a big field, and consequently, this is a large book. Although it would require more than a single semester to cover all of the material offered, we have designed our book so that a number of paths may be taken through the material. By select- ing subsets of the material, we have used this text for single semester and full year (two semester) courses.

We assume that most students will have had introductory courses in discrete mathe- matics, including predicate calculus, set theory, counting, and graph theory. If this is not true, the instructor should spend more time on these concepts in the “optional” sections at the beginning of the introductory chapters (2.1, 3.1, and 5.1). We also assume that students have had courses in data structures including trees, graphs, and recursion-based search, using stacks, queues, and priority queues. If they have not, then spend more time on the beginning sections of Chapters 3, 4, and 6.

In a one quarter or one semester course, we go quickly through the first two parts of the book. With this preparation, students are able to appreciate the material in Part III. We then consider the Prolog, Lisp, or the Java code in the supplementary materials for the book and require students to build many of the representation and search techniques of the second part of the book. Alternatively, one of the languages, Prolog, for example, can be introduced early in the course and be used to test out the data structures and search tech- niques as that are encountered. We feel the meta-interpreters presented in the language materials are very helpful for building rule-based and other knowledge-intensive problem solvers. Prolog, Lisp, and Java are excellent tools for building natural language under- standing and learning systems; these architectures are presented in Parts II and III and there are examples of them in the supplementary course materials.

In a two-semester or three-quarter course, we are able to cover the application areas of Parts IV and V, especially the machine learning chapters, in appropriate detail. We also expect a much more detailed programming project from students. We also think that it is very important in the second semester for students to revisit many of the primary sources of the AI literature. It is crucial for students to see both where we are in the evolution of the AI enterprise, as well as how we got here, and to have an appreciation of the future promises of artificial intelligence. We use materials from the WWW for this purpose or select a collected set of readings, such as, Computation and Intelligence (Luger 1995).

xiv PREFACE

The algorithms of our book are described using a Pascal-like pseudo-code. This nota- tion uses the control structures of Pascal along with English descriptions of the tests and operations. We have added two useful constructs to the Pascal control structures. The first is a modified case statement that, rather than comparing the value of a variable with con- stant case labels, as in standard Pascal, lets each item be labeled with an arbitrary boolean test. The case evaluates these tests in order until one of them is true and then performs the associated action; all other actions are ignored. Those familiar with Lisp will note that this has the same semantics as the Lisp cond statement.

The other addition to our pseudo code language is a return statement which takes one argument and can appear anywhere within a procedure or function. When the return is encountered, it causes the program to immediately exit the function, returning its argu- ment as a result. Other than these modifications we used Pascal structure, with a reliance on the English descriptions, to make the algorithms clear.

Supplemental Material Available

The sixth edition has an attached web site maintained by my graduate students. This site, built originally by two UNM students, Alejandro CdeBaca and Cheng Liu, includes sup- plementary ideas for most chapters, some sample problems with their solutions, and ideas for student projects. Besides the Prolog, Lisp, and Java programs in the supplementary materials for this book, we have included many other AI algorithms in Java and C++ on the web site. Students are welcome to use these and supplement them with their own com- ments, code, and critiques. The web url is www.cs.unm.edu/~luger/ai-final/.

The Prolog, Lisp, and Java programs implementing many of the AI data structures and search algorithms of the book are available through your Addison-Wesley Pearson Education representative. There is also an Instructor’s Guide available which has many of the book's exercises worked out, several practice tests with solutions, a sample syllabus, and ideas supporting teaching the material. There are also a full set of PowerPoint presen- tation materials for use by instructors adopting this book. Again, consult your local A-W Pearson representative for access and visit www.aw.com/luger.

My e-mail address is luger@cs.unm.edu, and I enjoy hearing from my readers.

Acknowledgements

Although I am the sole author of the sixth edition, this book has always been the product of my efforts as Professor of Computer Science, Psychology, and Linguistics at the Uni- versity of New Mexico along with my fellow faculty, professional colleagues, graduate students, and friends. The sixth edition is also the product of the many readers that have e- mailed comments, corrections, and suggestions. The book will continue this way, reflect- ing a community effort; consequently, I will continue using the prepositions we, our, and us when presenting material.

I thank Bill Stubblefield, the co-author for the first three editions, for more than twenty years of contributions, but even more importantly, for his friendship. I also thank

www.cs.unm.edu/~luger/ai-final/
www.aw.com/luger
PREFACE xv

the many reviewers that have helped develop this and earlier editions. These include Den- nis Bahler, Leonardo Bottaci, Skona Brittain, Philip Chan, Peter Collingwood, Mehdi Dastani, John Donald, Sarah Douglas, Christophe Giraud-Carrier, Andrew Kosoresow, Terran Lane, Chris Malcolm, Ray Mooney, Marek Perkowski, Barak Pearmutter, Dan Pless, Bruce Porter, Stuart Shapiro, Julian Richardson, Jude Shavlik, John Sheppard, Carl Stern, Leon van der Torre, Marco Valtorta, and Bob Veroff. We also appreciate the numer- ous suggestions and comments sent directly by e-mail by readers. Finally, Chris Malcolm, Brendan McGonnigle, and Akasha Tang, critiqued Chapter 16.

From our UNM colleagues, we thank Dan Pless, Nikita Sakhanenko, Roshan Ram- mohan, and Chayan Chakrabarti for their major role in developing materials for Chapters 5, 9, and 13; Joseph Lewis for his efforts on Chapters 9 and 16; Carl Stern for his help in developing Chapter 11 on connectionist learning; Bob Veroff for his critique of the auto- mated reasoning material in Chapter 14; and Jared Saia, Stan Lee, and Paul dePalma for helping with the stochastic approaches to natural language understanding of Chapter 15.

We thank Academic Press for permission to reprint much of the material of Chapter 11; this first appeared in the book Cognitive Science: The Science of Intelligent Systems (Luger 1994). Finally, we thank more than two decades of students who have used various versions of this book and software at UNM for their help in expanding our horizons, as well as in removing typos and bugs.

We thank our many friends at Benjamin-Cummings, Addison-Wesley-Longman, and Pearson Education for their support and encouragement in completing the writing task of our six editions, especially Alan Apt in helping us with the first edition, Lisa Moller and Mary Tudor for their help on the second, Victoria Henderson, Louise Wilson, and Karen Mosman for their assistance on the third, Keith Mansfield, Karen Sutherland, and Anita Atkinson for support on the fourth, Keith Mansfield, Owen Knight, Anita Atkinson, and Mary Lince for their help on the fifth edition, and Simon Plumtree, Matt Goldstein, Joe Vetere, and Sarah Milmore for their help on the sixth. Katherine Haratunian of Addison Wesley has had a huge role in seeing that Professors received their Instructor Guides and PowerPoint presentation materials (These are maintained by Addison-Wesley Pearson and available only through your local sales representative). Linda Cicarella of the University of New Mexico helped prepare many of the figures.

We thank Thomas Barrow, internationally recognized artist and University of New Mexico Professor of Art (emeritus), who created the photograms for this book.

Artificial intelligence is an exciting and rewarding discipline; may you enjoy your study as you come to appreciate its power and challenges.

George Luger 1 January 2008 Albuquerque

xvi

PUBLISHER’S ACKNOWLEDGEMENTS

We are grateful to the following for permission to reproduce copyright material:

Figures 4.8 and 4.9, Tables 5.2 and 5.3 adapted from Figure 5.6, p. 157, Figure 5.18, p. 178, Figure 5.20, p. 180 and data on p. 167, from Speech and Language Processing: an introduction to natural language processing, computation linguistics, and speech recognition, Prentice Hall, (Pearson Education, Inc.), (Jurafsky, D., and Martin, J. H., 2000); Figure 5.3 adapted from Figure 5.12, p. 170, from Speech and Language Processing: an introduction to natural language processing, computational linguistics, and speech recognition, Prentice Hall, (Pearson Education, Inc.), (Jurafsky, D., and Martin, J.H., 2000), which was itself adapted from a figure from Artificial Intelligence: A Modern approach, 1st Edition, Prentice Hall, (Pearson Education, Inc.), (Russell, S. J., and Norvig, P., 1995); Figure 7.1 from figure from Expert Systems: Artificial Intelligence in Business, by Harmon, P. and King, D., John Wiley & Sons, Inc., Copyright © 1985 Paul Harmon and David King. This material is used by permission of John Wiley & Sons, Inc.; Figures 7.6, 7.9, and 7.10 from figures from “Inference and the computer understanding of natural language,” in Artificial Intelligence, Vol. 5, No. 4, 1974, pp. 373–412, Copyright © 1974, reprinted with permission from Elsevier Science, (Schank, R. C., and Reiger, C. J., 1974); Figure 7.27 from figures from Analogy-Making as Perception: A Computer Model, The MIT Press, (Mitchell, M., 1993); Figure 9.2 from “an improved algorithm for non-monotonic dependency net update,” in Technical Report LITH-MAT-R-82-83, reprinted by permission of the author, (Goodwin, J., 1982); Figure10.21 adapted from figure from “models of incremental concept formation,” in Artificial Intelligence, Vol. 40, Nos. 1–3, 1989, pp. 11-62, Copyright © 1989, reprinted with permission from Elsevier Science, (Gennari, J. H., Langley, P., and Fisher, D., 1989); Figure 11.18 from part of Figure 6.2, p. 102, from Introduction to Support Vector Machines: and other kernel-based learning methods, Cambridge University Press, (Cristianini, N., and Shawe-Taylor, J., 2000).

Academic Press for Chapter 11, adapted from Cognitive Science: The Science of Intelligent Systems, (Luger, G. F., 1994). American Society for Public Administration for an abridged extract from “Decision-making and administrative organization,” in Public Administration Review, Vol. 4, Winter 1944, (Simon, H. A., 1944).

In some instances we have been unable to trace the owners of copyright materials, and we would appreciate any information that would enable us to do so.

xvii

BRIEF CONTENTS

Preface vii Publisher’s Acknowledgements xv

PART I ARTIFICIAL INTELLIGENCE: ITS ROOTS AND SCOPE 1

1 AI: HISTORY AND APPLICATIONS 3

PART II ARTIFICIAL INTELLIGENCE AS REPRESENTATION AND SEARCH 35

2 THE PREDICATE CALCULUS 45

3 STRUCTURES AND STRATEGIES FOR STATE SPACE SEARCH 79

4 HEURISTIC SEARCH 123

5 STOCHASTIC METHODS 165

6 CONTROL AND IMPLEMENTATION OF STATE SPACE SEARCH 193

PART III CAPTURING INTELLIGENCE: THE AI CHALLENGE 223

7 KNOWLEDGE REPRESENTATION 227

xviii BRIEF CONTENTS

PART III (continued)

8 STRONG METHOD PROBLEM SOLVING 277

9 REASONING IN UNCERTAIN SITUATIONS 333

PART IV MACHINE LEARNING 385

10 MACHINE LEARNING: SYMBOL-BASED 387

11 MACHINE LEARNING: CONNECTIONIST 453

12 MACHINE LEARNING: GENETIC AND EMERGENT 507

13 MACHINE LEARNING: PROBABILISTIC 543

PART V ADVANCED TOPICS FOR AI PROBLEM SOLVING 573

14 AUTOMATED REASONING 575

15 UNDERSTANDING NATURAL LANGUAGE 619

PART VI EPILOGUE 671

16 ARTIFICIAL INTELLIGENCE AS EMPIRICAL ENQUIRY 673

Bibliography 705 Author Index 735 Subject Index 743

xix

CONTENTS

Preface vii Publisher’s Acknowledgements xv

PART I ARTIFICIAL INTELLIGENCE: ITS ROOTS AND SCOPE 1

1 AI: HISTORY AND APPLICATIONS 3 1.1 From Eden to ENIAC: Attitudes toward Intelligence, Knowledge, and

Human Artifice 3 1.2 Overview of AI Application Areas 20 1.3 Artificial Intelligence—A Summary 30 1.4 Epilogue and References 31 1.5 Exercises 33

PART II ARTIFICIAL INTELLIGENCE AS REPRESENTATION AND SEARCH 35

2 THE PREDICATE CALCULUS 45 2.0 Introduction 45 2.1 The Propositional Calculus 45 2.2 The Predicate Calculus 50 2.3 Using Inference Rules to Produce Predicate Calculus Expressions 62 2.4 Application: A Logic-Based Financial Advisor 73 2.5 Epilogue and References 77 2.6 Exercises 77

xx CONTENTS

3 STRUCTURES AND STRATEGIES FOR STATE SPACE SEARCH 79 3.0 Introduction 79 3.1 Graph Theory 82 3.2 Strategies for State Space Search 93 3.3 Using the State Space to Represent Reasoning with the Predicate Calculus 107 3.4 Epilogue and References 121 3.5 Exercises 121

4 HEURISTIC SEARCH 123 4.0 Introduction 123 4.1 Hill Climbing and Dynamic Programming 127 4.2 The Best-First Search Algorithm 133 4.3 Admissibility, Monotonicity, and Informedness 145 4.4 Using Heuristics in Games 150 4.5 Complexity Issues 157 4.6 Epilogue and References 161 4.7 Exercises 162

5 STOCHASTIC METHODS 165 5.0 Introduction 165 5.1 The Elements of Counting 167 5.2 Elements of Probability Theory 170 5.3 Applications of the Stochastic Methodology 182 5.4 Bayes’ Theorem 184 5.5 Epilogue and References 190 5.6 Exercises 191

6 CONTROL AND IMPLEMENTATION OF STATE SPACE SEARCH 193 6.0 Introduction 193 6.1 Recursion-Based Search 194 6.2 Production Systems 200 6.3 The Blackboard Architecture for Problem Solving 187 6.4 Epilogue and References 219 6.5 Exercises 220

PART III CAPTURING INTELLIGENCE: THE AI CHALLENGE 223

7 KNOWLEDGE REPRESENTATION 227 7.0 Issues in Knowledge Representation 227 7.1 A Brief History of AI Representational Systems 228

CONTENTS xxi

7.2 Conceptual Graphs: A Network Language 248 7.3 Alternative Representations and Ontologies 258 7.4 Agent Based and Distributed Problem Solving 265 7.5 Epilogue and References 270 7.6 Exercises 273

8 STRONG METHOD PROBLEM SOLVING 277 8.0 Introduction 277 8.1 Overview of Expert System Technology 279 8.2 Rule-Based Expert Systems 286 8.3 Model-Based, Case Based, and Hybrid Systems 298 8.4 Planning 314 8.5 Epilogue and References 329 8.6 Exercises 331

9 REASONING IN UNCERTAIN SITUATIONS 333 9.0 Introduction 333 9.1 Logic-Based Abductive Inference 335 9.2 Abduction: Alternatives to Logic 350 9.3 The Stochastic Approach to Uncertainty 363 9.4 Epilogue and References 378 9.5 Exercises 380

PART IV MACHINE LEARNING 385

10 MACHINE LEARNING: SYMBOL-BASED 387 10.0 Introduction 387 10.1 A Framework for Symbol-based Learning 390 10.2 Version Space Search 396 10.3 The ID3 Decision Tree Induction Algorithm 408 10.4 Inductive Bias and Learnability 417 10.5 Knowledge and Learning 422 10.6 Unsupervised Learning 433 10.7 Reinforcement Learning 442 10.8 Epilogue and References 449 10.9 Exercises 450

11 MACHINE LEARNING: CONNECTIONIST 453 11.0 Introduction 453 11.1 Foundations for Connectionist Networks 455 11.2 Perceptron Learning 458 11.3 Backpropagation Learning 467 11.4 Competitive Learning 474

xxii CONTENTS

11.5 Hebbian Coincidence Learning 484 11.6 Attractor Networks or “Memories” 495 11.7 Epilogue and References 505 11.8 Exercises 506

12 MACHINE LEARNING: GENETIC AND EMERGENT 507 12.0 Genetic and Emergent Models of Learning 507 12.1 The Genetic Algorithm 509 12.2 Classifier Systems and Genetic Programming 519 12.3 Artificial Life and Society-Based Learning 530 12.4 Epilogue and References 541 12.5 Exercises 542

13 MACHINE LEARNING: PROBABILISTIC 543 13.0 Stochastic and Dynamic Models of Learning 543 13.1 Hidden Markov Models (HMMs) 544 13.2 Dynamic Bayesian Networks and Learning 554 13.3 Stochastic Extensions to Reinforcement Learning 564 13.4 Epilogue and References 568 13.5 Exercises 570

PART V ADVANCED TOPICS FOR AI PROBLEM SOLVING 573

14 AUTOMATED REASONING 575 14.0 Introduction to Weak Methods in Theorem Proving 575 14.1 The General Problem Solver and Difference Tables 576 14.2 Resolution Theorem Proving 582 14.3 PROLOG and Automated Reasoning 603 14.4 Further Issues in Automated Reasoning 609 14.5 Epilogue and References 666 14.6 Exercises 667

15 UNDERSTANDING NATURAL LANGUAGE 619 15.0 The Natural Language Understanding Problem 619 15.1 Deconstructing Language: An Analysis 622 15.2 Syntax 625 15.3 Transition Network Parsers and Semantics 633 15.4 Stochastic Tools for Language Understanding 649 15.5 Natural Language Applications 658 15.6 Epilogue and References 630 15.7 Exercises 632

CONTENTS xxiii

PART VI EPILOGUE 671

16 ARTIFICIAL INTELLIGENCE AS EMPIRICAL ENQUIRY 673 16.0 Introduction 673 16.1 Artificial Intelligence: A Revised Definition 675 16.2 The Science of Intelligent Systems 688 16.3 AI: Current Challanges and Future Direstions 698 16.4 Epilogue and References 703

Bibliography 705 Author Index 735 Subject Index 743

1

P A R T I

ARTIFICIAL INTELLIGENCE: ITS ROOTS AND SCOPE

Everything must have a beginning, to speak in Sanchean phrase; and that beginning must be linked to something that went before. Hindus give the world an elephant to support it, but they make the elephant stand upon a tortoise. Invention, it must be humbly admitted, does not consist in creating out of void, but out of chaos; the materials must, in the first place, be afforded. . . .

—MARY SHELLEY, Frankenstein

Artificial Intelligence: An Attempted Definition

Artificial intelligence (AI) may be defined as the branch of computer science that is concerned with the automation of intelligent behavior. This definition is particularly appropriate to this book in that it emphasizes our conviction that AI is a part of computer science and, as such, must be based on sound theoretical and applied principles of that field. These principles include the data structures used in knowledge representation, the algorithms needed to apply that knowledge, and the languages and programming tech- niques used in their implementation.

However, this definition suffers from the fact that intelligence itself is not very well defined or understood. Although most of us are certain that we know intelligent behavior when we see it, it is doubtful that anyone could come close to defining intelligence in a way that would be specific enough to help in the evaluation of a supposedly intelligent computer program, while still capturing the vitality and complexity of the human mind.

As a result of the daunting task of building a general intelligence, AI researchers often assume the roles of engineers fashioning particular intelligent artifacts. These often come in the form of diagnostic, prognostic, or visualization tools that enable their human users to perform complex tasks. Examples of these tools include hidden Markov models for language understanding, automated reasoning systems for proving new theorems in math- ematics, dynammic Bayesian networks for tracking signals across cortical networks, and visualization of patterns of gene expression data, as seen in the applications of Section 1.2.

2 PART I / ARTIFICIAL INTELLIGENCE: ITS ROOTS AND SCOPE

The problem of defining the full field of artificial intelligence becomes one of defining intelligence itself: is intelligence a single faculty, or is it just a name for a collection of dis- tinct and unrelated abilities? To what extent is intelligence learned as opposed to having an a priori existence? Exactly what does happen when learning occurs? What is creativity? What is intuition? Can intelligence be inferred from observable behavior, or does it require evidence of a particular internal mechanism? How is knowledge represented in the nerve tissue of a living being, and what lessons does this have for the design of intelligent machines? What is self-awareness; what role does it play in intelligence? Furthermore, is it necessary to pattern an intelligent computer program after what is known about human intelligence, or is a strict “engineering” approach to the problem sufficient? Is it even pos- sible to achieve intelligence on a computer, or does an intelligent entity require the rich- ness of sensation and experience that might be found only in a biological existence?

These are unanswered questions, and all of them have helped to shape the problems and solution methodologies that constitute the core of modern AI. In fact, part of the appeal of artificial intelligence is that it offers a unique and powerful tool for exploring exactly these questions. AI offers a medium and a test-bed for theories of intelligence: such theories may be stated in the language of computer programs and consequently tested and verified through the execution of these programs on an actual computer.

For these reasons, our initial definition of artificial intelligence falls short of unambig- uously defining the field. If anything, it has only led to further questions and the paradoxi- cal notion of a field of study whose major goals include its own definition. But this difficulty in arriving at a precise definition of AI is entirely appropriate. Artificial intelli- gence is still a young discipline, and its structure, concerns, and methods are less clearly defined than those of a more mature science such as physics.

Artificial intelligence has always been more concerned with expanding the capabili- ties of computer science than with defining its limits. Keeping this exploration grounded in sound theoretical principles is one of the challenges facing AI researchers in general and this book in particular.

Because of its scope and ambition, artificial intelligence defies simple definition. For the time being, we will simply define it as the collection of problems and methodologies studied by artificial intelligence researchers. This definition may seem silly and meaning- less, but it makes an important point: artificial intelligence, like every science, is a human endeavor, and perhaps, is best understood in that context.

There are reasons that any science, AI included, concerns itself with a certain set of problems and develops a particular body of techniques for approaching these problems. In Chapter 1, a short history of artificial intelligence and the people and assumptions that have shaped it will explain why certain sets of questions have come to dominate the field and why the methods discussed in this book have been taken for their solution.

3

AI: EARLY HISTORY AND APPLICATIONS

All men by nature desire to know...

—ARISTOTLE, Opening sentence of the Metaphysics

Hear the rest, and you will marvel even more at the crafts and resources I have contrived. Greatest was this: in the former times if a man fell sick he had no defense against the sickness, neither healing food nor drink, nor unguent; but through the lack of drugs men wasted away, until I showed them the blending of mild simples wherewith they drive out all manner of diseases. . . .

It was I who made visible to men’s eyes the flaming signs of the sky that were before dim. So much for these. Beneath the earth, man’s hidden blessing, copper, iron, silver, and gold—will anyone claim to have discovered these before I did? No one, I am very sure, who wants to speak truly and to the purpose. One brief word will tell the whole story: all arts that mortals have come from Prometheus.

—AESCHYLUS, Prometheus Bound

1.1 From Eden to ENIAC: Attitudes toward Intelligence, Knowledge, and Human Artifice

Prometheus speaks of the fruits of his transgression against the gods of Olympus: his purpose was not merely to steal fire for the human race but also to enlighten humanity through the gift of intelligence or nous: the rational mind. This intelligence forms the foundation for all of human technology and ultimately all human civilization. The work of Aeschylus, the classical Greek dramatist, illustrates a deep and ancient awareness of the extraordinary power of knowledge. Artificial intelligence, in its very direct concern for Prometheus’s gift, has been applied to all the areas of his legacy—medicine, psychology, biology, astronomy, geology—and many areas of scientific endeavor that Aeschylus could not have imagined.

1

4 PART I / ARTIFICIAL INTELLIGENCE: ITS ROOTS AND SCOPE

Though Prometheus’s action freed humanity from the sickness of ignorance, it also earned him the wrath of Zeus. Outraged over this theft of knowledge that previously belonged only to the gods of Olympus, Zeus commanded that Prometheus be chained to a barren rock to suffer the ravages of the elements for eternity. The notion that human efforts to gain knowledge constitute a transgression against the laws of God or nature is deeply ingrained in Western thought. It is the basis of the story of Eden and appears in the work of Dante and Milton. Both Shakespeare and the ancient Greek tragedians portrayed intellectual ambition as the cause of disaster. The belief that the desire for knowledge must ultimately lead to disaster has persisted throughout history, enduring the Renaissance, the Age of Enlightenment, and even the scientific and philosophical advances of the nine- teenth and twentieth centuries. Thus, we should not be surprised that artificial intelligence inspires so much controversy in both academic and popular circles.

Indeed, rather than dispelling this ancient fear of the consequences of intellectual ambition, modern technology has only made those consequences seem likely, even imminent. The legends of Prometheus, Eve, and Faustus have been retold in the language of technological society. In her introduction to Frankenstein, subtitled, interestingly enough, The Modern Prometheus, Mary Shelley writes:

Many and long were the conversations between Lord Byron and Shelley to which I was a devout and silent listener. During one of these, various philosophical doctrines were discussed, and among others the nature of the principle of life, and whether there was any probability of its ever being discovered and communicated. They talked of the experiments of Dr. Darwin (I speak not of what the doctor really did or said that he did, but, as more to my purpose, of what was then spoken of as having been done by him), who preserved a piece of vermicelli in a glass case till by some extraordinary means it began to move with a voluntary motion. Not thus, after all, would life be given. Perhaps a corpse would be reanimated; galvanism had given token of such things: perhaps the component parts of a creature might be manufactured, brought together, and endued with vital warmth (Butler 1998).

Mary Shelley shows us the extent to which scientific advances such as the work of Darwin and the discovery of electricity had convinced even nonscientists that the work- ings of nature were not divine secrets, but could be broken down and understood system- atically. Frankenstein’s monster is not the product of shamanistic incantations or unspeakable transactions with the underworld: it is assembled from separately “manufac- tured” components and infused with the vital force of electricity. Although nineteenth-cen- tury science was inadequate to realize the goal of understanding and creating a fully intelligent agent, it affirmed the notion that the mysteries of life and intellect might be brought into the light of scientific analysis.

1.1.1 A Brief History of the Foundations for AI

By the time Mary Shelley finally and perhaps irrevocably joined modern science with the Promethean myth, the philosophical foundations of modern work in artificial intelligence had been developing for several thousand years. Although the moral and cultural issues raised by artificial intelligence are both interesting and important, our introduction is more

CHAPTER 1 / AI: HISTORY AND APPLICATIONS 5

properly concerned with AI’s intellectual heritage. The logical starting point for such a history is the genius of Aristotle, or as Dante in the Divine Comedy refers to him, “the master of them that know”. Aristotle wove together the insights, wonders, and fears of the early Greek tradition with the careful analysis and disciplined thought that were to become the standard for more modern science.

For Aristotle, the most fascinating aspect of nature was change. In his Physics, he defined his “philosophy of nature” as the “study of things that change”. He distinguished between the matter and form of things: a sculpture is fashioned from the material bronze and has the form of a human. Change occurs when the bronze is molded to a new form. The matter/form distinction provides a philosophical basis for modern notions such as symbolic computing and data abstraction. In computing (even with numbers) we are manipulating patterns that are the forms of electromagnetic material, with the changes of form of this material representing aspects of the solution process. Abstracting the form from the medium of its representation not only allows these forms to be manipulated com- putationally but also provides the promise of a theory of data structures, the heart of mod- ern computer science. It also supports the creation of an “artificial” intelligence.

In his Metaphysics, beginning with the words “All men by nature desire to know”, Aristotle developed a science of things that never change, including his cosmology and theology. More relevant to artificial intelligence, however, was Aristotle’s epistemology or analysis of how humans “know” their world, discussed in his Logic. Aristotle referred to logic as the “instrument” (organon), because he felt that the study of thought itself was at the basis of all knowledge. In his Logic, he investigated whether certain propositions can be said to be “true” because they are related to other things that are known to be “true”. Thus if we know that “all men are mortal” and that “Socrates is a man”, then we can con- clude that “Socrates is mortal”. This argument is an example of what Aristotle referred to as a syllogism using the deductive form modus ponens. Although the formal axiomatiza- tion of reasoning needed another two thousand years for its full flowering in the works of Gottlob Frege, Bertrand Russell, Kurt Gödel, Alan Turing, Alfred Tarski, and others, its roots may be traced to Aristotle.

Renaissance thought, building on the Greek tradition, initiated the evolution of a dif- ferent and powerful way of thinking about humanity and its relation to the natural world. Science began to replace mysticism as a means of understanding nature. Clocks and, even- tually, factory schedules superseded the rhythms of nature for thousands of city dwellers. Most of the modern social and physical sciences found their origin in the notion that pro- cesses, whether natural or artificial, could be mathematically analyzed and understood. In particular, scientists and philosophers realized that thought itself, the way that knowledge was represented and manipulated in the human mind, was a difficult but essential subject for scientific study.

Perhaps the major event in the development of the modern world view was the Copernican revolution, the replacement of the ancient Earth-centered model of the universe with the idea that the Earth and other planets are actually in orbits around the sun. After centuries of an “obvious” order, in which the scientific explanation of the nature of the cosmos was consistent with the teachings of religion and common sense, a drastically different and not at all obvious model was proposed to explain the motions of heavenly bodies. For perhaps the first time, our ideas about the world were seen as fundamentally

6 PART I / ARTIFICIAL INTELLIGENCE: ITS ROOTS AND SCOPE

distinct from that world’s appearance. This split between the human mind and its sur- rounding reality, between ideas about things and things themselves, is essential to the modern study of the mind and its organization. This breach was widened by the writings of Galileo, whose scientific observations further contradicted the “obvious” truths about the natural world and whose development of mathematics as a tool for describing that world emphasized the distinction between the world and our ideas about it. It is out of this breach that the modern notion of the mind evolved: introspection became a common motif in literature, philosophers began to study epistemology and mathematics, and the system- atic application of the scientific method rivaled the senses as tools for understanding the world.

In 1620, Francis Bacon’s Novum Organun offered a set of search techniques for this emerging scientific methodology. Based on the Aristotelian and Platonic idea that the “form” of an entity was equivalent to the sum of its necessary and sufficient “features”, Bacon articulated an algorithm for determining the essence of an entity. First, he made an organized collection of all instances of the entity, enumerating the features of each in a table. Then he collected a similar list of negative instances of the entity, focusing espe- cially on near instances of the entity, that is, those that deviated from the “form” of the entity by single features. Then Bacon attempts - this step is not totally clear - to make a systematic list of all the features essential to the entity, that is, those that are common to all positive instances of the entity and missing from the negative instances.

It is interesting to see a form of Francis Bacon’s approach to concept learning reflected in modern AI algorithms for Version Space Search, Chapter 10.2. An extension of Bacon’s algorithms was also part of an AI program for discovery learning, suitably called Bacon (Langley et al. 1981). This program was able to induce many physical laws from collections of data related to the phenomena. It is also interesting to note that the question of whether a general purpose algorithm was possible for producing scientific proofs awaited the challenges of the early twentieth century mathematician Hilbert (his Entscheidungsproblem) and the response of the modern genius of Alan Turing (his Turing Machine and proofs of computability and the halting problem); see Davis et al. (1976).

Although the first calculating machine, the abacus, was created by the Chinese in the twenty-sixth century BC, further mechanization of algebraic processes awaited the skills of the seventeenth century Europeans. In 1614, the Scots mathematician, John Napier, cre- ated logarithms, the mathematical transformations that allowed multiplication and the use of exponents to be reduced to addition and multiplication. Napier also created his bones that were used to represent overflow values for arithmetic operations. These bones were later used by Wilhelm Schickard (1592-1635), a German mathematician and clergyman of Tübingen, who in 1623 invented a Calculating Clock for performing addition and subtrac- tion. This machine recorded the overflow from its calculations by the chiming of a clock.

Another famous calculating machine was the Pascaline that Blaise Pascal, the French philosopher and mathematician, created in 1642. Although the mechanisms of Schickard and Pascal were limited to addition and subtraction - including carries and borrows - they showed that processes that previously were thought to require human thought and skill could be fully automated. As Pascal later stated in his Pensees (1670), “The arithmetical machine produces effects which approach nearer to thought than all the actions of animals”.

CHAPTER 1 / AI: HISTORY AND APPLICATIONS 7

Pascal’s successes with calculating machines inspired Gottfried Wilhelm von Leibniz in 1694 to complete a working machine that become known as the Leibniz Wheel. It inte- grated a moveable carriage and hand crank to drive wheels and cylinders that performed the more complex operations of multiplication and division. Leibniz was also fascinated by the possibility of a automated logic for proofs of propositions. Returning to Bacon’s entity specification algorithm, where concepts were characterized as the collection of their necessary and sufficient features, Liebniz conjectured a machine that could calculate with these features to produce logically correct conclusions. Liebniz (1887) also envisioned a machine, reflecting modern ideas of deductive inference and proof, by which the produc- tion of scientific knowledge could become automated, a calculus for reasoning.

The seventeenth and eighteenth centuries also saw a great deal of discussion of episte- mological issues; perhaps the most influential was the work of René Descartes, a central figure in the development of the modern concepts of thought and theories of mind. In his Meditations, Descartes (1680) attempted to find a basis for reality purely through intro- spection. Systematically rejecting the input of his senses as untrustworthy, Descartes was forced to doubt even the existence of the physical world and was left with only the reality of thought; even his own existence had to be justified in terms of thought: “Cogito ergo sum” (I think, therefore I am). After he established his own existence purely as a thinking entity, Descartes inferred the existence of God as an essential creator and ultimately reas- serted the reality of the physical universe as the necessary creation of a benign God.

We can make two observations here: first, the schism between the mind and the phys- ical world had become so complete that the process of thinking could be discussed in iso- lation from any specific sensory input or worldly subject matter; second, the connection between mind and the physical world was so tenuous that it required the intervention of a benign God to support reliable knowledge of the physical world! This view of the duality between the mind and the physical world underlies all of Descartes’s thought, including his development of analytic geometry. How else could he have unified such a seemingly worldly branch of mathematics as geometry with such an abstract mathematical frame- work as algebra?

Why have we included this mind/body discussion in a book on artificial intelligence? There are two consequences of this analysis essential to the AI enterprise:

1. By attempting to separate the mind from the physical world, Descartes and related thinkers established that the structure of ideas about the world was not necessar- ily the same as the structure of their subject matter. This underlies the methodol- ogy of AI, along with the fields of epistemology, psychology, much of higher mathematics, and most of modern literature: mental processes have an existence of their own, obey their own laws, and can be studied in and of themselves.

2. Once the mind and the body are separated, philosophers found it necessary to find a way to reconnect the two, because interaction between Descartes mental, res cogitans, and physical, res extensa, is essential for human existence.

Although millions of words have been written on this mind–body problem, and numerous solutions proposed, no one has successfully explained the obvious interactions between mental states and physical actions while affirming a fundamental difference

8 PART I / ARTIFICIAL INTELLIGENCE: ITS ROOTS AND SCOPE

between them. The most widely accepted response to this problem, and the one that provides an essential foundation for the study of AI, holds that the mind and the body are not fundamentally different entities at all. On this view, mental processes are indeed achieved by physical systems such as brains (or computers). Mental processes, like physi- cal processes, can ultimately be characterized through formal mathematics. Or, as acknowledged in his Leviathan by the 17th century English philosopher Thomas Hobbes (1651), “By ratiocination, I mean computation”.

1.1.2 AI and the Rationalist and Empiricist Traditions

Modern research issues in artificial intelligence, as in other scientific disciplines, are formed and evolve through a combination of historical, social, and cultural pressures. Two of the most prominent pressures for the evolution of AI are the empiricist and rationalist traditions in philosophy.

The rationalist tradition, as seen in the previous section, had an early proponent in Plato, and was continued on through the writings of Pascal, Descartes, and Liebniz. For the rationalist, the external world is reconstructed through the clear and distinct ideas of a mathematics. A criticism of this dualistic approach is the forced disengagement of repre- sentational systems from their field of reference. The issue is whether the meaning attrib- uted to a representation can be defined independent of its application conditions. If the world is different from our beliefs about the world, can our created concepts and symbols still have meaning?

Many AI programs have very much of this rationalist flavor. Early robot planners, for example, would describe their application domain or “world” as sets of predicate calculus statements and then a “plan” for action would be created through proving theorems about this “world” (Fikes et al. 1972, see also Section 8.4). Newell and Simon’s Physical Symbol System Hypothesis (Introduction to Part II and Chapter 16) is seen by many as the arche- type of this approach in modern AI. Several critics have commented on this rationalist bias as part of the failure of AI at solving complex tasks such as understanding human lan- guages (Searle 1980, Winograd and Flores 1986, Brooks 1991a).

Rather than affirming as “real” the world of clear and distinct ideas, empiricists con- tinue to remind us that “nothing enters the mind except through the senses”. This con- straint leads to further questions of how the human can possibly perceive general concepts or the pure forms of Plato’s cave (Plato 1961). Aristotle was an early empiricist, emphasiz- ing in his De Anima, the limitations of the human perceptual system. More modern empir- icists, especially Hobbes, Locke, and Hume, emphasize that knowledge must be explained through an introspective but empirical psychology. They distinguish two types of mental phenomena perceptions on one hand and thought, memory, and imagination on the other. The Scots philosopher, David Hume, for example, distinguishes between impressions and ideas. Impressions are lively and vivid, reflecting the presence and existence of an exter- nal object and not subject to voluntary control, the qualia of Dennett (2005). Ideas on the other hand, are less vivid and detailed and more subject to the subject’s voluntary control.

Given this distinction between impressions and ideas, how can knowledge arise? For Hobbes, Locke, and Hume the fundamental explanatory mechanism is association.

CHAPTER 1 / AI: HISTORY AND APPLICATIONS 9

Particular perceptual properties are associated through repeated experience. This repeated association creates a disposition in the mind to associate the corresponding ideas, a pre- curser of the behaviorist approach of the twentieth century. A fundamental property of this account is presented with Hume’s skepticism. Hume’s purely descriptive account of the origins of ideas cannot, he claims, support belief in causality. Even the use of logic and induction cannot be rationally supported in this radical empiricist epistemology.

In An Inquiry Concerning Human Understanding (1748), Hume’s skepticism extended to the analysis of miracles. Although Hume didn’t address the nature of miracles directly, he did question the testimony-based belief in the miraculous. This skepticism, of course, was seen as a direct threat by believers in the bible as well as many other purvey- ors of religious traditions. The Reverend Thomas Bayes was both a mathematician and a minister. One of his papers, called Essay towards Solving a Problem in the Doctrine of Chances (1763) addressed Hume’s questions mathematically. Bayes’ theorem demon- strates formally how, through learning the correlations of the effects of actions, we can determine the probability of their causes.

The associational account of knowledge plays a significant role in the development of AI representational structures and programs, for example, in memory organization with semantic networks and MOPS and work in natural language understanding (see Sections 7.0, 7.1, and Chapter 15). Associational accounts have important influences of machine learning, especially with connectionist networks (see Section 10.6, 10.7, and Chapter 11). Associationism also plays an important role in cognitive psychology including the sche- mas of Bartlett and Piaget as well as the entire thrust of the behaviorist tradition (Luger 1994). Finally, with AI tools for stochastic analysis, including the Bayesian belief network (BBN) and its current extensions to first-order Turing-complete systems for stochastic modeling, associational theories have found a sound mathematical basis and mature expressive power. Bayesian tools are important for research including diagnostics, machine learning, and natural language understanding (see Chapters 5 and 13).

Immanuel Kant, a German philosopher trained in the rationalist tradition, was strongly influenced by the writing of Hume. As a result, he began the modern synthesis of these two traditions. Knowledge for Kant contains two collaborating energies, an a priori component coming from the subject’s reason along with an a posteriori component com- ing from active experience. Experience is meaningful only through the contribution of the subject. Without an active organizing form proposed by the subject, the world would be nothing more than passing transitory sensations. Finally, at the level of judgement, Kant claims, passing images or representations are bound together by the active subject and taken as the diverse appearances of an identity, of an “object”. Kant’s realism began the modern enterprise of psychologists such as Bartlett, Brunner, and Piaget. Kant’s work influences the modern AI enterprise of machine learning (Section IV) as well as the con- tinuing development of a constructivist epistemology (see Chapter 16).

1.1.3 The Development of Formal Logic

Once thinking had come to be regarded as a form of computation, its formalization and eventual mechanization were obvious next steps. As noted in Section 1.1.1,

10 PART I / ARTIFICIAL INTELLIGENCE: ITS ROOTS AND SCOPE

Gottfried Wilhelm von Leibniz, with his Calculus Philosophicus, introduced the first sys- tem of formal logic as well as proposed a machine for automating its tasks (Leibniz 1887). Furthermore, the steps and stages of this mechanical solution can be represented as move- ment through the states of a tree or graph. Leonhard Euler, in the eighteenth century, with his analysis of the “connectedness” of the bridges joining the riverbanks and islands of the city of Königsberg (see the introduction to Chapter 3), introduced the study of representa- tions that can abstractly capture the structure of relationships in the world as well as the discrete steps within a computation about these relationships (Euler 1735).

The formalization of graph theory also afforded the possibility of state space search, a major conceptual tool of artificial intelligence. We can use graphs to model the deeper structure of a problem. The nodes of a state space graph represent possible stages of a problem solution; the arcs of the graph represent inferences, moves in a game, or other steps in a problem solution. Solving the problem is a process of searching the state space graph for a path to a solution (Introduction to II and Chapter 3). By describing the entire space of problem solutions, state space graphs provide a powerful tool for measuring the structure and complexity of problems and analyzing the efficiency, correctness, and gener- ality of solution strategies.

As one of the originators of the science of operations research, as well as the designer of the first programmable mechanical computing machines, Charles Babbage, a nine- teenth century mathematician, may also be considered an early practitioner of artificial intelligence (Morrison and Morrison 1961). Babbage’s difference engine was a special- purpose machine for computing the values of certain polynomial functions and was the forerunner of his analytical engine. The analytical engine, designed but not successfully constructed during his lifetime, was a general-purpose programmable computing machine that presaged many of the architectural assumptions underlying the modern computer.

In describing the analytical engine, Ada Lovelace (1961), Babbage’s friend, sup- porter, and collaborator, said:

We may say most aptly that the Analytical Engine weaves algebraical patterns just as the Jac- quard loom weaves flowers and leaves. Here, it seems to us, resides much more of originality than the difference engine can be fairly entitled to claim.

Babbage’s inspiration was his desire to apply the technology of his day to liberate humans from the drudgery of making arithmetic calculations. In this sentiment, as well as with his conception of computers as mechanical devices, Babbage was thinking in purely nineteenth century terms. His analytical engine, however, also included many modern notions, such as the separation of memory and processor, the store and the mill in Bab- bage’s terms, the concept of a digital rather than analog machine, and programmability based on the execution of a series of operations encoded on punched pasteboard cards. The most striking feature of Ada Lovelace’s description, and of Babbage’s work in gen- eral, is its treatment of the “patterns” of algebraic relationships as entities that may be studied, characterized, and finally implemented and manipulated mechanically without concern for the particular values that are finally passed through the mill of the calculating machine. This is an example implementation of the “abstraction and manipulation of form” first described by Aristotle and Liebniz.

CHAPTER 1 / AI: HISTORY AND APPLICATIONS 11

The goal of creating a formal language for thought also appears in the work of George Boole, another nineteenth-century mathematician whose work must be included in any discussion of the roots of artificial intelligence (Boole 1847, 1854). Although he made contributions to a number of areas of mathematics, his best known work was in the mathematical formalization of the laws of logic, an accomplishment that forms the very heart of modern computer science. Though the role of Boolean algebra in the design of logic circuitry is well known, Boole’s own goals in developing his system seem closer to those of contemporary AI researchers. In the first chapter of An Investigation of the Laws of Thought, on which are founded the Mathematical Theories of Logic and Probabilities, Boole (1854) described his goals as

to investigate the fundamental laws of those operations of the mind by which reasoning is performed: to give expression to them in the symbolical language of a Calculus, and upon this foundation to establish the science of logic and instruct its method; …and finally to collect from the various elements of truth brought to view in the course of these inquiries some proba- ble intimations concerning the nature and constitution of the human mind.