Abduction, Reason, and Science

AI Magazine Volume 23 Number 1 (2002) (© AAAI)

Book Reviews

been conceptually challenging but also shows a (terminological) confusion Book Reviews with its close neighbor, induction. This book contributes to the rapidly growing literature on the topic of abductive reasoning. By placing abduc- Abduction, Reason, and tion at the heart of the foundations of AI from philosophical, cognitive, and computational perspectives, the au- Science: Processes of Discovery thor makes us aware that abduction is not at all a new topic of research. In and Explanation—A Review addition, by introducing fine distinctions in abductive kinds, it shows its relevance as a recent topic of research in all these fields. Atocha Aliseda The importance of abduction has been recognized by leading researchers in all relevant fields. Al- though for Jaakko Hintikka, abduction is the fundamental problem of roadly speaking, abduction is a his discovery does not work. His helio- contemporary epistemology, in which reasoning process invoked to ex- centric view allowed him to think that abductive inferences are assembled as Bplain a puzzling observation. the sun, so near to the center of the “answers to the inquirer’s explicit or There are however, a variety of differ- planetary system, and so large, must (usually) tacit question put to some ent approaches that claim to capture somehow cause the planets to move as definite source of answers (informa- the true nature of this concept. One they do. In addition to this strong con- tion)” (p. 129), for Herbert Simon, the reason for this diversity lies in the fact jecture, he also had to generalize his nature of the retroductive process (an- that abductive reasoning occurs in a findings for Mars to all planets by as- other term for abduction) “is the main multitude of contexts. It concerns cases suming that the same physical condi- subject of the theory of problem solv- that cover the simplest selection of al- tions could be obtained throughout ing” (p. 16). For Paul Thagard, several ready existing hypotheses to the gener- the solar system. kinds of abduction play a key role as ation of new concepts in science. It also Research on abduction in AI dates heuristic strategies in the program PI (for “processes of induction”), a work- concerns cases where the observation is back to the 1970s, but it is only fairly puzzling because it is novel versus cases ing system devoted to explaining in in which the surprise concerns an computational terms the main problems of philosophy of science, such as anomalous observation. For example, if Abduction, Reason, and Science: Pro- we wake up, and the lawn is wet, we cesses of Discovery and Explanation, scientific discovery, explanation, and might explain this observation by as- Lorenzo Magnani, New York, evaluation (p. 49). suming that it must have rained or that Kluwer Academic/Plenum Pub- In agreement with other current ap- the sprinklers have been on. This is a lishers, 2001, 205 pages, ISBN 0- proaches to abduction, for the author practical setting found in our day-to- 306-46514-0 (hardback). of this book, there are two main episte- day commonsense reasoning when a mological meanings of the word ab- novel phenomenon is needed for an duction: (1) abduction that only gener- recently that it has attracted great in- explanation. Abduction also occurs in ates plausible hypotheses and (2) ab- terest in areas such as logic program- more theoretical scientific contexts. For duction as inference to the best ming, knowledge assimilation, and di- example, it has been claimed (Hanson explanation that also evaluates them 1961; Peirce 1958) that Johannes Ke- agnosis. It has been a topic of several to further obtain the best one. In this pler’s great discovery that the orbit of workshops at AI conferences (1996, book, the first meaning is further divid- the planets is elliptical rather than cir- 1998, 2000 European Conference on ed into selective or creative. Selection cular was a prime piece of abductive Artificial Intelligence; 1997 Interna- takes place in contexts such as medical reasoning. What initially led to this dis- tional Joint Conference on Artificial diagnosis, in which the task is to select covery was his anomalous observation Intelligence) as well as model-based a diagnosis from a precompiled set of that the longitudes of Mars did not fit reasoning conferences (1998, 2001 diagnostic entities. Creativity is present circular orbits. Moreover, before even Model-Based Reasoning Conference). in issues such as the discovery of a new dreaming that the best explanation in- It has also been at the center of recent disease. The latter meaning, abduction volved ellipses instead of circles, he publications (Flach and Kakas 2000; as inference to the best explanation, is tried several other forms. Kepler had to Josephson and Josephson 1994). In all described by the complete abduction- make some other assumptions about these places, the discussion about the deduction-induction cycle, represented the planetary system, without which different aspects of abduction has in an epistemological model for diag-

Copyright © 2002, American Association for Artificial Intelligence. All rights reserved. 0738-4602-2002 / $2.00 SPRING 2002 113 Book Reviews nostic reasoning (ST-MODEL) in which se- necessarily sentential or formal). of the chapters are dedicated to the lective abduction introduces a set of This book presents as well applica- analysis of the role of inconsistencies plausible hypotheses, followed by de- tions of the several distinctions of ab- in scientific discovery (chapter 6) and duction to explore their consequences duction introduced earlier. For exam- the makeup of hypotheses withdrawal and induction to test them, to either ple, the case of medical diagnostic rea- in science (chapter 7). increase their likelihood or refute all soning is described as an instance of Abduction, Reason, and Science is a but one. Even though the selection and theoretical abduction, showing that its book for those interested in the sub- evaluation phases of abduction are in machinery fits the epistemological ject of discovery who are willing to get fact integrated in cognitive models of model mentioned earlier; the patient an integrated picture from philoso- human thought, it is nevertheless a data are abstracted and used to select phy, cognitive science, and computer useful methodological distinction for hypotheses, and these hypothetical so- science, all disciplines concerned with the development of real artificial rea- lutions in turn provide starting condi- the question of creative reasoning. A soning systems. tions for the forecast of expected conse- beginner can get a pretty good idea of Moreover, Magnani proposes an in- quences, which are later compared to the recent status of research on abduc- teresting distinction between theoret- the patient’s data to evaluate (corrobo- tion, and at the same time, it pleases ical abduction and manipulative ab- rate or eliminate) those hypotheses the expert by informing him/her duction. Although theoretical abduc- that come from deduction. A particular about approaches in other fields and tion “is the process of reasoning in application is the system NEOANEMIA, a offering a detailed and interesting no- which explanatory hypotheses are working diagnostic system for disorders tice about the original views of its au- formed and evaluated” (p. 18), and such as anemia developed at the home thor, Lorenzo Magnani. can, in turn, be sentential or model university of the author (Pavia). Visual I found it interesting that the ST-MOD- based (to be explained later), manipu- and temporal aspects of abduction are EL resembles the American pragmatist lative abduction is action based, the also taken into account and presented Charles S. Peirce’s complete framework case in which “the suggested hypothe- as cases of model-based creative abduc- of abduction, which considers not only ses are inherently ambiguous until ar- tion. This perspective highlights the its inferential logical structure (which is ticulated into configurations of real or importance of both spatial reasoning what most approaches take) but also “imaginated” [SIC] entities (p. 54). and anomaly resolution for hypotheses two further aspects, namely, (1) testing The interplay of these two aspects generation and scientific discovery. The and (2) economy. I think that it has be- “consists of a superimposition of inter- concrete example is the discovery of come quite clear by now—and this nal and external, where the elements non-Euclidean geometries, about book makes a strong point in this direc- of the external structures gain new which it is argued that some of the hy- tion —that the study of the role of test- meanings and relationships to one an- potheses created by Lobachevsky were ing in abduction is fundamental for its other, thanks to the constructive ex- indeed image based, something that understanding. planatory theoretical activity” (p. 59). helped to deal with the fifth parallel References Magnani offers an impressive and postulate by manipulation of symbols. comprehensive overview of the logical Interestingly, the author claims that Flach, P., and Kakas, A., eds. 2000. Abduc- approaches to abduction, covering the the anomalous (or problematic) aspect tive and Inductive Reasoning: Essays on Their Relation and Integration. Applied Logic Se- deductive model, the abductive logic of the Euclidean fifth postulate lies in ries. New York: Kluwer Academic. programming paradigm, and ap- that “we cannot draw or ‘imagine’ the Hanson, N. R. 1961. Patterns of Scientific proaches linking abduction to the two lines at infinity” (p. 165), which, in Discovery. Cambridge, U.K.: Cambridge well-known framework of belief revi- contrast to the rest of the postulates, is University Press. sion, all of which are considered as empirically unverifiable and, thus, Josephson, J., and Josephson, S., eds. 1994. cases of theoretical sentential abduc- opens the door to the possibility of cre- Abductive Inference: Computation, Philosophy, tion, given their logical character. His ating alternative geometries to that of Technology. New York: Cambridge Universi- conclusion is that all these logical ap- Euclid. ty Press. proaches deal primarily with the selec- To summarize, in the first three Peirce, C. S. 1958. Collected Papers of Charles tive and explanatory aspects of abduc- chapters of this book, certain frame- Sanders Peirce, Volumes 1–6, eds. C. tion, leaving aside many other cre- works for the various facets of abduc- Hartshorne and P. Weiss. Volumes 7–8, ed. ative processes, such as conceptual tive reasoning are put forward, cover- A. W. Burke. Cambridge, Mass.: Harvard change, an essential ingredient to ing the selection of explanations as University Press. study revolutionary changes in sci- well as their evaluation from the stand- Atocha Aliseda ([email protected]. ence. Thus, “we have to consider a point of theoretical abduction (chapter mx) is a full professor at the Institute for broader inferential one (view) which 2) and manipulative abduction (chap- Philosophical Research of the National Au- encompasses both sentential and what ter 3). In the following chapters, vari- tonomous University of Mexico and "breedtestrategiepostdoc" at the Faculty of I call model-based sides of creative ab- ous applications are presented in the Philosophy in Groningen, The Netherlands. duction” (p. 36). (The term model- area of diagnosis (chapter 4) as well as She obtained her Ph.D. from Stanford Uni- based reasoning is used here to indicate in discovery in science with special versity in 1997. Her main research interests the construction and manipulation of emphasis on visual and temporal as- are abductive logic and the connection be- various kinds of representations, not pects of abduction (chapter 5). The rest tween philosophy of science and AI.

114 AI MAGAZINE Book Reviews

complicated. For example, a system The Logic of Knowledge Bases might know that Mary has a teacher without knowing the identity of such an individual. Explicitly distinguishing A Review what is true from what is known makes it possible for a system to deal correctly with these scenarios. Enrico Motta The introduction of the epistemic operator, K, requires the adoption of a possible world semantics for the interaction language (the language used by A knowledge-based system (KBS) con- a function of the task in hand; for ex- a reasoning engine to interact with a tains (by definition) an explicitly codi- ample, see Bylander and Chandra- knowledge base at the knowledge lev- fied body of knowledge, which causal- sekaran’s (1988) discussion on the in- el). In short, the abstract state of the ly determines its behavior. Hence, at a teraction hypothesis. As a result, they knowledge of an agent (that is, its epis- coarse-grained level of abstraction, KB- would argue, it is not possible to dis- temic state) can be characterized as the Ss can be characterized in terms of two cuss the knowledge of a system inde- collection of all possible worlds that components: (1) a knowledge base, en- pendently of the task context in which coding the knowledge embodied by are consistent with the knowledge the system is meant to operate. I won’t held by the agent. If the knowledge of the system, and (2) a reasoning engine, go into too many details here because which is able to query the knowledge the agent is complete, then the epis- a detailed discussion of the declarative temic state contains only one world. A base, infer or acquire knowledge from versus the procedural argument is well external sources, and add new knowl- nice feature of Levesque and Lakemey- beyond the scope of this review. The edge to the knowledge base. Levesque er’s treatment of epistemic logic is that important point to make is that and Lakemeyer’s The Logic of Knowledge in contrast to many other treatments Levesque and Lakemeyer’s approach is Bases deals with the “internal logic” of of modalities, the discussion is reason- situated in a precise AI research a KBS: It provides a formal account of ably easy to follow for people who are the interaction between a reasoning paradigm, which considers knowledge not experts in the field. This is the re- engine and a knowledge base. Clearly, bases as declaratively specified, task-in- sult of two main features of this analy- this analysis is not the same as provid- dependent representations of knowl- sis: First, the authors introduce some ing a formal account of the behavior of edge. simplifying assumptions, such as the a KBS as a whole. A knowledge-level ac- use of standard names to identify indi- count of a KBS (that is, a competence- The Logic of Knowledge Bases, Hector J. viduals in the universe of discourse, centered, implementation-indepen- Levesque and Gerhard Lakemeyer, that do not affect the substance and Cambridge, Massachusetts, The MIT dent description of a system), such as the general applicability of the pro- Press, 282 pp., $45.00, ISBN 0-262- posed models. Second, although the Clancey’s (1985) analysis of first-gener- 12232-4. ation rule-based systems, focuses on proposed language extends first-order logic with an epistemic operator, the task-centered competence of the Starting from such a declarativist Levesque and Lakemeyer succeed in system; that is, it addresses issues such standpoint, Levesque and Lakemeyer reconciling their analysis within a as what kind of problems the KBS is de- view a knowledge base as an epistemic standard first-order–logic framework. signed to tackle, what reasoning meth- agent, and they set out to specify for- ods it uses, and what knowledge it re- Thus, the reader is not forced into mally what knowledge can be attribut- quires. In contrast with task-centered learning a new syntax, and the under- ed to such a system. To talk about the analyses, Levesque and Lakemeyer fo- lying model theory is a “reasonably epistemic state of a knowledge base, cus on the competence of the knowl- conservative” extension of standard Levesque and Lakemeyer introduce an edge base rather than that of the whole model theory for first-order logic. More extra logical symbol, K, to be able to system. Hence, their notion of compe- importantly, the previous statement distinguish what is known from what tence is a task-independent one: It is can be given a strong interpretation is true. At this point, a reader might the “abstract state of knowledge” (p. because the representation theorem, 49) denoted by the contents (implicit wonder why the authors can’t simply discussed in chapter 7, shows that for or explicit) of a knowledge base at any stick to classical first-order logic and finite knowledge bases both Tell and particular time in its life cycle. This is describe what is true about the world? Ask operations (see later for more de- an interesting assumption, which the The reason, argue Levesque and Lake- tails on the Tell and Ask protocol) can “proceduralists” in the AI community meyer, has to do with incomplete always be realized using objective sen- might object to: According to the pro- knowledge. If a system has complete tences, that is, sentences expressed in cedural viewpoint of knowledge repre- knowledge, then of course there is no standard first-order logic. The ability to sentation, the knowledge modeled in difference between what is known and reduce formal treatments of modalities an application, its representation, and what is true: The two sets coincide. to standard first-order logic is an im- the associated knowledge-retrieval However, when a system has incom- portant result, given that standard mechanisms have to be engineered as plete knowledge, things become more first-order logic is far better understood

Copyright © 2002, American Association for Artificial Intelligence. All rights reserved. 0738-4602-2002 / $2.00 SPRING 2002 115 Book Reviews and computationally tractable than Robert Moore’s epistemic logic and by Bill Clancey, Mark Stefik, and Bal- modal logics. Having said so, one John McCarthy’s situation calculus. Of aknishman Chandrasekaran in the slightly confusing aspect of Levesque particular interest are chapters 12 and United States and by Bob Wielinga and and Lakemeyer’s analysis is that al- 13, which discuss a semantic approach others in Europe (Bylander and Chan- though Levesque and Lakemeyer talk to logical omniscience. A standard drasekaran 1980; Clancey 1985; Fensel about modeling knowledge, they are problem in knowledge representation et al. 1999; Stefik 1995; Schreiber et al. actually modeling beliefs. For those is the trade-off between expressivity 2000). Au contraire, the reader gets the used to “classical” theories of knowl- and computational efficiency. Most feeling that the authors are actually not edge, this is slightly confusing because solutions err either on the side of ex- that interested in applications of in Levesque and Lakemeyer’s treat- pressivity (that is, you might wait a knowledge representation technology: ment, an agent might know some- long time for an answer) or efficiency The examples in the book tend to be of thing that is not true, but in many oth- (that is, you are guaranteed an answer, the “Tweety is a bird” variety, and none er treatments, everything known must but there is a lot that you are prevent- of the 143 references seems to be relat- be true (that is, the real world is always ed from representing). Levesque and ed to some application. one of the possible worlds available to Lakemeyer show that it is possible to In conclusion, this is an excellent an epistemic agent). This is a bit dis- distinguish between explicit and im- book, which is very much grounded in concerting at first, but eventually the plicit beliefs by providing a four-val- the AI tradition of symbolic representa- reader gets used to it; that is, he/she ued semantics (in addition to standard tion of knowledge. Anybody interested eventually rests assured that it is true and false truth assignments, sen- in formal representations of knowledge Levesque and Lakemeyer’s epistemic tences can now be either true and false and epistemic agents should definitely agents who might be affected by solip- or neither true nor false). The distinc- read this text. However, those readers sism rather than he or she. tion between implicit and explicit be- who are primarily interested in knowl- The problem of how to handle for- liefs enables very efficient decision edge-based systems viewed as task-per- mally and effectively incomplete procedures for explicit beliefs in the forming agents should definitely note knowledge is one of the two main goals context of a very expressible language. that the word systems does not follow of the book. The other has to do with Thus, what is the general assessment the words knowledge base in the title of precisely characterizing the behavior of of this book, and what audience is the the book. This omission is indeed a sig- a knowledge base at the knowledge lev- book relevant to? The answer to the nificant one! el. The perspective of a knowledge-level first question is quite easy. This is clear- References view of intelligent systems was first ly a very good book, which provides a proposed by Allen Newell in 1981 and powerful, formal, and detailed (but rea- Bylander, T., and Chandrasekaran, B. 1988. has since informed much work in sonably easy-to-follow) logical treat- Generic Tasks in Knowledge-Based Reason- ing: The Right Level of Abstraction for knowledge representation. In contrast ment of some of the thorniest issues in Knowledge Acquisition. In Knowledge Ac- with Newell’s goal-oriented view of a knowledge representation: incomplete quisition for Knowledge-Based Systems, Vol- knowledge-level system as an epistemic knowledge, nonmonotonic reasoning, ume 1, eds. B. Gaines and J. Boose, 65–77. agent, as already mentioned, Levesque reasoning about actions, and logical San Diego, Calif.: Academic. and Lakemeyer are not concerned with omniscience. The answer to the second Clancey W. J. 1985. Heuristic Classification. goal-driven behavior. For them, a question is a bit more complicated. The Artificial Intelligence 27(1): 289–350. knowledge base is essentially a task-in- book is definitely going to be required Fensel, D.; Benjamins, V. R.; Motta, E.; and dependent body of knowledge to be in- material for anybody interested in for- Wielinga, B. 1999. UPML: A Framework for teracted with by means of two basic op- mal knowledge representation or in Knowledge System Reuse. Paper presented erations: Tell, to add new knowledge, formal theories of knowledge. Howev- at the Sixteenth International Joint Confer- and Ask, to find out what the system er, what about the wider world of ence on Artificial Intelligence (IJCAI-99), 31 knows or what is true about the world. knowledge-based systems? After all, July–5 August, Stockholm, Sweden. Chapter 5 in the book formally speci- one would expect a book with the Schreiber, G.; Akkermans, J. M.; Anjewier- fies Tell and Ask as operations that take words knowledge base in the title to be den, A. A.; de Hoog, R.; Shadbolt, N. R.; Van as arguments a sentence and an epis- of interest to the wider community of de Velde, W.; and Wielinga, B. J. 2000. Knowledge Engineering and Management: The temic state and return either an ele- researchers and practitioners in the COMMONKADS Methodology. Cambridge, ment in the set {yes, no} (Ask) or a new area of knowledge-based systems. In Mass.: MIT Press. epistemic state (Tell). addition, it would be nice if such inter- Stefik M. 1995. Introduction to Knowledge The first half of the book covers the est was not going to be driven purely Systems. San Francisco, Calif.: Morgan Kauf- basics, and the second half shows ap- by intellectual curiosity but also by the mann. plications of the framework to non- possibility of applying these results to monotonic reasoning, tractable rea- the engineering of real systems in real Enrico Motta is the director of the Knowl- edge Media Institute (KMi) of the Open soning, and reasoning about actions. I contexts. Unfortunately, no attempt is University in the United Kingdom. His won’t go into too many details here, made in the book to link the analysis main interest is in knowledge technologies, but essentially Levesque and Lakemey- either to concrete applications or to re- and his current research focuses on the er show how their framework can be search in task-performing knowledge- specification of reusable knowledge-based used to reconstruct approaches such as based systems, for example, the work components.

116 AI MAGAZINE Book Reviews

customer. The flight terrain applica- Heterogeneous Agent Systems tion is primarily used to maintain the path of a flight, taking into considera- tion the current position and three-di- A Review mensional terrain information provided by satellites and terrain databases, respectively. The supply-chain management application takes care of the P. Ravi Prakash inventory levels and the ordering logistics for production companies. All the applications, especially the last two, have characteristics that The notion of software agents has lacking coherence and consistency. Al- make them suitable for using agents. been around for more than a decade. though written by seven authors, the First, they consist of multiple, largely Since its beginning, the definition of treatment in this book is quite coher- independent, and well-defined tasks agent, like the definition of intelli- ent and consistent. to be done. Second, there are indepen- gence, has been quite controversial The book starts off with the follow- dent data sources, and these sources and often provoked hot discussions. ing questions: What is an agent? If a are independently updated. Third, the Questions such as the following nor- piece of software is not an agent, how actions executed can vary depending mally come up in such arguments: do you make it an agent? As a re- on the circumstances. Fourth, the en- What is an agent? Should a piece of sponse, the authors give 10 desiderata tities in the applications need to rea- software be categorized as an agent by for agents and agent platforms. These son with uncertain data and the looking at its behavioral characteris- desiderata roughly cover issues such as beliefs they hold about other entities tics or by the methodology using the accessing of heterogeneous data in the domain. which it was produced? Is a printer sources, a declarative framework for This book can roughly be divided daemon an agent? If a piece of soft- specifying actions, types of reasoning, ware is not an agent, is there a way to security, efficiency, reliability, and val- into three parts: (1) basic concepts, (2) make it an agent? Many attempts have idation. Of these issues, validation of implementation, and (3) advanced been made to define the notion of an infrastructure by deploying a set of concepts. agent or agency, ranging from quite applications seems out of place. The The first part discusses basic con- generic definitions to restrictive defi- main contributions of this book are an cepts such as the IMPACT architecture, nitions. approach to making a normal pro- service description language, the con- This book adopts a generic defini- gram an agent and providing a practi- verting of legacy data and software into tion of an agent: a piece of code that agents, and the development of agents. does a small and well-defined job by In chapter 2, the IMPACT architecture is Heterogeneous Agent Systems, V. S. Subrah- discussed. One distinguishing feature offering services. Some of its character- manian, Piero Bonatti, Jürgen Dix, of IMPACT is that it supports fuzzy istics are declarative specification, au- Thomas Eiter, Sarit Kraus, Fatmaözcan, matchmaking using concept hierar- tonomous behavior, and interactive- and Robert Ross, Cambridge, Mas- ness. This book primarily concentrates sachusetts, MIT Press, 580 pp., $60.00, chies. The service description language on abstracting the data sources and re- ISBN 0-262-19436-8. is discussed in chapter 3. Chapter 4 lated software to make them agents, contains the core concepts of the book—converting, by use of code calls, hence this less restrictive definition. cally implementable formal theory of legacy data and software application Although this book might not put an agent construction and agent interac- program interfaces (APIs) into services. end to the debates mentioned earlier, tions. After converting the APIs and defining it tries to answer the most practical To discuss and illustrate the con- question of how to convert a normal the constraints, the agent code looks cepts discussed throughout the book, program into an agent. similar to a PROLOG program. Chapter 6 three motivating examples are consid- Until now, most of the books in this discusses the components of agent pro- ered. This approach is one of the good field have discussed the issues in an in- grams—action base and action con- features of this book because using formal manner or only theoretically straints—which is followed by a discus- these examples throughout gives the without talking about concrete imple- sion of the syntax and semantics of the mentations, leaving the reader won- book a sense of continuity. The moti- agent programs. Part 1 is not very for- dering, “It is all fine, but how do I im- vating examples are a department mal and is highly readable. plement an agent?!” This book fills the store application, a flight terrain ap- The second part discusses imple- gap to some extent. It extensively talks plication, and a supply-chain man- mentation issues. The IMPACT server about the implementation of agents, agement application. The department implementation and protocol details using the IMPACT agent development store application is supposed to proac- are given in chapter 5. Chapter 12 dis- environment. Also, most of the books tively provide information and make cusses the implementations for issues in this field are a collection of articles multimedia presentations of the prod- such as the compiling of the agent written by different authors, often ucts depending on the profile of the programs and safety checks.

plication developed by the authors to for the motivating examples. There is illustrate the concepts discussed in the an impressive list of references at the book. This chapter is quite weak; it end of the book. One thing lacking in does not cover most of the concepts this book is exercises. Including them discussed in the book, such as belief for at least some chapters would have reasoning and temporal reasoning. made this book more useful for aca- Considering the supposedly signifi- demic purposes. The editing should cant features of agents introduced ear- have been tighter: A few grammatical ly on in the book, a more complex ap- errors have slipped through. Most of plication that demonstrates the use- the pages look a bit cluttered because fulness of the features should have of mathematics symbols. Indentation been chosen. for the items in the enumerated and One noticeable exclusion is agent bulleted lists and a bold font for theo- communication mechanisms. There is rem, deﬁnition, and example headings no mention of communicative acts; would have improved the readability. performatives; or FIPA (Foundation for This book might not be suitable as a Intelligent Physical Agents), the up- primary textbook for a course on coming agents standard. agents because of its specificity and There are significant advantages to lack of exercises. However, it will be applying this paradigm to make nor- useful as a supplementary textbook or mal programs into agents, especially a reference book. It will be of use to Simulating legacy code. The ﬁrst and foremost is the agent programmers (practitioners) Organizations the declarativeness that is achieved. It who want to get a concrete idea about allows easier program comprehension; the implementation issues. This book Computational Models of modification; and correctness check- also might interest those who want to ing, manual as well as automatic. It al- acquire an understanding of the prac- Institutions and Groups so allows you to plug high-level rea- tical implementations of the advanced soning capabilities, such as constraint, concepts such as reasoning with be- Edited by Michael J. Prietula, belief, and temporal and uncertainty liefs, uncertainty, and security with re- Kathleen M. Carley, and Les Gasser reasoning. into the program (agent). spect to agents. The bulk of the book, However, it is obvious that not all pro- I feel, is for the researchers with a bent 6 x 9, 350 pp., $45.00, ISBN 0-262-66108-X grams require such complex and so- (Prices higher outside the U.S. and subject to change for formal treatment. phisticated mechanisms: For example, without notice.) the department store example does Acknowledgment To order, call 800-356-0343 (US and Canada) or not need it, but applying this para- I would like to thank my colleague, M. digm to the flight terrain application (617) 625-8569. Sasikumar, for his valuable comments might be quite worthwhile. If the Distributed by The MIT Press, 5 Cambridge Cen- on this review. ter, Cambridge, MA 02142 flight terrain application was pro- grammed in a traditional way, it would not have the advantages of declarative programming mentioned earlier. Also, the advantages of an P. Ravi Prakash is a senior staff scientist at agent architecture such as a match- the National Centre for Software Technolo- making service will be missed, al- gy, India. He received his B.Tech in com- The third part (chapters 7 to 11), though it is debatable whether fuzzy puter science and engineering from JNTU, consuming roughly half the book, is matching is appropriate for all applica- Kakinada, India. His research interests in- devoted to advanced concepts such as tions. Thus, if a program is complex clude intelligent agents, multiagent sys- belief, temporal and uncertainty rea- enough, then using this paradigm will tems, planning and scheduling, soft com- soning, and security. This part is very be more advantageous than the tradi- puting, and parallel and distributed com- formal; it is meant for researchers in- tional way of programming. puting. His e-mail address is ravi@ncst. terested in the formal treatment of the Each chapter starts with an over- ernet.in. theory. Others are advised not to ven- view and ends with a section on relat- ture into reading these chapters, at ed work containing brief discussions least during the first read, lest they of other approaches, along with a might think that programming agents comparison with their approach. For is an esoteric discipline that is beyond all the algorithms discussed in this the reach of the everyday program- book, the complexity analysis, as well mer! as experimental results, are given. Chapter 13 discusses a logistics ap- There is an appendix giving the codes

118 AI MAGAZINE Book Reviews

ming, that is, the study of the actions Dynamic Logic that programs perform and the correctness of these actions. This has been a major issue in computer sci- A Review ence since Dijkstra’s (1968) attacks on the GOTO statement. Perhaps the most popular formal approach aimed at proving program correctness is Andrés Silva Hoare’s (1969), which is based on correctness assertions. In Hoare’s logic, statements of the form {a}P{b} say that if program P starts in an input state satisfying a, then if and when P halts, he real world is dynamic, and logic provide alternative applications it does so in a state satisfying b. Hoare any intelligent perception of the of modal logic to program specifica- provided some inference rules used to Tworld should include the con- tion and verification. The main differ- infer assertions about programs from cept of time. Remember that time and ence between the two is that temporal assertions about other programs. space are a priori conditions of human logic is endogenous, and dynamic In 1976, Pratt (1976) made the con- perception in Kant’s philosophy. On logic is exogenous. A logic is exoge- nection between program logic and the one hand, time is inherent to ac- modal logic, an older tradition in tion and change; on the other, action which classical logic is extended with and change are possible because of the modalities. The two most important passage of time. According to McDer- Harel, D.; Kozen, D.; and Tiuryn, J. Dy- modalities used in modal logic are ne- mott, “Dealing with time correctly namic Logic. Cambridge, Massachusetts, cessity and possibility, whose respec- would change everything in an AI pro- The MIT Press, 2000, 459 pp., $50.00, tive modal operators are ❒ and <>. gram” (McDermott 1982, p. 101). ISBN 0-262-08289-6. Therefore, if f is a formula, then so are It should not be surprising then that ❒ f and <>f. ❒ f should be read as “it is temporal reasoning has always been a necessary that f,” and <>f should be very important topic in many fields of nous if programs (actions) are explicit read as “it is possible that f.” Semanti- AI, particularly areas dealing with in the language. Temporal logic is en- cally, modal formulas are interpreted change, causality, and action (plan- dogenous, so its programs (actions) are according to Saul Kripke’s semantics, ning, diagnosis, natural language un- never explicit in the language. Dy- best known as Kripke frames. Basically, derstanding, and so on). AI develop- namic logic subsumes temporal logic. an interpretation in modal logic con- ments based on temporal reasoning Some cross-fertilization has already lead to general theories about time sists of a collection of many possible take place between AI, temporal logic, and action, such as McDermott’s worlds or states. Pratt’s discovery, fur- and dynamic logic. I focus on dynam- (1982) temporal logic, Vilain’s (1982) ther developed by other authors, led ic logic, which is the topic that is cov- theory of time, and Allen’s (1984) the- to the association of programs with ered at length in the book under re- ory of action and time. Work on the modal operators. As a result, program view. Dynamic logic is an approach to application of these results has taken logic can now make use of the well-de- program verification with strong AI place in fields such as planning and veloped corpus of modal logic. medical knowledge-based systems. potential. One of the most prominent Briefly, the dynamic logic approach However, action and change are not uses of dynamic logic in AI was to program logic is as follows: The as- an exclusive interest of AI. In main- Moore’s (1990) approach. Moore for- sociation of a modal operator, [] and stream computer science, any execu- malized some issues related to agency, <>, with a program P, gives birth to tion of a “traditional” computer pro- with a focus on what an agent needs the operators [P] and

. The exoge- gram is considered to perform an ac- to know to be able to perform an ac- nous characteristic of dynamic logic is tion that leads to a change of state. tion. For more information on this now clear. If f is a formula (proposi- From this point of view, the field of topic and a clear demonstration of the tional or ﬁrst order), then [P]f should program veriﬁcation, traditionally fo- usefulness of dynamic logic for agent be read as “necessarily, halting execu- cused on the correctness of actions reasoning and action, see the survey tions of P result in a state satisfying f.” carried out by program executions, by Meyer (1999). Also, some research However,

f should be read as “pos- can potentially provide AI with many in knowledge engineering inspired by, sibly, halting executions of P result in approaches suitable for dealing with or making use of, dynamic logic has a state satisfying f.” Therefore, Hoare’s action and change. Temporal logic been published van Harmelen and logic statements such as {a}P{b}, in dy- and dynamic logic are two of the ap- Balder (1992) and Fensel (1995). namic logic are expressed as a → [P]b. proaches that have been used in the Dynamic logic is an eclectic ap- Actually, dynamic logic subsumes ﬁelds of both AI and program veriﬁca- proach to program verification, as is Hoare logic and temporal logic as well. tion, temporal logic being the most evidenced by its history. This history The semantics of dynamic logic are popular. Both temporal and dynamic starts with the pragmatics of program- based on Kripke frames, demonstrat-

application of dynamic logic to AI or knowledge representation. However, Network and Netplay: AI researchers who want to deepen their understanding of the capabilities Virtual Groups and limits of dynamic logic will find on the Internet useful information in the book. References Edited by Fay Sudweeks, Margaret McLaughlin and Allen, J. 1984. Toward a General Theory of Sheizaf Rafaeli Action and Time. Artificial Intelligence Foreword by Ronald Rice 23:123–154. Dijkstra, E. W. 1968. Go To Statement Con- The vast, international web of computer networks that is the internet offers millions of users sidered Harmful. Communications of the the opportunity to exchange electronic mail, photographs, and sound clips; to search ACM 11(3): 147–148. databases for books, CDs, cars, and term papers; to participate in real-time audio-and video- conferencing; and to shop for products both virtual and physical. This huge conglomerate Fensel, D. 1995. The Knowledge-Based Ac- of links, hyperlinks, and virtual links is not just a technology for linking computers—it is quisition and Representation Language KARL. a medium for communication. The convergence of computer and communication tech- New York: Kluwer Academic. nologies creates a social convergence as well. People meet in chat rooms and discussion groups to converse on everything from automechanics to post-modern art; networked Hoare, C. A. R. 1969. An Axiomatic Basis groups form virtually and on-the-fly, as common interests dictate. Like interpersonal com- for Computer Programming. Communi- munication, the networks are participatory, their content made up by their audience. Like cations of the ACM 12(10): 576–580, 583. mass-mediated communication, they involve large audiences. But the networks are neither purely interpersonal nor purely mass—they are a new phenomenon. McDermott, D. 1982. A Temporal Logic for Network and Netplay addresses the mutual influences between information technology Reasoning about Processes and Plans. Cog- and group information and development, to assess the impact of computer-mediated communications on both work and play. Areas discussed include the growth and features of the nitive Science 6(2): 101–155. internet, network norms and experiences, and the essential nature of network communi- Meyer, J.-J.Ch. 1999. Dynamic Logic Rea- cations. soning about Actions and Agents. Paper 6 x 9, 320 pp. ISBN 0-262-69206-5 presented at the Workshop on Logic-Based Artificial Intelligence, 14–16 June, Wash- To order, call 800-356-0343 (US and Canada) or (617) 625-8569. ington, D.C. Distributed by The MIT Press, Cambridge, MA 02142 Moore, R. C. 1990. A Formal Theory of Knowledge and Action. In Readings in Plan- ning, eds. J. F. Allen, J. Hendler, and A. Tate, 480–519. San Francisco, Calif.: Morgan Kaufmann. ing that dynamic logic is built on solid to dynamic logic, such as temporal Pratt, V. R. 1976. Semantical Considera- modal logic foundations. logic, process logic, and Kleene algebra tions on Floyd-HCare Logic. In Proceedings Here we have a book that provides a (but, strangely enough, these topics of the Seventeenth Symposium on the deep insight into the topic of dynamic are covered in the last chapter of the Foundations of Computer Science, 109– logic. However, readers of this maga- book). The second part introduces 121. Washington, D.C.: IEEE Computer So- zine should be warned: This book does propositional dynamic logic, covering ciety. not provide tips on how to apply the syntax, semantics, properties, com- van Harmelen, F., and Balder, J. R. ML2: A concepts of dynamic logic to AI be- pleteness, complexity, and so on. The Formal Language for KADS Models of Exper- cause the main focus of the authors is third part, on first-order dynamic log- tise. Knowledge Acquisition 4(1): 127–167. the use of dynamic logic as a formal ic, is the most involved part of the Vilain, M. 1982. A System for Reasoning system for reasoning about programs. book and introduces syntax and se- about Time. In Proceedings of the Second This 460-page book is divided into mantics, uninterpreted and interpret- National Conference on Artificial Intelli- three parts: (1) fundamental concepts, ed levels, complexity, axiomatization, gence, 197–201. Menlo Park, Calif.: Ameri- can Association for Artificial Intelligence. (2) propositional dynamic logic, and expressive power of languages, and so (3) first-order dynamic logic. The first on. part provides readers with the neces- This book is a comprehensive Andrés Silva is assistant professor at Uni- sary background to understand dy- source of information on dynamic log- versidad Politécnica de Madrid. He started namic logic and makes the book self- ic. It is aimed at researchers, teachers, his research career at the Laboratory of Ar- contained. Despite the introductory and students of the subject. The book tifical Intelligence at Universidad Politécni- aim of this part, its contents are rather can be used in a dynamic logic course ca de Madrid and with the Group of Artifi- cial Intelligence and Learning (GRAIL) at deep, amounting to one-third of the because all chapters come with exer- Universidad Carlos III. In 1998 to 1999, he book. This first part covers mathemat- cises that teachers will find useful. If worked at the Joint Research Centre of the ical preliminaries, computability, com- you are interested in program logics European Commission in Ispra (Italy). He plexity, logic, and reasoning about and program verification using dy- holds a Ph.D. in computer science from programs. Also, the authors provide namic logic, this is your book. Do not Universidad Politécnica de Madrid. His e- an introduction to other topics related expect to find any information on the mail address is asilva@fi.upm.es.

120 AI MAGAZINE Book Reviews

world champion, Gary Kasparov. It is Automated Theorem Proving: Newborn’s background in the search issues of computer chess that led to his more recent interest in ATP. His book Theory and Practice provides an introduction to some of the basic logic, calculi, heuristics, and practicalities of first-order ATP. Rather A Review than working rigorously through the theory of mathematical logic, the Geoff Sutcliffe book focuses on just the necessary foundations for understanding how first-order ATP systems operate and links these foundations to the implementation of two ATP systems, HERBY and THEO. The book comes with a CD utomated Theorem Proving (ATP) statements deduced by the system, the containing the source code for HERBY deals with the development of overall scheme for controlling the de- and THEO and several suites of ATP Acomputer programs that show duction steps of the system, and the problems for HERBY and THEO to at- that some statement (the conjecture) heuristics that will control the fine- tempt.1 Thus, the reader is able to ex- is a logical consequence of a set of grained aspects of the deduction steps periment as both a user and a develop- statements (the axioms and hypothe- (the heuristics are most likely to deter- er of ATP systems. The provision of the ses). ATP systems are used in a wide va- mine the success or failure of an ATP ATP systems and problems sets this riety of domains: A mathematician system because they most directly book apart from most other introduc- might use the axioms of group theory tory books on ATP that provide a more to prove the conjecture that groups of in-depth treatment of the theory but Automated Theorem Proving: Theory and order two are commutative; a manage- Practice, Monty Newborn, Berlin, fail to get readers over the initial hur- ment consultant might formulate ax- Springer-Verlag, 231 pp., $54.95. ISBN dles of using an ATP system (a notable ioms that describe how organizations 0-387-95075-3. exception is Wos et al.’s book, Auto- grow and interact and, from these ax- mated Reasoning: Introduction and Ap- ioms, prove that organizational death plications [McGraw-Hill, 1992], which rates decrease with age; or a frustrated comes with the well-known ATP sys- control the system’s search for a solu- teenager might formulate the jumbled tem OTTER). Newborn’s book is suitable tion). Current state-of-the-art ATP sys- faces of a Rubik’s cube as a conjecture as an introduction to ATP for under- tems, such as E, VAMPIRE, E-SETHEO, and and prove, from axioms that describe graduate university students and inde- WALDMEISTER,have been developed legal changes to the cube’s configura- pendent, interested readers. with the benefit of years of experi- tion, that the cube can be rearranged After an introductory chapter ex- mentation and effort (information to the solution state. All these tasks plaining the structure of the book and can be performed by an ATP system, about these systems can easily be software installation, chapters 2 to 4 given an appropriate formulation of found on the World Wide Web). For a introduce first-order logic (in the syn- newcomer to the field, it is a daunt- the problem as axioms, hypotheses, tax used by HERBY and THEO) and the and a conjecture. Most commonly, ing, if not seemingly impossible, task basic mechanics and semantics of res- ATP systems are embedded as compo- to start on the road to building a de- olution-based ATP. Chapters 5 and 6 nents of larger, more complex soft- cent ATP system. It is almost impera- then provide the underlying theory ware systems, and in this context, the tive that a budding ATP system devel- and describe the calculi for the two ATP systems are required to au- oper should start with the benefit of ATP systems, chapter 5 corresponding tonomously solve subproblems that previous experience. Monty New- to HERBY and chapter 6 corresponding are generated by the overall system. To born’s book, Automated Theorem Prov- to THEO. The architecture, use, and im- build a useful ATP system, several is- ing: Theory and Practice, can contribute plementation of HERBY are described in sues have to carefully be considered, to this learning process. chapters 7, 8, and 11, respectively, and independently and in relation to each Monty Newborn is a professor of the same information is provided for other, and addressed in a synergetic computer science at McGill Universi- THEO in chapters 9, 10, and 12. The last manner. These issues include the ty, Canada. Newborn is probably bet- chapter steps aside to briefly discuss choice of logic that will be used to rep- ter known for his involvement with the CADE ATP System Competition resent the problems, the calculus that computer chess than with ATP. In par- (CADE [the Conference on Automated will be used for deduction, the pro- ticular, he is the long-standing chair- Deduction] is the major forum for the gramming language that will be used person of the ACM Computer Chess presentation of new research in all as- to write the ATP system, the data Committee, which organized the fa- pects of automated deduction). Vari- structures that will be used to hold the mous 1997 match in which IBM’s DEEP ants of HERBY and THEO participated in statements of the problem and the BLUE program defeated the human the competitions in 1997 and 1998.

based ATP is Herbrand’s theorem, among the linear resolution–based which shows that logical consequence systems available now. Chapter 9, as can be established by considering in- well as describing the architecture of stances of a set of clauses. The seman- THEO, describes some strategies that tic tree calculus implemented in HERBY can be used by a broad range of resolu- is a direct application of Herbrand’s tion-based systems to improve perfor- theorem. Herbrand’s theorem and the mance. Some of these strategies come semantic tree calculus are presented in at the cost of completeness; that is, if such a strategy is used, then some the- ADVANCES IN chapter 5. The semantic tree calculus is appropriate for an introductory orems can no longer be proved. How- KNOWLEDGE DISCOVERY book on ATP because it provides a way ever, it is well known that almost all to see the immediate consequences of contemporary high-performance ATP AND DATA MINING Herbrand’s theorem in action. It also systems use incomplete strategies (sometimes even unintentionally). Usama M. Fayyad, provides a clear framework in which to illustrate some issues of heuristic The chapters describing the imple- Gregory Piatetsky-Shapiro, control in an ATP system, as is done in mentations of HERBY and THEO will be Padhraic Smyth, and chapter 7. Unfortunately, the seman- too technically messy for a reader who Ramasamy Uthurusamy, tic tree calculus fails to benefit from is primarily interested in the deductive editors the abstraction inherent in the resolu- issues and will also be inadequate for a ISBN 0-262-56097-6 632 pp., index. tion inference rule, thus making it programmer who wants to understand The AAAI Press hard to develop a high-performance the internal workings of the systems Distributed by The MIT Press ATP system based on this calculus. The (although the programmer does have Massachusetts Institute of Technology, the option of examining the source 5 Cambridge Center Cambridge, reader should note that the examples code provided on the CD). These Massachusetts 02142 in chapter 5 all use a finite Herbrand chapters, however, will give all readers To order, call toll free: universe and that a semantic tree can (800) 356-0343 or (617) 625-8569. obviously get a lot deeper when the a feel for the complexity of source code and data structures that are re- MasterCard and VISA accepted. Herbrand universe is infinite, as exem- quired to implement an ATP system in plified by the example in chapter 8. an imperative programming language Each chapter ends with a set of exercis- The resolution calculus, the basis such as C. es that are useful for testing the read- for THEO, is described in chapter 6 with Overall, Newborn’s book meets the er’s understanding of the chapter ma- some nice examples. The explanation needs of its intended audience as a terial. of how a resolution proof can be ex- straightforward and practical intro- Many first-order ATP systems and tracted from a closed semantic tree duction to ATP. If it whets your ap- calculi, including those described in provides an excellent link to the pre- petite, you can take advantage of the this book, use the clause normal form ceding chapter and Herbrand’s theo- bibliography of appropriate further (CNF) of first-order logic statements. A rem. Linear resolution, a refinement of material. If not, you’ll walk away with procedure for converting a problem in the resolution calculus, is introduced at least a basic understanding of ATP. first-order form to CNF is described in next. Linear resolution is significant as chapter 3, along with instructions for the basis for the PROLOG programming Note using the conversion program sup- language (where the format of the 1. The first-order problems in the WFF di- plied on the CD. The use of CNF is problem is restricted to Horn clauses, rectory on the CD use the keyword theo- largely motivated by the resolution in- providing completeness for linear-in- rem to mark the conjecture to be proved, ference rule, which generalizes the input resolution) and its use in various but the compile program, for converting tuitively understandable modus po- high-performance ATP systems, for ex- problems to clause normal form, expects to nens rule: If A is true, and it is true that ample, PTTP, PROTEIN, and METEOR. Lin- find the keyword conclusion. To use the compile program, it is necessary to edit ei- A implies B, then B is true. THEO is a ear resolution also parallels the tableau ther the note.WFF files or the compile resolution-based ATP system. Chapter calculus used successfully by the E- source code file lexwff.c in the COMPSC di- 4 describes the binary-resolution and SETHEO system. Interestingly, Newborn rectory. factoring inference rules, which, in chooses to extend linear resolution to combination, implement the full-reso- linear-merge resolution. Historically, lution inference rule. Chapter 4 also resolution with merging was devel- Geoff Sutcliffe is a faculty member in the introduces the very important sub- oped separately but around the same Department of Computer Science at the sumption rule, which is used to elimi- time (late 1960s) as linear resolution. University of Miami. His research focuses nate redundant clauses that might be The link between them was noticed on the evaluation and appropriate applica- inferred (although the book admits later, and several ATP systems exploit- tion of automated theorem-proving (ATP) that HERBY and THEO implement only a ed the link until the late 1970s when systems, including the development of par- weakened form of subsumption, interest in developing their combina- allel and distributed ATP systems, and easy- to-use ATP system interfaces. His e-mail ad- called s–subsumption). tion seemed to disappear. The THEO dress is [email protected]. The underpinning for resolution- system is thus uniquely interesting

122 AI MAGAZINE