Book Reviews AI Magazine Volume 10 Number 3 (1989) (© AAAI) BookReviews

Automated Reasoning: author writes with a freshness and an price of a (computer science) book is Thirty-Three Basic obvious love for logic. inversely proportional to its content. Between the listing of the prob- This book confirms this law: It is Research Problems lems and their detailed discussion is delightfully low priced. As Georg Ulrich Wend1 a review of AR. It seems to be too Christoph Lichtenberg, a German sketchy for the novice, but a com- physicist of the eighteenth century To read the book Automated Reason- panion book by the same author said: “If you have two trousers, sell ing: Thirty-Three Basic Research Prob- (Wos, L.; Overbeek, R.; Lusk, E.; and one, and buy this book.” lems (Prentice Hall, Englewood Cliffs, Boyle, J. 1984. Automated Reasoning: Ulrich Wend1 is at the SCS Information- N.J., 1987, 300 pp., $11.00) by Larry Introduction and Applications. Engle- stechnik GmbH, Hoerselbergstr. 3, 8000 Wos “it is not necessary to be an wood Cliffs, N.J.: Prentice-Hall.) gives Munchen 80. expert in mathematics or logic or far more details. computer science” (from the preface). Chapter 5, the in-depth discussion, However, even if you are such an is remarkable for its style. It is fasci- expert, you will read it with interest, nating to immerse oneself into the Logical Foundations of and likely, with enjoyment. details guided more by questions Artificial Intelligence The book is outstanding for its pre- from the author than answers; sentation of the theme. Following the indeed, a whole paragraph consists of Drew McDermott introductory chapter, Wos discusses nothing but questions! You will some obstacles to the automation of enjoy this chapter and the whole Knowledge is important to intelligent reasoning in Chapter 2. In Chapter 3, book, especially if you dislike the programs: Just about everyone in AI he lists the research problems (with ‘definition-theorem-proof’ style. The would agree, but the agreement short descriptions) in nine groups: six author’s presentation lets you forget doesn’t extend much further. There is problems on strategy, five on infer- that completely solving one of the a weak sense of “know” in which a ence rules, six on demodulation, one problems is considered equivalent to computer program knows P if its cor- on subsumption, three on knowledge finishing a Ph.D.! rectness depends on l? For instance, representation, two on global A non-expert will have the most an airline reservation system might approach, one on logic program- difficulty with the given test prob- “know” every passenger has a 0.9 ming, two on self-analysis, and six on lems primarily because they are all probability of showing up for a other areas. After a short review of out of the field of mathematics. reserved flight in the sense that it automated reasoning (AR) in Chapter Again, it is not trivial to see a corre- books 11 percent too many passen- 4, these problems are discussed in spondence to the problems or the gers. However, more interesting pos- detail in Chapter 5. Chapter 6 gives obstacles. At this point, it is likely the sibilities exist. A program might have some sets of test problems, all con- author’s opinion-“zeal, interest, and a notation that facts can be cerning a mathematical discipline. curiosity” suffice as prerequisites-is expressed, and it might consult a An appendix as interesting as the bib- too optimistic. However, this book is database of such facts to move liography follows. Last but not least also a workbook, and the appendix through problems. In this case, we there is an excellent index. shows the reader how to get the have a more explicit concept that a The discussion of the obstacles in appropriate software. Finally, besides program knows P if P is in the Chapter 2 is relaxing; although some a machine such as a VAX, you only database (or can be derived from it repetition exists in the book, the need one thing: a lot of time. Per- when required). This concept is the author spares you a puffy pseu- haps the best thing to say about the idea of declarative knowledge; the more dophilosophical treatise on AI in gen- book is that it tempted me to wait for mundane concept is called procedural eral and AR specifically. After reading a copy of the companion software. knowledge. The book Logical Founda- this chapter, you are convinced there The book as a whole is a rarity in tions of Artificial Intelligence (Morgan is no general problem solver, and you that it successfully serves several Kaufmann, Los Altos, Calif., 1987, have a pleasant introduction to the audiences at the same time. The 406 pp., $48.95) by Michael Gene- belly of the beast. layperson gets a real background sereth and Nils Nilsson is about the A non-expert might have some knowledge of one of the main disci- declarative version. (Declarative problems in mapping the eight obsta- plines of AI, and the theorist gets a knowledge could be false, and we cles described to the problems and good occasion to put the theory to would do better to call it belief, as I problem areas cited earlier. However, practice. Most astonishingly, both often do in this review.) the detailed discussion that forms the can enjoy it. Declarative knowledge requires a heart of the book provides the reader One last point is worth mention- notation, often called a knowledge with ample reward. It is clear that the ing. It seems to be a law that the representation system. Those who

FALL I989 103 Book Reviews study declarative knowledge repre- concept learning, in which new rules warned that the foundations being sentations are divided about whether are inferred from facts that would explored are not exactly the founda- such notations ought to be thought follow from them. The algorithms tions of AI. of as variants of the predicate calcu- described-based on Tom Mitchell’s The authors’ viewpoint causes dis- lus and related systems of mathemat- idea of version spaces-are specific to tortions throughout the book. The ical logic invented by logicians, the domain of the concept of learn- chapter on planning (Chapter 12) is mathematicians, and philosophers in ing. Similarly, Chapter 8 describes typical. The planning problem is this century. Lately, those who think Bayesian inference, from a priori defined as finding a constructive they ought to be so regarded seem to probabilities to a posteriori probabili- proof that a series of actions will be winning. Genesereth and Nilsson ties given new evidence. bring about a desired state of the have no doubt and take it for granted In competition with this diversity world. It is assumed that the right that the semantic tools of mathemat- is the idea of a unified model of infer- way to find this proof is to tailor a ical logic are indispensable for ana- ence. The desire for such a model is general-purpose theorem prover with lyzing knowledge representations. strong among those who study a few specialized strategies. This They adopt the label Iogicism for this declarative representations, and description of the planning problem doctrine. Genesereth and Nilsson are no excep- is quite remote from any description In Chapter 2, they develop these tion. As are most of their colleagues, that active researchers in the field tools from the point of view of AI, they are drawn to the model of infer- would produce. Perhaps Genesereth which is rather different from the ence as the derivation of conclusions and Nilsson feel that techniques for logician’s point of view. They intro- that are entailed by a set of beliefs. solving planning problems will come duce the idea of conceptualization, They wander from this idea in a few and go, but the foundations can be which is roughly what a philosopher places but not for long. It is not hard secure. If so, they are too casual calls the intended model of a formal to see why: Deduction is one of the about the implicit claim. theory. A formal theory provides fews kinds of inference for which we I admit to bias on these questions. predicates and functions for talking have an interesting general theory. My skepticism should be balanced by about (some part of) the world and The authors made a conscious the agreement of many perfectly rea- axioms involving these symbols. decision not to talk much about sonable people with Genesereth and Formal semantics specifies a mapping search processes. Unfortunately, a Nilsson’s view. However, it bothers from the symbols to the entities in deductive process is always confront- me that questions about the scope of the world. It is one of the key ed with the problem of what to their enterprise are treated so superfi- insights (and disturbing insights) of deduce next; it is impossible to cor- cially in this book. mathematical logic that unique map- rectly choose with any confidence, so Let me put such doubts aside and pings are rare. Any given theory has programs that deduce must try vari- pretend from now on that deduction many interpretations that make its ous options. Sometimes, they can is the foundation of AI. Within this axioms true, that is, several models. generate many useless inferences perspective, the book has some Mathematicians cheerfully study before finding useful ones. Chapters strengths and some weaknesses. them all, but philosophers worry 4 and 5 are devoted to the topic of Chapter 6 is an excellent discussion more about how the correct model implementing deductive processes of nonmonotonic logic, a family of can be picked out. Fortunately, in AI, using the resolution method with logic extensions that allow defeasible we can ignore these problems and various refinement strategies. Howev- conclusions, which can be blocked by treat ontological commitment as an er, throughout the rest of the book, knowing more. Most real-world infer- engineering decision. It doesn’t little focus is given to the actual com- ence is nonmonotonic because know- matter that the robot we build putational consequences of relying ing more usually causes one to cannot intend a model; its builders on these methods. Usually, proofs are change one’s mind in some way. For can. Genesereth and Nilsson give a given with a passing warning that example, if told that person P is an lucid explanation of this topic. finding the proof might be expensive. adult living in suburbia, you would When they reach the topic of infer- It’s clear what the authors are probably infer P owns a car. Now, ence, however, they seem to lose thinking. They are studying founda- suppose the further information that their way. Chapters 3 through 5 are tions; so, what’s important is what P is blind. It is a problem with the concerned with this topic, but it is should be inferred in a situation, not logicist approach to infer that logic central to the entire book. Inference what actually can be practically lacks this ability; if a proposition is can be considered as the deriving of inferred (compared, say, to theoreti- entailed by a set of beliefs, it is probably useful, probably correct cal mechanics). However, as I argued entailed by any superset. Various conclusions from a set of beliefs. This earlier, what should be inferred is nonmonotonic variants of logic have characterization is vague, but the probably different from what is been proposed. They are all covered vagueness is inevitable. Almost any entailed by current beliefs because in Chapter 6, especially John algorithm can be thought of as doing not all that is entailed is useful and McCarthy’s circumscription, which inference in some sense, and if we not all that is probably correct is augments a standard theory with a arrange that its premises and conclu- entailed. The authors are forced to special axiom schema that can sions are expressed in a logical nota- acknowledge these discrepancies in change nonmonotonically as new tion (not a difficult requirement), their detours to alternative inference beliefs are added. This description of then it can be thought of as doing techniques in Chapters 6, 7, and 8, circumscription is the best I have inference on a declarative knowledge but their heart isn’t in it. When it seen; no one could ask for more from representation. The authors give sev- comes to actual applications, they a textbook. eral examples throughout the book. always revert to classical deduction. Chapter 7, on induction, is too Chapter 7, for instance, is devoted to The unsophisticated reader should be short. Induction is defined as the

104 AI MAGAZINE Book Reviews problem of finding a hypothesis that this connection, the value of meta- Response to Drew McDermott’s entails observed data. An excellent reasoning as a programming or repre- Review of Logical Foundations description follows of version spaces, a sentation technique has seldom been of Artificial Intelligence useful framework for thinking about demonstrated. Chapter 10 is mainly sets of hypotheses that have not been concerned with pointing out various Nils Nilsson ruled out by data observed so far. alternative architectures for metarea- However, no mention is made of soning. Little discussion is given of McDermott makes some valid points Ehud Shapiro’s or Gordon Plotkin’s what it is for. One possibility is for in his review. I acknowledge some of generalizations of the idea to a logical reasoning about the reasoning of these in this response. It’s too bad framework, which is an amazing other agents, which connects to the that he chose to embed the helpful omission. It would have been nice to syntactic belief calculus of Chapter 9. comments in the context of his by- see a discussion of explanation-based However, the possibility is not really now-tiresome doubts about the value generalization, a more recent idea of explored, except for a baffling detour of logic in AI. (The reader who wants Mitchell’s in which a problem solu- into an alternative formalization of to be saturated with the arguments tion, expressed as a proof, is general- belief that is never related to the surrounding these doubts should see ized and stored as a new lemma. approaches of Chapter 9. McDermott’s 1977 A Critique of Pure Perhaps, this idea is too recent, but it Chapters 11 and 12 are about tem- Reason and the accompanying com- would fit the authors’ world view well. poral reasoning and planning. They mentary in Computational Intelligence, Chapter 8, on probabilities, is are out of date, based on the situa- 3(3). weaker. It deals with two topics, with tion calculus devised by John McDermott accurately summarizes little hint about how they are related McCarthy and Patrick Hayes in the our book as being about the represen- or how much of the problem of rea- late 1960s. A lot of work has been tation and use of declarative knowl- soning under uncertainty is covered. done in the last 20 years in the logi- edge in AI. He duly notes that The first topic is Bayesian inference, cist tradition on formalizing tempo- declarative knowledge requires a which underlies much work on ral reasoning to handle continuous notation and that some AI expert systems. The discussion is time and alternative possible worlds. researchers think most of the nota- clear and helpful. The second topic is It is mentioned only in the biblio- tional schemes being used are vari- Nilsson’s probabilistic logic, an graphic notes. ants of the predicate calculus. He attempt to generalize logical entail- Chapter 13 is entitled Intelligent- even says, “Lately those who think ment so that the probability of a con- Agent Architecture, but it is nothing they ought to be so regarded seem to clusion can be found given the of the sort. It is hard to say what it is be winning.” Under these circum- probabilities of some premises. It is about. I think it was meant to be a stances, it does seem odd for McDer- not clear what the theoretical or prac- nod toward robotics. mott to devote much space to tical significance of probabilistic logic In summary, this book has some complaining about the logical basis is. The maximum-entropy method is excellent parts, notably its treat- of a book whose very title proclaims briefly mentioned here; I wish it had ments of formal semantics, non- it is about logical foundations. In any been covered in more depth because monotonic logic, and logics of case, given such a title, it wouldn’t it is of interest in its own right. knowledge and belief. However, it seem necessary that readers “should Chapter 9 is a description of the omits detailed discussion of the work be warned that the foundations being logic of knowledge and belief using of many logicists, including James explored are not exactly the founda- two different conceptualizations: Allen, Alan Bundy, Ernie Davis, tions of [all of] AI.” Kurt Konolige’s syntactic approach Patrick Hayes, Robert Kowalski, Ray It seems one of McDermott’s main and J.K.K. Hintikka’s possible worlds Reiter, and Udi Shapiro. Opportuni- complaints about the logical approach. This discussion is lucid and ties were missed for tracing a consis- approach is based on his view that thorough. In the syntactic approach, tent thread through the topics “logicists” think of inference as being belief is a predicate on formulas in an covered. Nonmonotonic reasoning is simply sound deductive inference. agent’s database. In the possible rarely mentioned after Chapter 6, Actually, we happen to agree with worlds approach, an agent does not even though it is relevant in several McDermott’s statement that “almost believe P if for all the agent knows, P subsequent chapters. Reasoning any algorithm can be thought of as might be false; and this statement is about reasoning is covered in Chap- doing inference in some sense, and if formalized as “There exists a world, ters 9, 10, and 13, each time from a we arrange that its premises and con- possible as far as the agent believes, different perspective. Finally, the clusions are expressed in a logical in which P is false.” In either book is unself-conscious about the notation (not a difficult require- approach, we can make inferences importance of logic in AI, which ment), then it can be thought of as such as “If Fred knows that Mary’s might be less than the authors doing inference on a declarative phone number is the same as John’s, believe. knowledge representation.” He inap- and he knows Mary’s phone number, propriately and unfairly characterizes then he knows John’s” without our treatment of various nondeduc- Drew K McDermott is a professor at the knowing what the phone number is. tive inference techniques (nonmono- Chapter 10 is about metareasoning, Computer Science Department, Yale tonic reasoning, inductive reasoning, or reasoning about reasoning. This University, P.0. Box 21.58 Yale Station, and probabilistic reasoning) as mere concept mesmerizes many in AI, New Haven, CT 06520. detours from what he would like to probably because it seems intimately regard as our commitment to deduc- connected with our ability to con- tive inference. We had hoped it would sciously introspect and observe our- be obvious that our commitment selves thinking. However, except for regarding inference isn’t exclusively

FALL 1989 10.5 Book Reviews

Have KEE”- Will Travel

RAPID PROTOTYPING of Robust, Extensible Knowledge-Based Systems

AMOS OSHRIN Knowledge Systems Consultant 4 Morning Sun Court l Mountain View, CA 94043 (415) 969-9600

KEE is a registered trademark of IntelliCorp RESULTS, Not Just Reports

For free information, circle no. 101

This publication is available in microform to soundness but to an understand- chapter or more on temporal reason- from University ing of the underlying theoretical ing and how time-based (rather than properties of various inference meth- state-based) techniques might be ods. Soundness is just one property used in planning would also have that an inference method might or been useful. might not have; minimal-model McDermott appropriately observes entailment and version-space proper- that our chapter on induction was ties are others We are curious about too short. what subtleties McDermott found in The last-chapter in the book, Chap- Chapters 6, 7, and 8 to make him ter 13 on agent architectures, is think that “our heart wasn’t in it.” admittedly speculative and strayed McDermott’s specific criticisms are somewhat close to research frontiers better than his general ones. He (as did some of the other chapters, makes some good points about our perhaps more successfully). We think chapters on planning. Our goal in that the material in Chapter 13 pre- 1 Please send information about these titles: these chapters was to present the fun- sents an interesting way to think damental conceptualization of the about an agent which is embedded in state-based approach and the situa- a world that it must sense and in Name tion calculus based on it. No founda- which it must act. We like to think of Company/Institution tions text could omit these basic and it more as a foundation for thinking classical ideas. We agree with McDer- about robots than as “a nod toward Address mott that it would have been good to robotics.” City include more material on practical state Zip planning (for example, hierarchical Nils J. Nilssorz is a professor and chair- Phone 1 I planning, opportunistic planning, Cd talLfree 600-521-3044. In M~chtgan. and constraint-based planning). It is man of the Computer Science Depart- Alaska and Hawaii call collect 313-761-4700 Or mail inquiry to: University Microfilms International hard, though, to decide which of this ment, Stanford University, Stanford, CA 300 North Zeeb Road. Ann Arbor MI 46106 additional material is foundational. A 94305. For free information, circle no. 120 106 AI MAGAZINE