Browse Books

Reviewer: Doris C. Appleby

Artificial Intelligence (AI) is, or at least has been, a hot topic at colleges, universities, and industrial laboratories. Students are eager to enroll in courses with AI in the title, and even some strictly undergraduate schools have developed liaisons with local industry involved in its applications. Thus, it would seem that a large number of undergraduate texts would be available. However, the six considered below pretty much cover the market. The books by Nilsson and Winston are largely theoretical in nature, written by major researchers in the field. The first is directed to the mathematically adept, while the second is intended for those in related fields. The texts by Charniak and McDermott and by Tanimoto are intended for beginners and include LISP programming with applications. Rich's uneven text attempts to combine clarity and complete explanations, with some success. Banerji's book assumes the the perils of AI applications can best be avoided through adherence to the methods of strict mathematical proof. Taken together, the books leave me somewhat uneasy. Although there seems to be agreement on which topics make up AI (see table 2), as one- or two-semester undergraduate offerings, the texts do not hang together very well. Charniak and McDermott address this issue in their preface, noting “how hard it is to extract coherent and lasting theories from a field developing as rapidly as AI” (p. viii). Winston excludes topics that are well-developed fields in and of themselves. Rather than being solid introductions to a well-defined field of knowledge, the books resemble those texts for liberal arts students that are intended to give the flavor of a technical field without requiring too much background. Interestingly, Fischler and Firschein's Intelligence: The Eye, the Brain, and the Computer [1], a book at the level of Scientific American, is being adopted as the primary text at a number of colleges. [For a review of this book, see , Rev. 8707-0538.] If one is looking for a text providing the necessary foundation material for work in AI, a better choice than any of these would be Genesereth and Nilsson's Logical Foundations for Artificial Intelligence [2], which assumes that an informed acquaintance with ideas expressed in mathematical form and with symbolic logic is necessary for further work. [For a review of this book, see , Rev. 8805-0321.] The six texts considered here paint the field with a broad brush, and as such are useful for a computer science student choosing an area for concentration or for a researcher in some other field who wants to survey Artificial Intelligence. Banerji [I hold to] the viewpoint that the field of Artificial Intelligence, like many fields of science, can and should be discussed with mathematical precision. . . . Much of the mathematical material is in the Appendix and. . . . the flavor of the arguments remains informal. . . . I have not assumed that the reader has more than a passing acquaintance with discrete mathematics. . . . This book can be used as a text in a second course in Artificial Intelligence (pp. ix–x). Given the intended audience, the number of errors in the second printing is unforgivable. If one follows through some of the complex diagrams, such as those on pages 33 and 177, one finds both typographical errors and incompleteness sufficient to confuse even the most diligent undergraduate. On pages 10 and 11, a formal statement is truncated so that the reader cannot possibly know what is intended unless he or she understood the matter beforehand. The diagram that follows, intended to illustrate the statement, is mislabeled in column-major rather than row-major form, so it is of no help at all. Nonetheless, this book is neatly organized to promote a valid position: that successful AI projects of any sort of generality need to be based firmly in correct mathematics. Banerji considers heuristics based on Newell, Shaw, and Simon's General Problem Solver [3]. This is followed by search reductions using notions from graph theory and concepts of machine learning based on the theories of Bruner [4] and their implementation by Pennypacker [5]. These topics lead nicely to the properties necessary for any language suitable for developing learning programs. Of the 36 exercises, 13 are theoretical, of the “prove that. . . .” type. The book would probably be best suited for use in part of a seminar or possibly for independent study. Charniak and McDermott [An AI textbook should] :9Binclude those aspects of the field that we felt would prove to be enduring. . . . present AI as a coherent body of ideas and methods. . . . of the last twenty years. acquaint the student not only with the classical programs of the field, but the underlying theory as well. . . . Our solution [to serving a diverse student population] has been to try to explain everything (pp. vii–viii). In trying to explain everything the authors have written excellent minitutorials on LISP, Bayesian statistics, the predicate calculus, quantification and skolemization, lambda-expressions, nonmonotonic logic, unification, resolution, and Gaussian convolutions. They do not try to summarize the calculus and list those sections and exercises requiring it as optional. However, do not be misled into choosing this as an elementary text for average students. Although little specific knowledge is assumed, a student will have to understand rapidly notions that are often developed throughout an entire course. For example, the reader, depending on his or her prior knowledge of statistics, may not understand the multivariate version of Bayes' theorem after seven pages of (admittedly well done) explanation. For the clever undergraduate, this book cannot be beat. It contains fascinating, up-to-date material, written in a clear and lively style, with compelling exercises at the end of each chapter. The authors also make reference to the relevant literature throughout the narrative and provide notes as well as a complete bibliography. The movement back and forth from theory to application, with excursions into the history of AI, is masterful. The book could serve equally well in a seminar, with students presenting material from the references; in a first AI course, with students completing substantial individual or team programming projects; or as a survey of AI concepts and realizations. As a reference, it provides introductory material on a wide variety of AI topics, with guides to the literature. Nilsson I have attempted here to describe fundamental AI ideas that underlie many of these applications. My organization of these ideas is not, then, based on the subject matter of their application, but is, instead, based on general computational concepts involving the kinds of data structures used, the types of operations performed on these data structures, and the properties of control structures used by AI systems (p. xi). Even though this is an old book on a rapidly changing subject, it is still worth owning, reading, selecting as a text in a graduate seminar, or using as a research sourcebook. The bibliography and notes are outstanding, as is the list of journals. This volume is theoretical, but it is not as basic as the one by Banerji. It does not attempt formal proofs in correct mathematical style, but it does explain in clear English how and why things work as they do. It also discusses why certain modes of thought are important and to which application areas they might lead. The exercises follow the presentation style, with instructions like “draw a sketch . . . ,” “explain why . . . ,” “find a counterexample . . . ,” and “write simple English sentences that express the meaning of . . . .” This book would be a good choice for a seminar, augmented by the assignment of journal papers. It would also make a good sourcebook in a course with a lab or programming component, although the instructor would be on his or her own here in devising suitable projects. Clearly, in this book Nilsson has collected some of the important notions an AI researcher must understand. In his preface, he invites readers to contact him with corrections or suggestions for the second edition, which, unfortunately, has not been forthcoming. An updated text and bibliography from this major researcher would be welcomed by many. Rich The goal of this book is to provide programmers and computer scientists with a readable introduction to the problems and techniques of artificial intelligence. . . . The references to the source literature guarantee that students have access to not just one approach, but to as many as possible of those whose eventual success still needs to be determined by further research (pp. xi–xii). This text was the first, to my knowledge, to present a loosely defined collection of the topics known as AI in a clear and interesting way, assuming little prior knowledge on the part of the reader. It was designed for an introductory graduate course, in conjunction with readings from original sources. In a previous review [8406-0422] the book was criticized for using examples from the early 1970s, with little mention of current issues. Its strengths include readability, well-executed diagrams, good problems and project suggestions, and an up-to-date (as of 1983) and thorough bibliography. It is a pleasure to read, and it explains in detail some of the concepts covered briefly in the Nilsson or Winston books, the only others that were available when this appeared. Although intended for graduate students then, it is probably more suitable for the knowledgeable undergraduates of today. The publisher indicates that a second edition may be forthcoming, but no publication date has been announced at this writing. I eagerly await a new edition. Tanimoto This book grew out of the perception that hands-on experimentation coordinated with textbook explanations of principles and of actual programs can provide an ideal learning combination for students of artificial intelligence. . . . [The text] integrates the presentation of principles with actual runnable LISP illustrations (p. vii). This text is intended for a two- or three-quarter course combining LISP language instruction with AI applications and principles. Many instructors follow this guideline, selecting a language manual separate from the AI text. Tanimoto claims that this book is for undergraduates, while Charniak and McDermott's is more suitable for graduate students. Yet he assumes prior courses in logic, calculus, and probability. Rumors from the front have it that students are having a lot of trouble with both texts. However, given three quarters or two semesters, upper-level students should do nicely with this. If only one semester is available, they will be frustrated. Previous reviewers have been at odds concerning its worth. Pollack [ , Rev. 8710-0831] found that it “fulfills its purpose quite well,” while Lee [ SIGART Newsl. 103 (Jan. 1988)] “cannot recommend this book as it stands for student use.” The problem seems to be that the author attempts to combine difficult concepts with programming exercises, sometimes within the same paragraph. Nonetheless, it is up-to-date, interesting, inclusive, and readable. The hands-on, laboratory approach may well be appreciated by undergraduates, who often miss the fine points of theory the first time around. Winston The central goals of Artificial Intelligence are to make computers more useful and to understand the principles which make intelligence possible. . . . The purpose of this book is to provide [psychologists, philosophers, linguists, and other people] with a picture of what is going on. Thankfully no special background is required, not even computer programming experience or sophisticated mathematical training (p. 1). To understand the ideas introduced in the first nine chapters in depth, there is a need to see some of them embodied in program form. A study of LISP, with examples drawn from the earlier material, addresses that need (p. 5). Winston's book is intended to introduce workers in fields other than computer science to AI. In particular, it is aimed at the non-mathematically oriented whose disciplines “intersect the natural concerns of people doing Artificial Intelligence” (p. 5). Riesbeck [ , Rev. 8409-0703] cautions that this text will not provide the necessary background for doing AI, but that it can provide an understanding of some topics. Winston avoids stressing linguistics, pattern recognition, decision theory, or automatic theorem proving, which require more mathematical background than the addressed audience has. Extensive, well-annotated references provide direction for further study. Concepts are presented in the first nine chapters, and an elementary LISP tutorial with examples from these makes up chapters 10–16. This is a thorough update of an enormously popular first edition. Coded algorithms have been successfully rewritten in English and will be more accessible for the non-technical reader. Exercises are collected in the final chapter. These range from thought questions, “find the bug,” and the programming of simple LISP functions to the writing of a compiler and interpreter for augmented transition networks. Several of the exercises have been supplied by other writers. Webber and Nilsson As can be seen in table 1, all the texts considered include extensive annotated bibliographies. These are hard to keep current in a field as active as AI. Thus, a collection of readings, such as that of Webber and Nilsson, could serve a wide audience if updated regularly—say every two years. As the authors note, many of the papers included are hard to come by, having previously been published only in limited-edition conference proceedings. This volume includes 31 titles in five categories: Search and Search Representation, Deduction, Problem-Solving and Planning, Expert Systems and AI Applications, and Advanced Topics. This last includes material on frames, nonmonotonic reasoning, and speech. Interestingly, there are no papers on vision or robotics, topics that are included in many AI texts.