Browse Books

Reviewer: Doris C. Appleby

Rather than being a third edition of <__?__Pub Fmt italic>Artificial intelligence,<__?__Pub Fmt /italic> this “textbook” is almost entirely new. The quotes are intentional. The book does have exercises and various suggested syllabi, but is more in the nature of an introduction or survey than a text on AI. It is twice as long as the second edition [1] and has more than twice as many exercises. Methods, algorithms, and programs are described in procedural English, which is a well-done half-English, half-program form. The reader interested in LISP details is directed to Winston and Horn's 1989 LISP text [2]. This edition has 29 chapters<__?__Pub Caret> divided into three parts, plus an appendix on relational databases, included to facilitate understanding the Rete procedure (fast-forward chaining) of chapter 7, and natural-language access to database information. The sections are “Representation and Method,” “Learning and Regularity Recognition,” and “Vision and Language.” Winston provides simple descriptions of 18 applications in use today, from the 1976 MYCIN rule-based system, to aid in the determination of appropriate antibiotic treatment for patients with bacterial infections, to the 1992 GENINFER program, which performs Bayesian probability analysis as an aid in genetic counseling. Where the book works best as a text is when a technique (such as backward chaining) is described, followed by a toy system, including software (such as Zookeeper, an identification-oriented deduction system), followed by the commercial application, in this case MYCIN. The supporting software is easily available via the Internet, as described in the text. As of October 1992, 16 programs were available, supporting 13 chapters of the text plus the database appendix. All are written in Common Lisp, and they run on a variety of platforms. Each program comes with at least four files, including a “.txt” file describing the experiments to be performed, an “.exp” file containing the LISP calls to run the experiments, one or more “.lsp” files containing the actual LISP code, and a “.ref” file with a transcript of a program run. The “.txt” files include directions on running the programs, their intended purposes, a number of “performance experiments” involving running the programs as is, “knowledge experiments” involving modifying the data (rules, frame definitions, and so on) used by the programs, and “LISP projects” involving programming to enhance the programs in various ways. Unfortunately, little description of the algorithms, data formats, or semantics of the user commands is included. The experiments and projects are generally interesting, although the performance experiments provide little more insight than would be obtained by visually examining the “.ref” files. The programs, and particularly the documentation, are of mixed quality. While several support the text examples quite well, such as those dealing with rule-based systems (chapter<__?__Pub Fmt hardspace>7) and version space learning (chapter 20), others are sorely lacking in documentation or, as in the case of the various search programs, use data in a form different from, and not always compatible with, the text examples. This book is intended for students of computer science or engineering as well as non-numeric practitioners such as biologists, psychologists, philosophers, and linguists. <__?__Pub Fmt nolinebreak>Winston<__?__Pub Fmt /nolinebreak> claims that people in this second group “need to know about artificial intelligence to understand the principles that make intelligence possible.” In this regard, he performs the same function for AI that Isaac Asimov did for the hard sciences. He does an excellent job of describing highly technical material in a form that laypeople can understand. As a popularizer, he is subject to the same criticism as was Asimov—oversimplification. This book is good, but it is not clear who would gain from it. Also, it may add to the number of people for whom a little knowledge is a dangerous thing. What will a commercial AI shop look like__?__ Will it have statisticians and mathematicians who know whether Fisher's exact test, which the book discusses, makes sense__?__ Will expert LISP programmers who can modify a commercial product to meet the needs of the actual application being developed be available__?__ Or will non-technical people blindly use inappropriate tools they do not have the background to understand__?__ One group that will benefit from this book is those who are using an existing AI system appropriately and want to know, to satisfy their own intellectual curiosity, why it works as it does. Another is those who want a general survey, with the intent of delving deeper into some topics, guided by the excellent reference list.