Browse Books

Reviewer: Stanley Wertheimer

There is no generally accepted definition of the word heuristic; however, most interested parties agree that a heuristic is a rule-of-thumb, an educated g uess, the application of intuition or common sense, or inductive intelligence. S eventeenth and eighteenth century mathematicians used heuristics in place of est ablished theory in analysis (Fourier series, integration theory, differential eq uations). They were called theorems but were not proved or proved incorrectly - which gave those in the nineteenth century some cleaning up to do. Two points should be made: some of the greatest mathematicians postulated and used these unproved theorems; the theorems were usually correct. If mathematics had waited for its friend rigor before sallying forth, rigor's companion mortis would have kept progress to a minium for long periods of time. This is not to say that un tried half-baked ideas should be allowed to proliferate, but that informed exper ience-based ideas should be carefully used by practitioners while others try to provide theoretical justification. The author of the book under review is both practitioner and justifier. Physics, chemistry, and the biological sciences rely on the infusion of new unpr oven ideas for progress. Often, however, the experiments (if not of the Gedanke n kind) require large quantities of money and time. A sizeable computational co puter science problem also often requires new ideas; the basis for these ideas i s not physical experimentation, however, but analytic insight, as in mathematics , so cost is not so much of a question and work proceeds faster. The insight te nds to take the form of a heuristic method. One is not concerned with conceptua lization (e.g., the theory of relativity) as much as with solution (e.g., comput e a satellite's orbit); in this sense one would like to prove that the solution method will always yield the correct answers. However, as is the case in so man y computational problems, the solver must often be happy with a sufficiently goo d answer most of the time. In the classical applied science problems in the continuous world, the underlyin y structures to be solved are such things as linear and nonlinear systems and eq uations; the theoretical and numerical properties of such systems are usualy wel l understood due to long and continued effort. The structures underlying many hard computer science problems are trees and grap hs; solution of these problems often requires searching for a distinctive node o r path. Exhaustive search is theoretically possible but often practically impos sible. Problem solution becomes possible with the use of heuristics to guide th e way; even so, there is a chance that the problem will not be solved, or the so lution will only be close to optimal. Pearl has gathered together, for the firs t time, almost all that is current in solving search problems in computer scienc e using heuristics. Pearl give some well known examples of the kinds of problems which with the book is concerned: the 8-Queens problem; the 8-Puzzle; the Traveling Salesman proble m; the Counterfeit Coin problem. He also devotes the last third of the book to game-playing programs. The first two parts are: Problem Solving Strategies and the Nature of Heuristic Search (an introduction, typical problems, descriptive taxonomy of methods); and Performance Analysis of Heuristic Methods (mathematica l analysis of several heuristics). In the game-playing section Pearl applied he uristic search methods, analyzes them, and presents some experimental results. The mathematics involved is mostly probability theory. Important search methods discussed are hill-climbing, depth-first, backtracking, best-first, and hybrid methods is carefully laid out. (Pearl several times includes verbatim sections of his own papers, mostly written over the past five years.) An attempt is made at analyzing the way in which heuristic functions are discovered. There is a p articularly interesting discussion about look-ahead; why it might be theoretical ly unjustified but practically effective. In the game-playing sections Pearl co mpares pruning (a backtracking procedure), SSS* (a best-first search for an opti mal playing strategy), and SCOUT (a cautious test-before-evaluate strategy of hi s own). This book can be used by one wishing an introduction to the topic and a knowledg e of where the problems are as of publication, by graduate students (there are s ufficient and interesting problems at the end of each of the ten chapters), or b y rsearchers in the field, as a reference (there are eight pages in the bibliogr aphy, roughly 150 entries). Pearl has taken the work of others, notably Nils Ni lsson, as of the late 1970's, added what he and his colleagues at the UCLA Cogni tive Systems Laboratory has since done, and come up with an eminently readable b ook (the first on the subject) which is sure to be a standard reference for seve ral years.