Browse Books

Reviewer: Nigel G. Ward

The authors take induction to encompass “all inferential processes that expand knowledge in the face of uncertainty.” The purpose of this book is to present their “framework for understanding induction,” which is centered around the if-then rule, but also includes such notions as category, hierarchy, default, prediction, message, search, support, competition, and variation. Chapter 1 is a reasonable introduction to the problem of induction. The authors lay out the important issues, argue against naive syntactic approaches, and lay out their framework. Chapter 2 explains how rules can model the environment. Chapter 3 is perhaps the key chapter. The first topic is rule modification, with the focus on modifying rule strengths and the credit-assignment problem. A partial solution, the bucket-brigade algorithm, is presented poorly, with confusing analogies and suggestive examples. Discussion of the second topic, learning new rules, is vague, unfortunately. Chapter 4 is on computational models. The account of Holland's genetically inspired classifier systems (also hard to read) is at a much lower level than the rest of the book. The discussion of PI, a program under development, serves to situate learning in the context of problem solving. One of the authors' main points is that rules compete among themselves for the right to analyze the environment (both for categorization and prediction) and for the right to suggest behavior. Chapter 5 illustrates this by showing how some data on the conditioning of rats are best explained in these terms. Chapter 6, on category formation, reviews the important ideas on categories and some unimportant new experiments. Chapter 7 sketches the findings on naive physics and on models of personality, including the fundamental attribution error. Chapter 8 discusses the role of knowledge of variability in generalization—how knowledge of the typical variability of categories (among members of a species, or among inhabitants of a region) affects the propensity to generalize from a small number of examples. Chapter 9 relates learning to education. Statistics courses help people reason about uncertainty, whereas logic courses do not help people reason formally very much. Experiments show the cause: naive people are sensitive to variability, but do not think logically. For example, “If A then B” is often thought of in terms of a “pragmatic schema” of permission. Chapter 10 surveys research in analogy. Chapter 11 discusses the process of scientific discovery and stresses the importance of mental models to scientists. The book has its flaws: It is annoyingly repetitive. The authors digress frequently for no reason. Some sections have no reason to be here. The authors' framework is too loose to lead them to say anything new about most of the fields they survey. The book casts no new light on the traditional problem of induction. The book is directed toward cognitive scientists in general. As such, it reports a large number of psychology experiments and presents a broad range of ideas on learning at a fairly accessible level. It's very sensible and never simplistic. People interested in Machine Learning, Computer-Aided Instruction, and Learning in general should find this book interesting.