One of the features of a computer program that attempts concept learning is described by the term bias. The bias of a learning program refers to the collection of factors that are brought to bear upon the selection and consideration of partially formed hypotheses pertaining to the concept being learned. For most learning programs, the bias is fixed and provided by the program’s author. Utgoff investigates how a learning program may modify its bias by considering when a shift in bias should be attempted, by proposing a method for accomplishing the shift, and by implementing a program that demonstrates procedures for performing a shift.

The book begins with an introduction to machine learning and bias and a discussion of related work. The third chapter identifies the design choices made by the author and introduces the reader to the Recommend-Translate-Assimilate (RTA) method. The fourth chapter contains a description of the LEX program and the STABB program. LEX is a well-known learning program in AI and is the environment for testing ideas about bias and the shifting of bias. Shift-To-A-Better-Bias (STABB) is Utgoff’s program for shifting the bias within LEX.

The fifth and sixth chapters present two procedures for the shifting of bias. The Least Disjunction algorithm and the Constraint Back-Propagation algorithm both modify bias by introducing additional concept descriptions to the existing collection of concept descriptions. Examples, using LEX, of both of these procedures are provided. A seventh chapter provides a summary and contains a discussion of the issues pertaining to these topics. An appendix is provided that contains the LISP code for STABB and relevant parts of LEX.

Utgoff’s book is based upon his doctoral dissertation and is of an advanced nature. The book’s purpose is to present an open question in machine learning and investigations directed toward understanding and resolving that question. This purpose is adequately met. The relative shortness of the book can be explained by the nature of its origin and is appropriate for the specific subject.

The major highlight of this book is the collection of examples that illustrate the two algorithms under consideration. An additional highlight is the inclusion of the LISP code, which enables a reader to view this work on another level. The weakness of the book is in the content of the preparatory material. The consideration of other machine learning work is often cursory.

Overall, this volume should be viewed positively. The main topic of the book is an important one, and it is treated carefully and comprehensively by the author. The book is also a welcome addition to the literature on machine learning because of its expanse and the detail that it contains.

This book is intended for those with an interest or need for information on machine learning at an advanced level. AI researchers, computer scientists, and graduate students comprise the intended audience. To fully appreciate the examples in the text, the reader should have some rudimentary awareness of the techniques of the integral calculus.

A fairly extensive reference list is provided, and an index is included. The layout of the book is nicely done, the number of typographical errors is minimal, and the various algorithmic figures and mathematical expressions are very clear.