Browse Books

Reviewer: Harvey Cohn

Today one of the salient features of artificial intelligence is machine learning, as exemplified by such applications of it as neural networks and the use of outside input for the improvement of software. The authors energetically argue that the program in question does not learn, but evolves, and that the relevant paradigm for it is Darwinian evolution, complete with the concept of survival of the fittest. The subject is very new. Most references are only a few years old. The book has less the feel of a textbook and more that of a discussion guide for a seminar. The exercises at the end of each chapter are more like projects than questions. The statistical tools for testing and comparing performances are laid out in great detail for empirical evaluation of the “fittest” of the programs. Ideally, the program that is to evolve will consist of many almost independent components (like genes). The testing process involves the use of trees to represent the augmentation or moderation of programming components for better performance. The type of problem to which this approach might be well suited would be image recognition, but the convincing applications seem to lie with future machines. The authors pursue the biological model to the point of including analogies to parenting, genetic crossover, ontogeny, and (of course) DNA. Such a radical approach will attract many fans, and with them, the experience leading to more refined presentations. This book will be of interest to futurists, but not to those in search of immediate answers.