Browse Books

Reviewer: Louis Hodes

Genetic programming is part of artificial intelligence, specifically part of automatic programming. Since programming is considered more of an art than a science, it is not surprising that all the dozens of problems Koza tackles are specially invented impractical problems. This limitation is never explicitly expressed by Koza in this book or his earlier, equally large book on genetic programming [1]. There is no demonstration that extension of these methods will ever solve practical problems. I distinguish two kinds of potential readers of this book. One kind is somewhat familiar with this field. They know that genetic programming is not expected to produce programs for real problems. The other kind is more innocent and may think of genetic programming in connection with mathematical programming. They are looking for methods to apply to some practical task. It is this second group of potential readers who need to be informed of the limitations of genetic programming. For example, in molecular biology, Koza deals with the problem of detecting transmembrane regions in amino acid sequences of proteins. He describes the problem well and shows how it can be solved by scanning the sequence and, for each amino acid, computing the sums of the hydrophobicity indices of the 5 amino acids on both sides; more precisely, adding the indices of the 11 closest amino acids at each sequence position. These sums show up the transmembrane regions prominently. The programs generated by genetic programming have no way to develop the capability to solve this problem, so they are given a data set that simplifies the task. They are merely required to distinguish a transmembrane region from another randomly selected region of the protein of equal sequence length. Such a problem is not likely to arise in the real world. Thus, Koza is dealing with a problem that looks interesting but is not practical. When he compares his results with “conventional methods,” he is referring to other artificial intelligence methods on this simplified data set that do not use genetic programming. As in the earlier book, Koza expresses himself well and has some fascinating ideas. This work supersedes the earlier book by adding the capability of automatically defined functions (ADFs), or subroutines, in a syntactically appropriate way. In my review of the earlier book, I observed more conventional practice by summarizing the book briefly. This book follows similar lines, with more varied applications and the additional need to compare results without ADFs to the newer approach. ADFs improve results in tasks that can be broken down into easier subtasks. The book is well done and describes a great deal of work. The reader should be warned, however, that these methods are at the research stage. One disappointment for me was there was no longer an attempt to simultaneously learn two behavioral outcomes, as in his earlier book. I would think this could be done with subroutines. The video, as in the earlier work, is a shortcut for reading the book. It is again misleading in the transmembrane problem when it claims “better performance than human-written algorithms,” since it refers to the specially constructed classification problem discussed earlier. The lack of mention that these studies are experimental may lead to an adverse reaction by people who try to use these methods on real problems.