Book Reviews Genetic Algorithms + Data Structures = Evolution Programs by Zbigniew Michalewicz Springer-Verlag, 1992 Hard cover, 250 + xiv pages ISBN 3-540-55387-8 Reviewed by: Michael de la Maza Numinous Noetics Group MIT AI Lab Cambridge, MA 02139 Like a good politician, Michalewicz begins by reducing expectations (he quotes Anthony de Mello: "[The Master] only points the way - he teaches nothing") and follows with a book that is a hodgepodge of introductory material and original work. The subject of the book, as the title suggests, is the study of evolution programs, which are genetic algorithms augmented by novel genetic operators and domain dependent d a t a structures. Indeed, a better title for the book would be: Some of the errors in the book are serious. The genetic algorithm selection function, which assigns to each individual the probability that it will participate in the formation of the next generation, requires that the fitness of each individual be positive. At the beginning of chapter 2, Michalewicz writes that any bounded fitness function that assigns negative fitnesses to individuals can be changed into a function that is always positive by adding a positive constant to the function. This is undoubtedly true but, unfortunately, the most widely used selection function in genetic algorithms is not invariant with respect to this transformation. Choosing a constant is an important decision that can greatly impact the search behavior of the genetic algorithm, not a simple cosmetic change as Michalewicz seems to imply. Later, in chapter 4, Michalewicz briefly mentions that extreme choices of the constant change search behavior, but he does not connect this observation with his earlier discussion in chapter 2. After reading this book I am left with the feeling that Michalewicz has some excellent ideas but the exigencies of publishing forced him to couch them in a barely palatable form. The result is a book that is not as illuminating as a beginner would like it to be and not as incisive as an expert would like it to be. The expert reader is advised to skip or skim the first six chapters and start with the seventh, which contains the bulk of the original material in the book. The last section of the book is a set of concatenated review papers and should be read as such. The reader unfamiliar with evolution programs may want to begin with a short introduction to genetic algorithms, such as ch. 25 in Winston, and then use Michalewicz's book to acquire deeper understanding. Both the novice and the advanced reader will want to digest the excellent conclusion. Genetic Al9orithms * (Genetic Operators + Data Structures) = Evolution Programs. The book is divided into three parts: genetic algorithms, numerical optimization, and evolution programs. The first section, which spans the first four chapters, is a brief introduction to the theory and practice of genetic algorithms. Michalewicz quotes extensively from other authors in these opening chapters. Presumably this is not to "save the subject by magnificent quotations," in the words of T.S. Eliot, but rather to introduce new ideas parsimoniously and to give credit to the developers of the ideas. Unfortunately, these quotations are from seminal papers in the field and are not aimed at the novice reader, who Michalewicz says is part of the intended audience of the book. Those wanting a truly gentle introduction to genetic algorithms should consult ch. 25 in Pat Winston's 1992 book. Goldberg gives a more thorough introduction in his 1989 text. The second section contains chapter 7, the best chapter in the book. Michalewicz is past the introductory material, he is writing about what he is interested in, and it shows. The seventh chapter describes how to modify evolution strategies to handle the type of constraints found in numerical optimization. Michalewicz acknowledges and dismisses previous attempts to address this problem and then tells the reader the main thesis of the book: d a t a structures and genetic operators can be designed to enforce constraints in an efficient, extensible, and elegant way. The last section of the book supports this thesis by reviewing recent work on a variety of problems. An entire chapter is devoted to the traveling salesman problem which has stimulated much creative work in d a t a representation and genetic operators. Computational Morphology Review of v l reensts texts from by: Robert P. Goldman Computer Science Department Tulane University Richard Sproat, Morphology and Computation, ACL-MIT Press Series in Natural Language Processing, Bradford Books, The MIT Press, 1992, 295 + xv, ISBN 0-262-19314-0. $35.00. Graeme D. Ritchie, Graham J. Russell, Alan W. Black and Stephen G. Pulman, Computational Morphology: Practical Mechanisms ]or the English Lexicon, A C L - M I T Press Series in Natural Language Processing, Bradford Books, The MIT Press, 1992, 291 + x, ISBN 0-262-18146-0. $32.50. S I G A R T Bulletin, Vol. 4, No. 2

/lp/association-for-computing-machinery/book-review-computational-morphology-review-of-two-recent-texts-from-HLRRW8xzbf