Browse Books

Reviewer: H. Van Dyke Parunak

While the market regularly delivers dramatic increases in the power of computer hardware, software is often the bottleneck in fielding new systems. As systems become more complex, software developed using traditional techniques takes longer to design, is more likely to contain bugs, and is more difficult to maintain. Synthetic evolution is one of the most promising avenues for overcoming the software bottleneck. The basic idea is to change the essence of software from artifact to life form, and shift the role of the software engineer from builder to breeder. As early as the 1950s, some researchers were exploring such mechanisms. However, the emergence of symbolic computation as the dominant paradigm for intelligent programs left these efforts in a backwater from which they have only recently been emerging. In recent years, one of these approaches, the genetic algorithms of John Holland, has attracted considerable attention and imitation. Fogel's book will introduce readers to another approach, originated by his father in the early 1960s and developed by the continued efforts of the author and J. W. Atmar. The volume relates Fogel's approach, called evolutionary programming, to biological evolution, alternative applications of evolution to software, and the general quest for artificial intelligence. It offers a carefully organized, integrated exposition of a large body of research originally published elsewhere. Chapter 1, “Defining Artificial Intelligence,” picks up the debate where it left off when symbolic programming gained preeminence. Fogel reviews the Turing test and considers a number of projects that attempt to simulate human expertise. He also discusses neural networks before proffering his own definition of intelligence: the capability of a system to adapt its behavior to meet its goals in a range of environments. He argues that insight can be gained by tracing the commonalities among three different organizational forms of intelligence: the individual organism (ontogenetic intelligence); a sequence of organisms descended from one another (phylogenetic); and a group of organisms concurrently interacting with one another (sociogenetic). Traditional AI focuses on ontogenetic mechanisms. Fogel seeks to generalize and exploit phylogenetic mechanisms. Chapter 2, “Natural Evolution,” provides an overview of leading research questions in theoretical biology. A critical section of the chapter focuses on the question of the beneficiary of evolution: is it the individual gene, the genome of the individual organism, or the reproducing population of which the individual is a part__?__ Fogel favors a variety of the third option. The discussion is important because the main difference between his approach and the better-known genetic algorithms is explicit focus of the latter on the gene and genetic manipulations as an implementation strategy. Chapter 3, “Computer Simulation of Natural Evolution,” surveys early attempts to mimic evolution in computer software, and summarizes four current approaches. Fogel's evolutionary programming applies mutation to populations of contending algorithms (often represented as finite-state automata). Rechenberg and Schwefel's evolution strategy applies similar mechanisms to vectors in a multidimensional space in order to optimize continuous functions. Holland's genetic algorithms place less emphasis on mutation, and highlight the genetic operations of crossover and inversion. A variety of workers have explored the dynamics of evolution in programs such as Tom Ray's Tierra. Chapter 4, “Theoretical and Empirical Properties of Evolutionary Computation,” discusses whether the additional complexity of crossover (compared with pure mutation) is worth the trouble. It begins by analyzing the convergence properties of genetic algorithms and evolutionary algorithms using Markov chains. The complex nonlinear stochastic processes inherent in evolutionary models are difficult to model without making simplifying assumptions, so Fogel also reports on benchmark experiments comparing different variations of crossover with one another and comparing crossover with mutation. He concludes that crossover offers no consistent advantage. Chapter 5, “Intelligent Behavior,” illustrates the application of Fogel's methods to three problems: evolving linear equations to balance a pole on a rolling cart, evolving finite-state machines to play the prisoner's d<__?__Pub Caret>ilemma with one another, and evolving a neural network to play tic-tac-toe. Chapter 6, “Perspective,” returns to the question of intelligence, and argues that the ability to anticipate change in the environment, as evidenced in these experiments, is a hallmark of intelligence. Fogel contrasts AI's simulation of specific facets of intelligent behavior with the ability of evolutionary machines to emulate the process of life itself, resulting in an intelligence that is not at all artificial. The volume includes a glossary of biological terms that will be helpful to computer scientists new to the field, and a brief index. There is no integrated bibliography. Instead, each chapter closes with a list of references cited in that chapter. Fogel's prose is terse and academic, but the overall argument flows smoothly, and the book is a coherent and well-documented argument for Fogel's case as well as a useful compendium of the various flavors of research on evolutionary programming.