Reviews Cellular Automata: Theory and Experiment Edited by H o w a r d Gutowitz Cambridge, MA: M I T Press. Pp. xvii + 483. $37.50. Reviewed by James Reggia D e p a r t m e n t of C o m p u t e r Science University of Maryland College Park, M D 20742 this book of some interest: individuals concerned with parallel processing, learning or neural modelling. Cellular automata provide a massively parallel computational architecture, and can be viewed as the conceptual basis for at least one prominent parallel "supercomputer" directed in part at the AI community [Hillis D, The Connection Machine: A Computer Architecture Based on Cellular Automata, Physica IOD, 1984, 213-228]. In this context many of the theoretical papers in this book concerned with the behavior, properties and classification of cellular automata are of interest. These papers address the issues of computability and universality of different classes of cellular automata. Many also examine cellular automata viewed as discrete dynamical systems, characterizing the attractors, phase transitions, linearity, reversibility, periodicity, etc. of certain classes of these machines. One of this book's strongest points is that it brings together in one place so many theoretical contributions, with these papers being roughly half the papers in the book. A second smaller group of papers are of interest because they deal with learning in cellular automata. Two of these papers [by Y.C. Lee et al, and S. Qian et al: Adaptive Stochastic Cellular Automata ..., Parts I and II] provide both theoretical and experimental analysis of an approach to creating adaptive stochastic cellular automata. This approach to adaptive machines is closer to the "bottom-up" learning methods found in neural network models than it is to the traditional "top-down" symbol manipulation programs in mainstream AI. It relaxes the standard requirement of rule homogeneity throughout a cellular automata space to require only statistical homogenity. Using reinforcement learning, the authors demonstrate how cellular automata models can learn control tasks such as the pole-balancing problem. Another interesting paper [E Richards et al, Extracting Cellular Automata Rules Directly from Experimental Data] addresses the problem of "discovery" of cellular automata models that simulate physical processes. These authors demonstrate how a genetic algorithm can be used to derive local rules that model the dendritic solidification of NH4Br from a supersaturated solution. While this approach is presented in the context of a specific application, the techniques involved should be useful in general for learning models of spatial pattern evolution. Finally, two papers [Victor; Garzon] explore the relationship between cellular automata models and neural modelling, with neural/connectionist models viewed as a generalization of cellular automata models. These papers do present some new ideas, but are limited in that the authors seem unaware of several previous neural models that capture the discrete space and time features of cellular automata. In summary, this collection of papers covers a broad spectrum of issues related to self-processing, self-organization, massive parallelism, and emergent behaviors in the context of local interaction rules. While most of the papers are not of direct interest to individuals in AI, some are relevant to issues of parallel processing, learning and neural networks. This is a fine collection that provides a good introduction to contemporary work with cellular automata, and it has many pointers to relevant previous work. reggia@cs.umd.edu Cellular automata can be traced back to the 1950's when John yon Neumann used them to investigate the possibility of creating self-replicating machines. Today, cellular automata models of various dimensionality find applications in a wide range of technical and scientific disciplines. Von Neumann's original cellular automata model consisted of a two-dimensional uniform cellular space that resembles an infinite checkerboard. Each cell (square on the checkerboard) represents a small piece of the two dimensional space, and effectively serves as a finite state automaton that can be in one of a finite number of states usually designated 0, 1, 2 , . . . , n. State zero is often called the "quiescent" or "empty" state. At each instant of simulated time, each cell changes state as a function of its own current state and the state of immediately neighboring cells. State changes are governed by a set of rules which are collectively referred to as the transition function. Each individual rule is simple and based solely on locally-available information. However, as with many rule-based AI programs, the complete set of rules forming the transition function, through their application by all cells in the model simultaneously and repetitively over time, can produce very rich and at times striking behavior. Patterns of non-quiescent activity can form, spread, move, break apart, merge and in general self-organize in aesthetically pleasing ways that in some cases can serve as models for phenomena observed in nature. The book CelhdarAutomattz" Theoo, and Experiment is a collection of papers reprinted from the journal PhysicaD, which in turn were based on a conference of the same name held September 1989 at Los Alamos National Laboratory. The 35 papers are for the most part of high quality and range from highly theoretical to innovative applications. The papers are grouped into sections dealing with mathematical analysis, computational theory, characterization of variations and generalizations of cellular automata, adaptation, and applications to the natural sciences. Two appendices survey cellular automata software packages and provide a brief introductory overview of recent cellular automata literature. The obvious question one might have at this point is why such a book was sent to SIGART Newsletter What possible relevance could this material have to AI? The answer is that, for most individuals in AI, the material in this text is not of direct relevance. However, there are three groups of people who may find parts of S I G A R T Bulletin, Vol. 3, No. 3

/lp/association-for-computing-machinery/book-review-cellular-automata-theory-and-experiment-edited-by-howard-qglp9QloBj