Ten years of intensive effort on computer chess have produced notable progress. Although the background information and technical details that were written in 1975 for the first edition of this book are still valid in most essential points, hardware and software refinements have had a major impact on the effectiveness of these ideas. The current crop of chess machines are performing at unexpectedly high levels. The approach epitomized by the series of programs developed by David Slate and Larry Atkin at Northwestern in the middle 1970s (i. e. , a sophisticated search algorithm using very little chess knowledge) was expected to reach an asymptbtic level of performance no higher than that of a class A player (USCF rating between 1800 and 2000). This perspective was argued quite vigorously by Eliot Hearst in Chapter 8 of the first edition and was held at that time by many chess experts. Subsequent events have clearly demonstrated that the asymptotic performance level for this type of pro gram it at least as high as the master level (USCF rating between 2200 and 2400). Current discussions now focus upon whether the earlier reser vations were wrong in principle or simply underestimated the asymptote. If there is a real barrier which will prevent this type of program from attaining a world championship level of performance, it is not evident from the steady progress which has been observed during the last decade.

Ten years of intensive effort on computer chess have produced notable progress. Although the background information and technical details that were written in 1975 for the first edition of this book are still valid in most essential points, hardware and software refinements have had a major impact on the effectiveness of these ideas. The current crop of chess machines are performing at unexpectedly high levels. The approach epitomized by the series of programs developed by David Slate and Larry Atkin at Northwestern in the middle 1970s (i. e. , a sophisticated search algorithm using very little chess knowledge) was expected to reach an asymptbtic level of performance no higher than that of a class A player (USCF rating between 1800 and 2000). This perspective was argued quite vigorously by Eliot Hearst in Chapter 8 of the first edition and was held at that time by many chess experts. Subsequent events have clearly demonstrated that the asymptotic performance level for this type of pro gram it at least as high as the master level (USCF rating between 2200 and 2400). Current discussions now focus upon whether the earlier reser vations were wrong in principle or simply underestimated the asymptote. If there is a real barrier which will prevent this type of program from attaining a world championship level of performance, it is not evident from the steady progress which has been observed during the last decade.

Both chess play and psychological research offer rewards to their participants in the form of intellectual satisfaction. It seems to follow that combining these two forms of activity, by carrying out research into chess play, should be a particularly engaging enterprise. In the mid-1980s enough was now known for it to be feasible to tell a reasonably satisfying story by piecing together the accumulated results of experiments on chess. There were remaining gaps in knowledge, but the structure of chess skill had at least become sufficiently evident to exhibit where the gaps lay. Originally published in 1985, this book was an attempt to summarize the progress that had been made at the time, recounting some of the components of the research process while describing how the chessplayer seems to think, imagine, and decide.

For many years I have been interested in computer chess and have collected almost every learned paper and article on the subject that I could find. My files are now quite large, and a considerable amount of time, effort and expense has been required to build up this collection. I have often thought how difficult it must be for many computer chess enthusiasts to acquire copies of articles that they see referenced in some other work. Unless one has access to a good reference library, the task is almost impossible. I therefore decided to try to make available, in one volume, as many as possible of the most interesting and important articles and papers ever written on the subject. Such a selection is naturally somewhat subjective, and I hope that I will not offend authors whose works have been excluded. In particular I have decided to exclude any material which has appeared in the Journal of the International Computer Chess Association (ICCA), or in its precursor, the ICCA Newsletter. The reason is simply that the ICCA itself is in the process of compiling a compendium containing the most important material published in those sources. For further information on ICCA membership and publications the reader is invited to contact: Professor H. 1. van den Herik, or Dr Jonathan Schaeffer University of Limburg, Computing Science Dcpaitment, Department of Computer Science University of Alberta, 6200 MD Maastricht Edmonton Netherlands Alberta, Canada T6G 2HI.

Research in artificial intelligence, natural language processing and knowledge-based systems has blossomed during the past decade. At national and international symposia as well as in research centers and universities all over the world, these subjects have been the focus of intense debate and study. This is equally true in Israel which has hosted several international forums on these topics. The articles in this book represent a selection of contributions presented at recent AI conferences held in Israel. A theoretical model for a system that learns from its own experience in playing board games is presented in Learning from Experience in Board Games by Ze'ev Ben-Porat and Martin Golumbic. The model enables such a system to enhance and improve its playing capabilities through the use of a learning mechanism which extracts knowledge from actual playing experience. The learning process requires no external guidance or assistance. This model was implemented and tested on a variant of "Chinese Checkers. " The paper shows the feasibility and validity of the proposed model and investigates the parameters that affect its performance traits. The experimental results give evidence of the validity of the model as a powerful learning mechanism. Original and general algorithms for knowledge extraction and pattern matching were designed and tested as part of the prototype computer system. Analysis of the performance characteristics of these algorithms indicates that they can handle large knowledge bases in an efficient manner.

A synopsis of eminent computer chess programs reveal that they are designed around a 'brute force' approach. An argument is made that by continuing the 'brute force' search approach, computer chess development is moving away from human evaluation methods. Research is done into studies of evaluation methods, and a discovery is made that humans use a form of intuition, called their 'sense of beauty', to choose the best chess move. A paper by Margulies is cited which formulates principles of beauty which apply to chess. Three versions of a chess program are developed, using no heuristics, standard chess heuristics, and beauty heuristics formulated from Margulies principles. The performance of the three versions of the program are compared using chess puzzles, and rated for how quickly they find the solution, and how few nodes they evaluate. Graphs are produced from the results of these tests, showing that beauty heuristics are, on average, 15% faster at finding the solution, and evaluate 10% fewer nodes. An improvement is implemented in all versions of the program which biases the search towards better moves, resulting in the beauty heuristics success rising to an average of 25% faster to the solution, and evaluating 33% fewer nodes, than the other heuristics. It is concluded that the beauty heuristics are closer to the way that humans evaluate chess positions.

Problem solving is a central topic for both cognitive psychology and artificial intelligence (AI). Psychology seeks to analyze naturally occur ring problem solving into hypothetical processes, while AI seeks to synthesize problem-solving performance from well-defined processes. Psychology may suggest possible processes to AI and, in turn, AI may suggest plausible hypotheses to psychology. It should be useful for both sides to have some idea of the other's contribution-hence this book, which brings together overviews of psychological and AI re search in major areas of problem solving. At a more general level, this book is intended to be a contribution toward comparative cognitive science. Cognitive science is the study of intelligent systems, whether natural or artificial, and treats both organ isms and computers as types of information-processing systems. Clearly, humans and typical current computers have rather different functional or cognitive architectures. Thus, insights into the role of cognitive ar chitecture in performance may be gained by comparing typical human problem solving with efficient machine problem solving over a range of tasks. Readers may notice that there is little mention of connectionist ap proaches in this volume. This is because, at the time of writing, such approaches have had little or no impact on research at the problem solving level. Should a similar volume be produced in ten years or so, of course, a very different story may need to be told.