Collaborative Research: Solving Chess with Probabilistic Planning and Control

点击翻译按钮获取中文摘要

This is collaborative research with Hans Berliner at Carnegie Mellon University (IRI-9105202). Game-playing is an important domain for two reasons: first, it exhibits in a clean form two of the major problems facing AI systems, namely how to pay attention to only a small fraction of the possibilities in complex domains, and how to use a large quantity of evidential information to make decisions; second, it has long been considered a litmus test for ideas about problem solving and for the quality of AI ideas in general, particularly in comparison to human intelligence. Recent work on control of search by Russell and Wefald and on probabilistic inference in problem solving domains by Hansson and Mayer, suggests that the PI's will be able to break the grip of brute-force methods on high-performance systems. The object of the proposed research is to test out these ideas, with a view to demonstrating feasibility of an intelligent chess program with the potential to perform at a high level. Should the effort succeed, the PI's will continue with a concerted attempt to defeat the human world champion. The purely scientific benefits should be significant, in terms of demonstrating the feasibility of large-scale probabilistic inference and metalevel control; the benefits to the field of AI in the larger national context will be immeasurable, since several influential critics will be disarmed, and success will avoid the extremely deleterious conclusion the "all your need for AI is faster chips".