A Model of Strategy for the Game of Go Using Abstraction Mechanisms

摘要

While Chess programs are almost at the level of top world players, Go programs have yet to reach the beginner's one. Particularities of Go make it unsuitable for the arborescent approaches applied to Chess. The very large branching factor (200 to 350 possible moves), the length of the games (250 moves), and the difficulty to conceive global functions of evaluation adapted to it, are sufficient to explain the weak level of Go programs. Moreover, they totally lack the strategic dimension that is essential and crucial to the reasoning of the human player. Our work presents a new approach by using the possibility of abstraction to model the elementary strategy of Go. We have modeled, and implemented in the GOBELIN system, strategic mechanisms and basic concepts of Go. To this end, we have used mechanisms of abstraction that permit the search of a solution (a plan) in a simplified (abstracted) domain. This mehtod allows us not only to avoid combinatorial crash, but also to provide the system with a reasoning level higher than the tactical one and a strategic coherent behavior close to, yet humbly, what the Go players name the global sense.