| Research Abstract |
Chess programs, which were a target of game programming in Al, heve so far adopted brute force search. However, Go, which is now regarded as a new Al target, is different from Chess in that the former has a far larger search space than the latter. It is thus considered for Go programs to adopt a new approach, that is, knowledge-intensive approach. We, therefore, constructed a system which enabled to antomatically acquire various types of Go knowledge that human Go experts are considered to have.
First, we proposed a new evolutionary algorithm which was similar to a genetic algorithm. A normal genetic algorithm has two disadvantages. One is that the algorithm gets only a few types of rules which are optimal. The other is that it tends to get knowledge with fixed forms. Activation values to individuals and a new variation-making operator, spiltting, are introduced to this new algorithm to overcome these disadvantages. As a result, the proposed algorithm could automatically get various types of Go knowledge.
Then we evaluated this algorithm by applying it to the game of Go. As a result, it was found that the algorithm could acquire various pattern knowledge and, moreover, get sequences of moves by adding previous moves in splitteing a rule. We asked two Go experts to evaluate some of the acquired rules. As a result, we found from their judgemeny that 60-80 % of the rules presented were useful. In addition, we made our system to solve Tsume-Go problems by using the acquired knowledge of Tsume-Go and found that the performance of the system was almost the same as that of human sho-dan. The system is, therefore, considered to enable to acquire various useful Go knowledge.
|
