Flexible Acquisition of Various Types of Knowledge from Game Records: Application to the game of Go

A large amount of knowledge is considered very useful for systems playing games such as Go and Shogi (Japanese chess), for which the search space larger than that for chess. Knowledge acquisition is therefore necessary for systems playing these games. This paper introduces a new exible algorithm which acquires various types of knowledge from game records. This algorithm is especially applicable to Go because Go has the largest search space among major games. The algorithm enables computer systems to acquire patterns with a variety of shapes and sizes. Sequences of moves, which are useful for human players but lacking in the present game-playing systems, are also acquired by this algorithm. The possibility of acquiring other types of knowledge | such as maxims, patterns with situations, and patterns for tsume-go | is discussed. The possibility of applying the algorithm to games other than Go is also discussed.