Candidate Relevance Analysis for Selective Search in Shogi

Grimbergen, R. (1999)

9th Advances in Computer Chess Conference, Paderborn, Germany

Abstract

In this paper we present Candidate Relevance Analysis (CRA), a method for plausible move generation in two-player perfect information games. Plausible move generation has not been widely used because of the risks of losing important search candidates. However, there are search domains where the number of search candidates is too large to do a complete search. CRA limits the number of search candidates by analysing the relevance of the candidates with respect to the goal of a game, discarding all candidates judged irrelevant. CRA can easily be incorporated in alpha-beta search. Furthermore, the knowledge acquired by CRA can be used to improve the move ordering, thus improving the efficiency of alpha-beta search. In the domain of shogi (Japanese chess), CRA reduces the number of candidates by 39.7%, while maintaining an accuracy of 99%. Also, in shogi alpha-beta search with CRA outperforms full-width alpha-beta search even for small search depths.