Pattern Recognition for Candidate Generation in the game of Shogi

Grimbergen, R. and Matsubara, H. (1997)

Proceedings of the Workshop on Computer Games (W31) at IJCAI-97, pp. 7--12, Nagoya, Japan

Abstract

Chess has been an important research area in Artificial Intelligence for decades. It seems that the strongest programs will soon play better than the best experts, so it is time to look at other game domains as possible test domains for AI research. We feel that shogi is well suited for this, because it is similar to chess, yet significantly different. In this paper we give the main differences between chess and shogi and possible solutions for problems arising from these differences. It is our assumption that all problems can be solved using pattern recognition. We describe a method for pattern recognition to do this and give some preliminary results for candidate generation using pattern recognition. For candidate generation, our pattern recognition program currently reduces the number of candidates generated by at least 77\% and generates the optimal move in more than 75\% of the test positions, using about 5,000 patterns. To play shogi at a high level these numbers need to be improved upon, but as a preliminary result this is encouraging.