Learning an evaluation function for shogi from data of games

Abstract

This paper proposes a method for obtaining a reasonably accurate evaluation function of a shogi (Japanese chess) position through learning from data of games. An accurate evaluation function is indispensable for a strong shogi program. A shogi position is projected into a feature space which consists of feature variates charactering the position. Using such variates as input, we employ a multi-layer perceptron as a nonlinear evaluation function. Since it is not easy to obtain accurate evaluated values of positions, we employ reinforcement learning. Our experiments using hundreds of games show that the proposed method works well in obtaining a very accurate evaluation function for shogi, whose performance is comparable to that of a strong shogi program. © Springer-Verlag 2004.