Learning Evaluation Functions of Shogi Positions from Different Sets of Games

This paper addresses learning of a reasonably accurate evaluation function of Shogi (Japanese Chess) positions through learning from records of games. Accurate evaluation of a Shogi position is indispensable for a computer Shogi program. A Shogi position is projected into several semantic features characterizing the position. Using such features as input, we employ reinforcement learning with a multi-layer perceptron as a nonlinear function approximator. We prepare two completely different sets of games: games played by computer Shogi programs and games played by professional Shogi players. Then we built two evaluation functions by separate learning based on two different sets of games, and compared the results to find several interesting tendencies.