Application of a adaptive-learning situation evaluation function based on Feature-Matrix in Dots-and-Boxes

This paper proposes a method for obtaining a reasonably accurate evaluation function of a game situation through the data of games and the situation feature matrix. An accurate evaluation function is indispensable for a strong computer game program. A game situation is projected into a feature matrix which consists of feature variates charactering the situation. Using variates as input and employ a multi-layer perception as a nonlinear evaluation function. Since it is not easy to obtain accurate evaluated values of situations, the reinforcement adaptive-learning is employed. Experiments using 134 games show that the proposed method works well in obtaining a very accurate evaluation function for Dots-and-Boxes.