EVALUATION OF PATTERN SHAPES IN BOARD GAMES BEFORE MACHINE-LEARNING

摘要

Pattern matching is an important method to incorporate domain knowledge into board game programs, and it is useful for semi-random move selection in Monte-Carlo simulations. Each possible move is evaluated with the surrounding patterns of cells to decide the probability of play in a simulation. However, an important problem is to choose the best pattern shapes. Evaluating whether a pattern shape contains informative patterns is necessary before performing a costly machine learning on multiple game records. In this paper, we present a method to evaluate the importance of a pattern shape in a game board. From that, important pattern shapes can be extracted automatically from large, high quality data sets of game records. In experiments, we implement the proposed method in a data set of Othello game records, and use a state-of-the-art machine learning method, Bradley Terry Maximization-Minorization (BTMM), to evaluate the results.