COMPUTING “ELO RATINGS” OF MOVE PATTERNS IN THE GAME OF GO 1

Abstract

Move patterns are an essential method to incorporate domain knowledge into Go-playing programs. This article presents a new Bayesian technique for supervised learning of such patterns from game records. The technique is based on a generalization of Elo ratings. Each sample move in the training data is considered as a victory of a team of pattern features. The “Elo ratings” of individual pattern features are computed from these victories, and will be used in previously unseen positions to compute a probability distribution over legal moves. In this approach, several pattern features may be combined, without an exponential cost in the number of features. Despite a very small number of training games (652), this algorithm outperforms most previous pattern-learning algorithms, both in terms of mean log-evidence (–2.69), and prediction rate (34.9%). By using these patterns, the 19 × 19 Monte-Carlo program CRAZY STONE reached the level of the strongest classical programs.