Opponent modeling by analysing play

Abstract

Opponent modeling is useful for a player both to have a better chance to win and for teaching a novice player. In this paper we discuss opponent modeling by analysing the play of the opponent. We view this as two dual tasks: characterising the play of a particular opponent and recognising an unknown opponent. We learn understandable opponent models with a logical decision tree learning system. We illustrate our approach by an experiment in the game of go and discuss the application of similar ideas in other games.