Automatic Feature Construction and Optimization for General Game Player

In this paper, we describe our method that automatically constructs evaluation functions without any human analysis of a target game. Such automated constriction of evaluation functions is crucial to develop a general game player that can learn and play an arbitrary instance of a certain class of games. Our approach is to construct features written in logic programs from the game definition and translate them into specialized evaluator in order to get such efficiency that learning methods can try and test so many features that it produces accurate evaluation functions. We also introduced the decomposition of logical features called thin features in order to improve both accuracy and efficiency. Experiments on Othello endgames show that the accuracy and efficiency of evaluation functions generated by our method are approaching to those of the patten based evaluation function which is the state-of-the-art technique. keywords: automatic feature construction, logical feature