Inductive Inference of Chess Player Strategy

We investigate the problem of inferring, from records of chess games, some aspects of the strategy used to play the games. Initially, game records are generated from self-play by two simple chess programs, one of which does a one-ply search while the other does a four-ply quies-cent search. In each case, we are able to infer, from just the game records, good estimates of the weights used in the evaluation function. The approach is then applied to grandmaster games. Our one-ply and quiescent four-ply programs are now drastic simpliications of the true strategy used. Nonetheless, using inferred weights for these hypothetical models, we are still able to achieve some success (as measured by compression rates for the games) in predicting moves made by the players.