Evaluation Tuning for Computer Chess: Linear Discriminant Methods 1

Abstract

The evaluation function of a software version of the Deep Thought chess program is tuned against a database of master-level games. The agreement between the move chosen by the program and by a human master is maximized. A set of algorithms solving the maximization problem is presented. The trade-off between computational speed and the chance of getting stuck in a local maximum is extensively investigated.

The performance of the chess program with the tuned evaluation function is experimentally measured. The program has played several matches of 500 games each with different settings against a fixed program. The results show that 98% of the performance of the evaluation function can be achieved in about a week by tuning the evaluation function from scratch against the database. If the evaluation tuning is started from the best evaluation used so far, the new evaluation function improves the playing strength by 34 rating points.