Learning Piece-square Values using Temporal Differences

Abstract

This paper reports on the results obtained from using improved temporal difference learning methods to learn piece-square weight sets for use in a chess program. The learning takes place solely from self-play, starting from zero values. A comparison is made between values learnt from piece weights only, and piece weights plus positional weights. The weight sets obtained, when displayed as grey-scale diagrams matching chessboards, can be visually seen to correspond to various items of simple chess knowledge of the type found in elementary chess books, and regarded as basic information for beginning chess players. The paper also considers the effect of the squashing function used to map evaluations into probabilities-to-win.