Chess Neighborhoods, Function Combination, and Reinforcement Learning

Abstract

Over the years, various research projects have attempted to develop a chess program that learns to play well given little prior knowledge beyond the rules of the game. Early on it was recognized that the key would be to adequately represent the relationships between the pieces and to evaluate the strengths or weaknesses of such relationships. As such, representations have developed, including a graph-based model. In this paper we extend the work on graph representation to a precise type of graph that we call a piece or square neighborhood. Specifically, a chessboard is represented as 64 neighborhoods, one for each square. Each neighborhood has a center, and 16 satellites corresponding to the pieces that are immediately close on the 4 diagonals, 2 ranks, 2 files, and 8 knight moves related to the square. Games are played and training values for boards are developed using temporal difference learning, as in other reinforcement learning systems. We then use a 2-layer regression network to learn. At the lower level the values (expected probability of winning) of the neighborhoods are learned and at the top they are combined based on their product and entropy. We report on relevant experiments including a learning experience on the Internet Chess Club (ICC) from which we can estimate a rating for the new program. The level of chess play achieved in a few days of training is comparable to a few months of work on previous systems such as Morph which is described as “one of the best from-scratch game learning systems, perhaps the best” [22].