Connectionist Learning of Expert Backgammon Evaluations

A class of connectionist networks is described that have learned to evaluate the strength of backgammon moves. The networks were trained by “back-propagation” learning of a large set of sample positions evaluated by a human expert. A game-playing system using such a network evaluation function achieves an intermediate-to-advanced level of performance, and is significantly better than the best available commercial program. In order to achieve this level of performance, certain novel protocols for the training procedure and input coding scheme design were required. These protocols are expected to be of general utility in applications of network learning to large-scale “real-world” problems.