Learning to focus selectively on possible lines of play in checkers

Ongoing work in coevolution of neural networks for checkers play is described. An unusual aspect is that the fitness of a neural net is not reduced to a scalar fitness score. Instead, nets are sorted primarily in non-decreasing order of number of games lost to other members of the population, and secondarily in non-increasing order of number of wins. The first n nets in the sorted list are selected as parents. The main thrust of the present work, however, is to learn to prioritize the extension of possible lines of play in the minimax game tree. Neural nets not only assign static values to leaf nodes, as is common in minimax play, but priorities for dynamic evaluation, as well. At each iteration of the game tree evaluation, the highest-priority leaf node of the current game tree has its children generated. Thus the conventional minimax control strategy is replaced by one that seeks to extend lines of play selectively. Although none of the neural-net players is rated yet, tournaments pitting later generations against earlier indicate that quality of play is improving.