The Layered Learning Method and Its Application to Generation of Evaluation Functions for the Game of Checkers

摘要

In this paper we describe and analyze a Computational Intelligence (Cl)-based approach to creating evaluation functions for two player mind games (i.e. classical turn-based board games that require mental skills, such as chess, checkers, Go, Othello, etc.). The method allows gradual, step-by-step training, starting with end-game positions and gradually moving towards the root of the game tree. In each phase a new training set is generated basing on results of previous training stages and any supervised learning method can be used for actual development of the evaluation function. We validate the usefulness of the approach by employing it to develop heuristics for the game of checkers. Since in previous experiments we applied it to training evaluation functions encoded as linear combinations of game state statistics, this time we concentrate on development of artificial neural network (ANN)-based heuristics.