Learning Through the KRKa2 Chess Ending

Abstract

The chess ending (KRKa2) has been studied using decision trees, neural networks and human reasoning to build a classifier for this ending, and for the discovery of convenient chess attributes. These chess attributes will serve for testing new ideas in planning. The idea is to investigate whether good automatically learnt policies for a planning problem can be generated using training examples along with evolutionary algorithms. The training examples, used as input to the learning algorithm, describe specific descriptions of a number of solved instances in one domain; then to improve the learnt policies obtained from the training examples, the policies should evolve. We believe that the domain of games is well-suited for testing these new ideas in planning.