Using Decision Trees for State Evaluation in General Game Playing

Abstract

A general game playing agent understands the formal descriptions of an arbitrary game in the multi-agent environment and learns to play the given games without human intervention. In this paper, we present an agent that automatically extracts common features shared by the game winners and uses such learned features to build decision trees to guide the heuristic search. We present data to show the significant performance improvements contributed by the decision tree evaluation. We also show by using hash tables in knowledge reasoning, our agent uses 80% less time when compared to a widely available GGP agent written in the same language.