Heuristic Interpretation of Predicate Logic Expressions in General Game Playing

Unlike traditional game playing, General Game Playing (GGP) is concerned with agents capable of playing classes of games. Given the rules of an unknown game, the agent is supposed to play well without human intervention. For this purpose, agent systems using game tree search need to auto-matically construct a state value function to guide search. This state value function is often either probabilistic as in Monte-Carlo Systems and thus most likely unable to com-pete with deterministic functions in games like chess; or it is expensive in construction due to feature generation and learning-based selection and weighting mechanisms. In this work we present an alternative method that derives fea-tures from the goal conditions stated in the game rules, avoid-ing thereby the disadvantages of other approaches. The paper structures and generalizes some known approaches, shows new ways of deriving features and demonstrates how to use these features as part of an evaluation function. Experiments demonstrate both a high effectiveness and generality.