GENERAL GAME-PLAYING AND REINFORCEMENT LEARNING

This paper provides a blueprint for the development of a fully domain- independent single-agent and multiagent heuristic search system. It gi ves a graph-theoretic representation of search problems based on conce ptual graphs and outlines two different learning systems. One, an ''in formed learner'', makes use of the graph-theoretic definition of a sea rch problem or game in playing and adapting to a game in the given env ironment. The other, a ''blind learner'', is not given access to the r ules of a domain but must discover and then exploit the underlying mat hematical structure of a given domain. Relevant work of others is refe renced within the context of the blueprint. To illustrate further how one might go about creating general game-playing agents, we show how w e can generalize the understanding obtained with the Morph chess syste m to all games involving the interactions of abstract mathematical rel ations. A monitor for such domains has been developed, along with an i mplementation of a blind and informed learning system known as MorphII . Performance results with MorphII are preliminary but encouraging and provide a few more data points with which to understand and evaluate the blueprint.