Contextual Decision Making in General Game Playing

General Game Playing refers to designing Artificial Intelligence agents that are capable of playing different games without human intervention. The games are defined by sets of rules represented in logic descriptions and the agent players interact in a multi-agent system with a game server coordinating the legality of the operations and keeping the players informed of the state changes. This paper describes a general game agent that isolates the heuristic search coverage for contextual decision making by efficiently creating dynamic decision trees. The influence of certain game features is evaluated within the current decision context rather than on the whole game scale. The benefit of this approach is shown by performance comparison with agents that do search, learning, and learning with decision trees. We show this for a variety of games and have compared favorably against well known general game players and replicated actions of known human expert strategies.