Rule-Based Reasoning Via Abstraction

Abstract

The General Game Playing (GGP) project seeks to create software agents capable of receiving game rules hitherto unseen and, without human interaction, play that game effectively. For the time being, we have chosen to restrict the class of games to be multi-player, finite, deterministic and complete information. (This is precisely the class of games that can be modeled by a finite-state machine.) Due to the finiteness and deterministic restrictions, given the rules of a game it is in principle possible to completely expand the game model and reason about it using model-checking or traditional graph search techniques. In practice, however, the games can be intractably large, making these techniques inadequate; games might also have special structure or properties implied by the rules, which if discovered could mean considerable shortcuts in the search. For this reason, we prefer reasoning about the game rules and only expand the game tree when absolutely necessary, using any insight gained from the rules.