Predictive Sub-goal Analysis in a General Game Playing Agent

Abstract

General Game Playing (GGP) research aims at designing intelligent game playing agents that, given the rules of any game, automatically learn strategies to play and win without human intervention. Our GGP agent can play the wide variety of heterogeneous games provided by the IJCAI GGP competition framework, and without human intervention, learn from its own history to develop strategies toward achieving the game goals. It uses statistical analysis to identify important game features shared by the winners. To illustrate how the correct features are identified, we use game examples from different game categories, including Tic-Tac-Toe (territory taking game), Mini-Chess (strategy game), and Connect Four (board game on larger scale).