Applying GA to self-allotment of rewards in event-driven hybrid learning classifier systems

This paper describes our study into the concept of using rewards in a classifier system applied to the acquisition of decision-making algorithms for agents in a soccer game. Our aim is to respond to the changing environment of video gaming that has resulted from the growth of the Internet, and to provide bug-free programs in a short time. We have already proposed a bucket brigade algorithm and a procedure for choosing what to learn depending on the frequency of events with the aim of facilitating real-time learning while a game is in progress. We have also proposed a hybrid system configuration that combines existing algorithm strategies with a classifier system, and we have reported on the effectiveness of this hybrid system. This paper proposes applying genetic algorithms to the search for rewards in reinforcement learning where designers have hitherto used empirical trial-and-error methods. By pitting this new technique against an existing soccer game with algorithms designed by humans, we demonstrate the possibility of using genetic algorithms to automate the setting of rewards from the viewpoint of achieving a greater success rate and faster convergence than in cases where the success rewards for each play are set by the designers based on trial and error.