Applying GA for Reward Allotment in an Event-driven Hybrid Learning Classifier System for Soccer Video Games

摘要

This paper describes our study applying GA to search the reward values for reinforcement learning in a soccer video game using learning classifier systems. In particular, we report the result of promotion of efficiency by dividing searched space and searching the divided space alternately. We have already proposed that acquiring algorithms by using the event-driven hybrid learning classifier system. Moreover, we have proposed that using GA for setting the reward values which have no index for setting. As the result, a probability that the reward values can be set to appropriate value for learning was obtained. In prior studies, certain rewards became searching candidate. Meantime, if the kind of the success rewards increase, setting of success rewards is difficult in practical time because the search space of reward values become spread. To address this problem, in this paper, we propose that applying the technique of dividing the searched space, and searching divided space alternately with exchange information. By comparison with the technique of searching the reward values all at once, we show a possibility that this technique have effect to improve efficiency of learning the reward values.