Learning of keepaway task for RoboCup soccer agent based on Fuzzy Q-Learning

Behavior learning or skill acquisition is one of the important issues of reinforcement learning schemes, in order to realize the intelligent agent. Generally, simple tasks such as goal exploration can be easily acquired by the reinforcement learning techniques, as many simulation studies are demonstrated. However, complicated tasks such as behaviors in sports like soccer are difficult to acquire substantially. It is caused by difficulties of objective modeling and multi-agent environment. In this study, we developed a behavior acquisition system for keepaway task of 2-D RoboCup soccer agent based on the fuzzy Q-learning. We showed that the Fuzzy Q-Learning approach is promising to acquire behavior rules through numerical experiments. We discussed the issues of acquisition for behavior rules in terms of improvement of the learning performances.