Host: Japan Society for Fuzzy Theory and Intelligent Informatics (SOFT)
The introduction of the reinforcement learning to the multi-robot systems is researched. The performance of reinforcement learning is quite sensitive to the segmentation of state and action spaces. To overcome this problem, we have been developing a new technique, Bayesian-discrimination-function-based RL (BRL). BRL is capable of the adaptive division of the state space. In BRL, the over-fitting occasionally occurred. This paper introduces an extended BRL method to restrain over-fitting. The extended BRL method is the technique that introduced meta learning into BRL. This method coordinates learning rate to the learning degree of progress, and uses rule ignition entropy for an index of the learning degree of progress. The standard BRL and the extended BRL are comparing verified through the task of movement by the connected robots. Through a cooperative task by the connected robots, performance of the extended BRL is compared with standard BRL.