Abstract
A behavior acquisition mechanism based on reinforcement learning (RL) for an autonomous multi-robot system (MRS) is extended to improve the robustness in this paper. An important fundamental topic for a MRS is how to design an on-line behavior acquisition mechanism that is capable of developing a set of roles for coordinated behavior as well as assigning them to each robot in a timely manner in a physical environment. From this point of view, a robot controller is constructed with RL that has the mechanism of segmenting continuous state space and continuous action space simultaneously and autonomously to extend the adaptability to non-Markovian environment. This RL developed in our research group is called Bayesian-discrimination-function-based Reinforcement Learning (BRL). Although BRL is proved to be effective to develop a globally stable behavior to achieve a task for a MRS, it has the tendency of being brittle against disturbance after successful learning. This problem is generally called over-fitting. This paper shows an approach to extending BRL for overcoming this problem. This extended BRL is examined by three arm-type autonomous robots, the task of which is lifting an object without tilting it. The experiments conducted illustrate the robust performance of the proposed method by developing various types of cooperative behavior as a result of autonomous specialization.
