抄録
A multi-robot systems (MRS) is composed of many robots. So far MRS has been studied in various fields. For a robust MRS, we focus on Reinforcement Learning to control a MRS. Reinforcement Learning works well in a static environment. Since, it is difficult for Reinforcement Learning to adopt a dynamic environment. In addition to that, in general, a designer needs to appropriately decide how the discretization level of the state and action spaces is represented. However, there is no general design guideline. To overcome these problems, we have been developing BRL as a type of Reinforcement Learning techniques. In this paper, a technique is proposed for effectively using messy knowledge acquired using BRL for improving the robustness of MRS. The technique reconstructs the state space by using Support Vector Machine based on the input-output data acquired by BRL. To investigate our proposed technique, We conduct computer simulations of a cooperative carrying task with three autonomous mobile robots.
