2A1-L06 ヘテロジーニアス間転移学習のための知識再利用法の検討

Abstract

This paper describes an influence of heterogeneity for a transfer learning in heterogeneous multiple robots. A multi-agent robot system using reinforcement learning (Multi-agent reinforcement learning: MARL) is effective methodology to obtain efficient behaviors autonomously in dynamic environment. However, the MARL has a problem such as prolongation in learning time. The existing research indicated effectiveness of transfer learning for reducing of learning time in reinforcement learning. Transfer learning is a framework of reuse of knowledge which is obtained by reinforcement learning. In our prior research, we proposed and applied the parameter, transfer rate, to the transfer learning method for heterogeneous agents. The parameter adjusts the degree of reuse of past knowledge toward a newly obtained knowledge in a new task. However, transferability is depending on heterogeneity factors such as difficulty of tasks and agents' functions. In this paper, we investigate the effectiveness of transfer rate with heterogeneous agents and environments by conducting a computer simulation.