APPLICATION OF EVOLUTIONARY COMPUTATION FOR EFFICIENT REINFORCEMENT LEARNING

Abstract

In this paper, we propose a new method based on evolutionary computation for setting the metaparameters of reinforcement learning in order to match the demands of the task and reduce the learning time. In our method, we encode the metaparameters as the agent's genes and take the metaparameters of best-performing agents in the next generation. We investigate the influence of metaparameters on the agent learned policy and learning time. The results show that appropriate settings of metaparameters found by evolution have a great effect on the learning time and are strongly dependent on each other. In addition, by using the Cyber Rodent robot, we verified that metaparameters evolved in simulation are helpful for learning in real hardware.