Effects of Group Learning in Deep Reinforcement Learning

Abstract

Abstract:Inspired by human group learning, we propose a deep group reinforcement learning method using multiple agents with different performance. By combining a shared experience replay memory and an adaptive agent selection mechanism, the proposed method dynamically selects appropriate agents according to the situation, and improves learning efficiency and performance. In an experiment in the CartPole environment of OpenAI Gym, cooperative learning by the proposed method with agents with different number of neurons outperformed single learning in two aspects: learning speed and final performance. In particular, for the agents with few number of neurons, it was confirmed that learning performance was greatly improved in spite of the setting in which they could hardly accomplish the task alone.