Co-training by Experience Replay for Reinforcement Learning

Abstract

Listen

Translate article

In this paper, to improve the efficiency of the reinforcement learning model to explore the environment and get better results, a new method which involves the co-training process in reinforcement learning by sharing the experience pool of each agent in the training process has been developed. In this method, agents can gain a better understanding of the environment since agents use different policies to make action and explore the environment. At the same time, this paper designed an agent called Hard Memory Collector by modifying the value function and combining this agent and a normal agent for co-training. As an experimental result on the ViZDoom platform, the model achieved better results than the original Duel DQN network in terms of score, steps used per game and loss value.