Efficient Exploration by Decision Making Considering Curiosity and Episodic Memory in Deep Reinforcement Learning

Reinforcement learning is difficult in sparse environments where rewards are not easy to obtain, such as large scale real space. In recent years, methods to promote exploration by generating intrinsic rewards based on curiosity in such environments have received much attention. However, in large state spaces, efficiency is a problem. In this paper, we propose a DQN-based algorithm that takes into account both episodic memory and intrinsic rewards during making decision in order to improve the efficiency of exploration. We evaluate this method through some tasks.