Survey on Memory-augmented Deep Reinforcement Learning
WANG Chen-1,ZENG Fan-yu1,GUO Jiu-xia1,2
1(School of Computer Science and Engineering,University of Electronic Science and Technology of China,Chengdu 611731,China)2(College of Air Traffic Management,Civil Aviation Flight University of China,Guanghan 618307,China)
Abstract：In recent years,deep reinforcement learning has developed rapidly.To improve the performance of deep reinforcement learning(DRL)in high-dimensional state space and dynamic complex environment,researchers introduce memory-augmented neural networks(MANN)into DRL,and propose various memory-augmented deep reinforcement learning(MADRL)algorithms,which becomes a research hotspot.In this paper,according to the types of MANN,MADRL algorithms can be categorized into four classes:MADRL based on experience replay,MADRL based on memory network,MADRL based on episodic memory and MADRL based on differentiable neural computer.In addition,the training environments for DRL are introduced.Meanwhile,this paper systematically summarizes and analyzes the advantages and disadvantages of the research works on MADRL.Finally,the prospect and future research directions of MADRL are discussed.
汪晨,曾凡玉,郭九霞,. 记忆增强型深度强化学习研究综述[J]. 小型微型计算机系统, 2021, 42(3): 454-461.
WANG Chen-,ZENG Fan-yu,GUO Jiu-xia,. Survey on Memory-augmented Deep Reinforcement Learning. Journal of Chinese Computer Systems, 2021, 42(3): 454-461.