Overview of Deep Reinforcement Learning Improvements and Applications

Abstract

The deep reinforcement learning value has received a lot of attention from researchers since it was proposed. It combines the data representation capability of deep learning and the self-learning capability of reinforcement learning to give agents the ability to make direct action decisions on raw data. Deep reinforcement learning continuously optimizes the control strategy by using value function approximation and strategy search methods, ultimately resulting in an agent with a higher level of understanding of the target task. This paper provides a systematic description and summary of the corresponding improvements of these two types of classical method machines. First, this paper briefly describes the basic algorithms of classical deep reinforcement learning, including the Monte Carlo algorithm, the Q-Learning algorithm, and the most primitive deep Q network. Then the machine improvement method of deep reinforcement learning method based on value function and strategy gradient is introduced. And then the applications of deep reinforcement learning in robot control, algorithm parameter optimization and other fields are outlined. Finally, the future of deep reinforcement learning is envisioned based on the current limitations of deep reinforcement learning.