Prioritized Experience Replay Method Based on Experience Reward
Jiashan Gao, Xiaohui Li, Weihui Liu, Jingchao Zhao
Abstract
In recent years, artificial intelligence has been widely used in modern construction, and reinforcement learning methods have played an important role in it. The experience replay method is an important means to enable the reinforcement learning method to be widely used in real tasks. In order to improve the efficiency of the experience replay method, this article improves the traditional experience replay method and combines the past experience reward value with the timing difference error (TD error) to form a R- T experience prioritized parameter. Based on the data priority, we use R- T experience prioritized parameter as a standard to update and storage the memory pool data method. The improvement is used to form an improved experience replay method based on experience data priority. This paper applies the improved experience replay method to DDPG (Deep Deterministic Policy Gradient) algorithm and conducts simulation experiments. As a result, the improved experience replay method verified by experiments has a good effect on improving the efficiency of algorithm training and the stability of the algorithm training process.