Hierarchical Sampling Strategies and Their Application to Multi-Intelligent Body Deep Reinforcement Learning

Multi-intelligent reinforcement learning algorithms have received more attention as single intelligences are becoming increasingly difficult to meet increasingly complex business requirements. However, as the number of intelligences increases, algorithm convergence becomes more difficult. In this paper, a hierarchical sampling strategy (HSP) is proposed to solve this problem. The hierarchical sampling strategy divides the empirical samples into different clusters by clustering algorithm and dynamically adjusts the playback probability of the samples according to the importance and similarity of the samples, so as to improve the utilization efficiency of the samples and ensure that different types of samples are balanced during the training process, avoiding the neglect of certain key samples and thus improving the training effect. In order to verify the effectiveness of the strategy, experimental verification is carried out by combining it with the MATD3 algorithm, and the superior performance of HSP is verified through comparative experiments.