Q-value-based experience replay in reinforcement learning

Abstract

Experience replay has long been used in reinforcement learning to store and reuse past experiences. However, most existing experience replay methods sample experiences with non-uniform probabilities that are proportional to their values, such as temporal-difference errors, which often lead to biased learning. To address this issue, we propose a new experience replay method that hierarchically samples experiences based on Q-values. Specifically, the proposed method divides a set of uniformly sampled experiences into three groups of the same size according to the Q-values and then uniformly samples the same number of experiences from the three groups as a mini-batch. In this manner, the sampled experiences can effectively maintain diversity. Moreover, to estimate the Q-value accurately, we develop a new critic network based on the self-attention mechanism. We integrate the new experience replay method and critic network into the twin delayed deep deterministic policy gradient algorithm to form a new reinforcement learning algorithm. An extensive set of experiments using several standard benchmarks demonstrates the effectiveness of the proposed algorithm.