Variance Reduction Based Experience Replay for Policy Optimization

For reinforcement learning on complex stochastic systems, it is desirable to effectively leverage the information from historical samples collected in previous iterations to accelerate policy optimization. Classical experience replay, while effective, treats all observations uniformly, neglecting their relative importance. To address this limitation, we introduce a novel Variance Reduction Experience Replay (VRER) framework, enabling the selective reuse of relevant samples to improve policy gradient estimation. VRER, as an adaptable method that can seamlessly integrate with different policy optimization algorithms, forms the foundation of our sample-efficient off-policy algorithm known as Policy Optimization with VRER (PG-VRER). Furthermore, the lack of a rigorous theoretical understanding of the experience replay method in the literature motivates us to introduce a novel theoretical framework that accounts for sample dependencies induced by Markovian noise and behavior policy interdependencies. This framework is then employed to analyze the finite-time convergence of our VRER-based policy optimization algorithm, revealing a crucial bias-variance trade-off in policy gradient estimates: the reuse of old experience introduces increased bias while simultaneously reducing gradient variance. Extensive experiments have shown that VRER offers a notable acceleration in learning optimal policies and enhances the performance of state-of-the-art (SOTA) policy optimization approaches.