Recency-Weighted Acceleration for Continuous Control Through Deep Reinforcement Learning

摘要

Model-free reinforcement learning algorithms have been successfully applied to continuous control tasks. However, these algorithms suffer from severe instability and high sample complexity. Inspired by Averaged-DQN, this paper proposes a recency-weighted target estimator for actor-critic settings, which will construct a target estimator with more weight placed on recently learned value functions, obtaining a more stable and accurate value estimator. Besides, delaying policy updates with more flexible control is adopted to reduce per-update error because of value function errors. Furthermore, to improve the performance of prioritized experience replay (PER) for continuous control tasks, Phased-PER is proposed to accelerate training in different periods. Experimental results are given to demonstrate that using the same hyper-parameters and architecture the proposed algorithm is more robust and achieves better performance, surpassing the existing methods on a range of continuous control benchmark tasks.