Opportunistic Actor-Critic (OPAC) with Clipped Triple Q-learning

Srinjoy Roy, Saptam Bakshi, Tamal Maharaj, Swagatam Das

Keywords: Model-free Deep RL, Actor-Critic, Estimation Bias, Continuous Control

TL;DR: OPAC achieves higher average rewards than relevant baselines and mitigates the underestimation bias with the help of Clipped Triple Q-learning.

Abstract: Despite being the most successful model-free deep reinforcement learning (RL) algorithms in recent years, Soft Actor-Critic (SAC) and Twin Delayed Deep Deterministic Policy Gradient (TD3) have their respective downsides--TD3 performs well in simple tasks, while SAC does so in relatively complicated ones. However, they also suffer from underestimation due to Clipped Double Q-learning, i.e., taking a minimum of two Q-values. This paper introduces Opportunistic Actor-Critic (OPAC), an ensemble model-free deep RL algorithm that performs well in simple and complex tasks. OPAC combines the features of TD3 and SAC under one roof to retain their respective benefits. It also employs three critics and considers taking the mean of the smallest two Q-values for updating the shared target, dubbed Clipped Triple Q-learning. Our analytical results establish that Clipped Triple Q-learning incurs less underestimation than Clipped Double Q-learning. Furthermore, we have systematically evaluated OPAC in MuJoCo environments, and the empirical results indicate that OPAC attains higher average rewards than the current baselines.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)