Moderate Actor-Critic Methods: Controlling Overestimation Bias via Expectile Loss

Overestimation is a fundamental characteristic of model-free reinforcement learning (MF-RL), arising from the principles of temporal difference learning and the approximation of the Q-function. To address this challenge, we propose a novel moderate target in the Q-function update, formulated as a convex optimization of an overestimated Q-function and its lower bound. Our primary contribution lies in the efficient estimation of this lower bound through the lower expectile of the Q-value distribution conditioned on a state. Notably, our moderate target integrates seamlessly into state-of-the-art (SOTA) MF-RL algorithms, including Deep Deterministic Policy Gradient (DDPG) and Soft Actor Critic (SAC). Experimental results validate the effectiveness of our moderate target in mitigating overestimation bias in DDPG, SAC, and distributional RL algorithms.

On Practical Robust Reinforcement Learning: Practical Uncertainty Set and Double-Agent Algorithm

We study a robust reinforcement learning (RL) with model uncertainty. Given nominal Markov decision process (N-MDP) that generate samples for training, an uncertainty set is defined, which contains some perturbed MDPs from N-MDP for the purpose of reflecting potential mismatched between training (i.e., N-MDP) and testing environments. The objective of robust RL is to learn a robust policy that optimizes the worst-case performance over an uncertainty set. In this paper, we propose a new uncertainty set containing more realistic MDPs than the existing ones. For this uncertainty set, we present a robust RL algorithm (named ARQ-Learning) for tabular case and characterize its finite-time error bound. Also, it is proved that ARQ-Learning converges as fast as Q-Learning and the state-of-the-art robust Q-Learning while ensuring better robustness to real-world applications. Next, we propose {\em pessimistic} agent that efficiently tackles the key bottleneck for the extension of ARQ-Learning into the case with larger or continuous state spaces. Incorporating the idea of pessimistic agents into the famous RL algorithms such as Q-Learning, deep-Q network (DQN), and deep deterministic policy gradient (DDPG), we present PRQ-Learning, PR-DQN, and PR-DDPG, respectively. Noticeably, the proposed idea can be immediately applied to other model-free RL algorithms (e.g., soft actor critic). Via experiments, we demonstrate the superiority of our algorithms on various RL applications with model uncertainty.