An Empirical Evaluation of Posterior Sampling for Constrained Reinforcement Learning

TL;DR

This paper evaluates posterior sampling methods for constrained reinforcement learning, proposing two simple, efficient algorithms that outperform existing optimistic approaches in empirical tests.

Contribution

Introduces two novel posterior sampling algorithms for constrained RL that are simpler, more statistically efficient, and computationally cheaper than prior methods.

Findings

Posterior sampling achieves state-of-the-art performance.

The proposed algorithms outperform optimistic algorithms.

Simple algorithms are effective in constrained RL settings.

Abstract

We study a posterior sampling approach to efficient exploration in constrained reinforcement learning. Alternatively to existing algorithms, we propose two simple algorithms that are more efficient statistically, simpler to implement and computationally cheaper. The first algorithm is based on a linear formulation of CMDP, and the second algorithm leverages the saddle-point formulation of CMDP. Our empirical results demonstrate that, despite its simplicity, posterior sampling achieves state-of-the-art performance and, in some cases, significantly outperforms optimistic algorithms.