Lyapunov Robust Constrained-MDPs: Soft-Constrained Robustly Stable Policy Optimization under Model Uncertainty
This work addresses safety and robustness issues in reinforcement learning for applications requiring reliable performance under uncertainty, but it appears incremental as it combines existing frameworks.
The paper tackled the problem of ensuring both safety and robustness in reinforcement learning by proposing robust constrained MDPs (RCMDPs), which unify constrained MDPs for safety and robust MDPs for model uncertainty, and developed policy gradient algorithms with Lyapunov-based reward shaping to improve stability and convergence.
Safety and robustness are two desired properties for any reinforcement learning algorithm. CMDPs can handle additional safety constraints and RMDPs can perform well under model uncertainties. In this paper, we propose to unite these two frameworks resulting in robust constrained MDPs (RCMDPs). The motivation is to develop a framework that can satisfy safety constraints while also simultaneously offer robustness to model uncertainties. We develop the RCMDP objective, derive gradient update formula to optimize this objective and then propose policy gradient based algorithms. We also independently propose Lyapunov based reward shaping for RCMDPs, yielding better stability and convergence properties.