Reinforcement Learning with Guaranteed Robustness Under Dynamics Modeling Uncertainties

This paper presents robustness-guaranteed reinforcement learning (RGRL) under bounded dynamics modeling uncertainties in a wide range of dynamical systems, including fully actuated systems. Robust reinforcement learning enhances system robustness through bilevel optimization, while the RL controller still struggles to offer strict robustness guarantees. By describing the uncertainties and system dynamics models via ReLU neural networks, the states that most violate the proposed robustness conditions can be exactly found, and their robustness is prioritized during training. The solvability of the RL problem is particularly studied, yeilding a necessary condition that captures the adversarial relationship between control saturation and uncertainties. Subsequently, the attractive region can be pre-validated. Numerical simulations were performed to illustrate the effectiveness of the proposed method.