Predictive rollout–guided constrained reinforcement learning for low-conservatism robust vibration control under parametric uncertainty

Abstract

Active vibration control of mechanical structures is unavoidably affected by bounded parametric uncertainty caused by varying payloads, aging, and modeling errors. Such uncertainty can degrade nominal vibration suppression and, more importantly, trigger violations of hard safety limits such as travel, deflection, and strain bounds. This study proposes a Predictive Rollout–Guided Constrained Reinforcement Learning framework (PECRL) for empirical robustness enhancement with reduced nominal conservatism under bounded parameter variations. The controller is trained in a nominal environment, while uncertainty-aware behavior is promoted through differentiable finite-horizon closed-loop rollouts evaluated at the vertices of an uncertainty set. The rollouts provide two complementary training signals: a smooth vertex-based surrogate of predicted hard-limit exceedance and an envelope-width penalty for reducing parameter-induced trajectory dispersion. These terms are incorporated into policy learning through multi-constraint primal–dual Lagrangian updates. PECRL is intended as an empirical robustness-enhancing framework rather than a formal worst-case safety certificate. Numerical validation is conducted on a Van der Pol oscillator, a quarter-car active suspension with hard travel limits, and an uncertain cantilever-beam benchmark with actuator saturation.