Robust Learning for Structural Vibration Mitigation under Excitation Uncertainty

Abstract

Learning-based structural vibration mitigation must remain reliable under excitation uncertainty whose severity, spectral content, and temporal structure can deviate from nominal training conditions. Standard single-agent reinforcement learning (RL) typically optimizes expected return under a fixed disturbance model, which can yield policies that perform well in distribution yet exhibit brittle behavior and elevated tail risk under disturbance shift or strategically adverse realizations. This paper studies reliability-aware policy learning through a game-theoretic uncertainty abstraction: we formulate vibration mitigation as a two-player zero-sum Markov game between a controller and an admissible excitation agent, and train policies via tabular Minimax-Q learning with a state-wise matrix-game solve. Using an earthquake-motivated single-degree-of-freedom base-excitation structure, we compare uncontrolled and PD baselines, single-agent Q-learning trained under stochastic excitation models, and Minimax-Q learning trained against bounded admissible excitation actions. We evaluate policies under (i) a fixed learned minimax excitation policy and (ii) controller-specific worst-case excitation policies within the same admissible disturbance set, reporting both average response metrics and reliability-relevant measures. The results show that minimax training produces controllers that are harder to exploit and yield improved worst-case reliability profiles within the admissible disturbance set.