Adversarial Stabilization
Online, adaptive stabilization of unknown dynamics under worst-case, non-stochastic perturbations.
Safety-critical systems often need to stabilize unknown dynamics on the fly, without the luxury of an a priori stabilizing controller or stochastic assumptions on disturbances. This project studies online, adaptive simultaneous exploration and control under non-stochastic, potentially adversarial perturbations, addressing worst-case uncertainty while learning to stabilize the system from scratch in dynamic environments.
The core idea is a modular set-based online control framework: at each step, the controller adaptively either reduces uncertainty about the unknown system or applies inputs that guarantee desirable behavior regardless of adversarial actions. This creates a principled exploration-exploitation tradeoff for closed-loop learning, instead of relying on naive excitation that can be unsafe or too conservative.
The framework extends to unknown time-varying dynamics, safety-constrained systems, and large-scale networked systems under localized and delayed communication. We also develop adversarially robust policy gradient methods that exploit the geometry of the feasible policy space, connecting these guarantees to reinforcement-learning practice.
Selected outcomes. Adversarial stability for unknown time-varying systems (Yu et al., 2023); safe online learning with finite mistake bounds for power systems (Yeh et al., 2022); networked extensions with localized and delayed communication (Yu et al., 2023); and constrained game-theoretic robust RL (Yu et al., 2021).
Timeline. 2022 – present.