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.

References

2023

  1. Online Stabilization of Unknown Linear Time-Varying Systems
    J. Yu, V. Gupta, and A. Wierman
    In IEEE Conference on Decision and Control (CDC). 2023
  2. Online Adversarial Stabilization of Unknown Networked Systems
    J. Yu, D. Ho, and A. Wierman
    In ACM SIGMETRICS. 2023

2022

  1. Robust Online Voltage Control with an Unknown Grid Topology
    C. Yeh, J. Yu, Y. Shi, and A. Wierman
    In ACM International Conference on Future and Sustainable Energy Systems (ACM e-Energy). 2022

2021

  1. Robust Reinforcement Learning: A Constrained Game-theoretic Approach
    J. Yu, C. Gehring, F. Schäfer, and 1 more author
    In Annual Learning for Dynamics and Control Conference (L4DC). 2021