Minimax Control of Ambiguous Linear Stochastic Systems Using the Wasserstein Metric

Kihyun Kim, Insoon Yang

Introduction

In classical stochastic control, such as the Linear-Quadratic-Gaussian (LQG) regulator, designers typically assume perfect knowledge of the noise distribution affecting a system. However, in practice, we often rely on finite empirical data, which may not accurately represent real-world disturbances. This "distributional ambiguity"—uncertainty about the probability distribution itself—can lead to performance degradation or instability. This paper introduces a robust minimax control framework that utilizes the Wasserstein metric to bridge the gap between standard stochastic control and robust $H_\infty$ methods.

The Minimax Framework and Wasserstein Penalty

The authors formulate the control problem as a zero-sum minimax game between a controller and a hypothetical adversary. The controller seeks to minimize a quadratic cost function, while the adversary chooses the "worst-case" noise distribution to maximize it. To ensure the adversary remains realistic, a Wasserstein penalty is added to the cost function. This penalty, weighted by a parameter $\lambda$, measures the distance between the adversary's distribution and the observed empirical data.

Key Mathematical Contributions

  • Finite-Horizon Control: The derivation of a modified Riccati equation provides a closed-form, linear optimal control policy.
  • Infinite-Horizon Stability: The paper identifies conditions under which the system converges to a unique solution for the Algebraic Riccati Equation (ARE), ensuring long-term stability.
  • Computational Efficiency: By applying Kantorovich duality, the authors reduce an infinite-dimensional optimization problem over probability measures into a tractable finite-dimensional optimization solvable via standard Riccati recursions.

Bridging Control Paradigms

The parameter $\lambda$ serves as a tunable dial for robustness. As $\lambda \to \infty$, the penalty for the adversary becomes insurmountable, causing the framework to revert to standard LQG control. Conversely, as $\lambda$ decreases, the controller becomes increasingly robust against distributional shifts. This approach effectively generalizes $H_\infty$ control, which focuses on worst-case noise realizations, to handle worst-case noise distributions.

Experimental Results: Power Grid Frequency Regulation

The framework was evaluated using the IEEE 39 Bus System, a standard model for the New England power grid. Using 10 samples of disturbances as empirical data, the authors demonstrated the controller's effectiveness in maintaining frequency stability.

Frequency regulation performance across different control parameters, demonstrating reduced fluctuations under the minimax controller compared to standard methods.

The simulations show that the minimax controller significantly mitigates frequency fluctuations when the actual noise distribution deviates from the design samples. However, this robustness comes at a cost.

The trade-off between control energy and the robustness parameter λ, illustrating that increased robustness (lower λ) requires higher average control energy.

As shown in the figure above, there is a clear "cost of robustness." Lowering $\lambda$ to increase protection against ambiguity results in higher average control energy. This quantification allows engineers to make informed decisions regarding the trade-off between performance and safety margins.

Conclusion and Future Directions

This research provides a mathematically rigorous yet computationally practical method for controlling systems under distributional uncertainty. While the current work assumes full state observation and focuses on discrete-time systems, it lays the groundwork for future extensions into partial measurement scenarios and continuous-time stochastic differential equations. The relationship between sample size and the optimal selection of $\lambda$ remains a promising area for further theoretical exploration.

Summary updated

BibTeX
@inproceedings{kim2020minimax,
  author = {Kihyun Kim and Insoon Yang},
  title = {{Minimax Control of Ambiguous Linear Stochastic Systems Using the Wasserstein Metric}},
  booktitle = {CDC},
  year = {2020},
  url = {https://openreview.net/forum?id=J-z1VW81Apg},
}
Figure