Sensitivity of Online Feedback Optimization to time-varying parameters

Marta A. Zagorowska, Lars Imsland

Introduction

Online Feedback Optimization (OFO) has emerged as a powerful paradigm for the real-time control of complex physical systems. By treating optimization algorithms—such as gradient descent—as dynamic controllers, OFO allows systems to iteratively seek optimal operating points based on live measurements. While OFO is celebrated for its inherent robustness to model mismatches, the engineering community has lacked precise analytical tools to quantify its sensitivity. This paper provides the first mathematical framework to derive closed-form expressions for the sensitivity of OFO controllers to their internal tuning parameters and environmental model inaccuracies.

Problem Statement

In practical applications, engineers must choose specific parameters for OFO controllers, such as the step size ($\alpha$) and the scaling metric ($G$). Additionally, the controller often relies on an estimated gradient of the system, which may deviate from reality. Until now, determining how these choices or errors impacted the final objective function required exhaustive "what-if" simulations or trial-and-error. The goal of this research is to replace these brute-force methods with exact formulas that describe how control inputs ($u$) respond to changes in parameters ($p$) over time.

Methodology

The researchers focused on OFO implemented via Projected Gradient Descent. Their approach treats the iterative optimization process as a discrete-time dynamical system, utilizing the following steps:

  1. Quadratic Programming (QP) Reformulation: The projected gradient step is rewritten as a sequence of convex quadratic optimization problems.
  2. KKT Differentiation: By differentiating the Karush-Kuhn-Tucker (KKT) optimality conditions of these QPs, the authors derived exact sensitivities for the control action relative to the parameters at any given time step.
  3. Backpropagation in Time: Recognizing that OFO is a sequential process, the authors applied a chain-rule approach—analogous to backpropagation in neural networks—to track how a parameter change at time $t=1$ propagates through the system to affect the state at time $t=T$.
  4. Closed-Form Expressions: The result is a set of explicit matrix equations that allow for the direct calculation of sensitivity without the need for repeated numerical simulations.

Key Theoretical Insights

The "Forgetting" Property

A significant finding of the study is that OFO controllers exhibit a "forgetting" property. The sensitivity analysis reveals that errors or parameter variations introduced early in the control horizon have a diminishing impact on the final result as the system progresses. This provides a theoretical justification for the observed robustness of OFO in the face of poor initial data or transient disturbances.

Time-Dependent Sensitivity

Sensitivity is not a static value but is highly dependent on the time horizon. If a controller has not yet converged to the optimum, it remains highly sensitive to its initial settings. However, once the system settles near the optimal point, the sensitivity to individual time-step errors drops significantly, provided the step size is sufficiently large to maintain convergence.

Handling Constraints

Engineering systems are often governed by physical limits. The authors' formulas successfully capture the discontinuities in sensitivity that occur when a system hits a constraint boundary. This is critical for safety-constrained applications where abrupt changes in controller behavior must be anticipated.

Experimental Validation

Synthetic 1D Case Study

The mathematical framework was first validated against a synthetic 1D problem. The closed-form sensitivity formulas produced results that perfectly matched numerical finite difference calculations, confirming the mathematical integrity of the KKT differentiation approach.

Gas Lift Optimization

The framework was then applied to a realistic industrial benchmark: a Gas Lift Optimization problem involving five oil/gas wells and two platforms. The study analyzed "gradient mismatch"—a scenario where the controller’s internal model of the system is inaccurate. The sensitivity formulas predicted a potential 14% performance drop; in practice, the actual decrease was 6.1%. This demonstrates that the derived formulas provide a safe, conservative upper bound for performance loss, making them a reliable tool for risk assessment in industrial settings.

Conclusion and Future Work

This research provides a rigorous foundation for the systematic tuning of OFO controllers. By moving beyond empirical observations to closed-form sensitivity analysis, engineers can now predict how model errors and parameter choices will affect system performance.

Future work aims to extend these formulas to handle "degenerate" cases where constraints may be poorly behaved. Ultimately, these sensitivity expressions could enable the development of "self-tuning" controllers that autonomously adjust their parameters in real-time to optimize performance and maintain stability in changing environments.

Summary updated

BibTeX
@inproceedings{zagorowska2025sensitivity,
  author = {Marta A. Zagorowska and Lars Imsland},
  title = {{Sensitivity of Online Feedback Optimization to time-varying parameters}},
  booktitle = {ECC},
  year = {2025},
  doi = {10.23919/ecc65951.2025.11186978},
}
Figure