Robust Data Driven Control via Set Membership Uncertainty

377 concepts260 papers

Data-driven control offers an alluring promise: stabilizing and optimizing complex dynamical systems directly from operational trajectories, completely bypassing the arduous process of first-principles modeling. However, when deployed in the real world, this approach encounters a fundamental vulnerability: measurement noise.

In purely data-driven frameworks, noise is not merely a minor nuisance to be filtered out; it is a source of corruption that propagates directly into the control loop. As noted by Frank Allgöwer et al. in An overview of systems-theoretic guarantees in data-driven model predictive control, relying on raw, noise-corrupted input-output data introduces severe sensitivity, often leading to overfitting, poor tracking, or catastrophic instability. To build controllers that are truly safe, researchers have shifted away from finding a single "best-fit" model, moving instead toward frameworks that treat noise as a defining element of structural uncertainty.

The Pitfall of Point Estimates

Traditional system identification typically takes noisy data and attempts to find a single point-estimate model that minimizes prediction error. In a control context, this separation between identification and design can be dangerous. If the initial data collection is corrupted by stochastic noise, the resulting point-estimate model will be slightly wrong. Designing a controller tailored exclusively to this incorrect model can lead to unstable control actions when applied to the real physical system.

Even direct, non-parametric methods like Data-Enabled Predictive Control (DeePC)—which use historical trajectories directly to predict future behavior without an explicit model—suffer from this vulnerability. As Julian Berberich et al. demonstrate in Data-Driven Model Predictive Control with Stability and Robustness Guarantees, noise embedded within historical data matrices corrupts the predictions, causing the controller to overfit to past measurement errors rather than capturing the true underlying physics.

From Points to Sets: Set-Membership Identification

To overcome the limitations of point estimates, a powerful alternative paradigm has emerged: set-membership identification. Instead of asking, "What is the single best model for this data?" this approach asks, "What is the set of all possible models that do not contradict the data and our known noise bounds?"

As surveyed by Tim Martin et al. in Guarantees for data-driven control of nonlinear systems using semidefinite programming: A survey, this methodology constructs a consistency set containing every mathematical model capable of generating the observed trajectories under a specified noise bound.

In Robust data-driven state-feedback design, Julian Berberich et al. argue that treating measurement noise as a source of structural uncertainty rather than a signal to be filtered is the key to mathematical safety. Because the true physical system is guaranteed to live somewhere within this consistency set, any controller that can successfully manage the entire set is guaranteed to safely control the real system.

Traditional Approach (Point-Estimate):
Noisy Data ──> Single "Best" Model ──> Tailored Controller ──> Risk of Instability

Set-Membership Approach:
Noisy Data + Noise Bounds ──> Consistency Set (All Valid Models) ──> Robust Controller ──> Guaranteed Stability

Controlling the Entire Set Simultaneously

How do you design a single controller capable of stabilizing an entire set of candidate systems at once? The answer lies in robust optimization and control synthesis tools like Semidefinite Programming (SDP) and Linear Programming (LP).

The goal is to find a control law that satisfies stability criteria (such as Lyapunov or dissipativity conditions) for every single system configuration within the uncertainty description. For example, Steffen W. R. Werner et al. show in On the sample complexity of stabilizing linear dynamical systems from data that designing a controller to be simultaneously stabilizing across the entire identified set ensures robust performance despite model non-uniqueness.

This framework extends naturally to complex, non-linear dynamics. In Data-Driven Stabilization of Nonlinear Systems with Rational Dynamics, Robin Strässer et al. tackle systems governed by ratios of polynomials. By transforming these rational equations into polynomial forms and applying Sum-of-Squares (SOS) optimization, they create a robust framework that guarantees stability for any system consistent with the noisy data. Similarly, for systems where variables must remain strictly non-negative, Jared Miller et al. leverage the Extended Farkas Lemma in Data-Driven Control of Positive Linear Systems using Linear Programming to convert the robust stabilization of magnitude-bounded noise into a highly scalable linear program.

Why Regularization Works: The Robust Equivalence

While set-membership explicitly builds uncertainty sets for parametric models, non-parametric predictive controllers like DeePC often rely on a different mechanism to combat noise: regularization. In practice, control engineers frequently append heuristic $l_1$ or $l_2$ penalty terms to their optimization objectives to smooth out control actions and prevent overfitting.

Remarkably, recent theoretical breakthroughs have revealed that these heuristic regularizations are not just hacks—they are mathematically equivalent to robust optimization.

Linbin Huang et al. proved in Quadratic Regularization of Data-Enabled Predictive Control: Theory and Application to Power Converter Experiments that adding a quadratic regularization term to a data-driven controller is mathematically identical to solving a formal "min-max" robust optimization problem. This means the controller is implicitly optimizing performance against the worst-case noise scenario hidden within the data matrices.

Expanding on this, Jeremy Coulson et al. demonstrated in Regularized and Distributionally Robust Data-Enabled Predictive Control that these regularizations correspond to optimizing against a worst-case noise distribution via Distributionally Robust Optimization (DRO). This provides an intuitive explanation for why regularized data-driven controllers perform so reliably in stochastic environments: the regularization parameter directly dictates the size of the noise distribution the controller is guarded against.

The Power of Data Volume

If noise creates a large consistency set of possible systems, how do we make our controller less conservative? The solution is simple: collect more data.

As more trajectory data is observed, the number of constraints on the consistency set increases, causing the set of valid models to shrink around the true system dynamics. Robin Strässer et al. identified a clear phase transition in Data-Driven Stabilization of Nonlinear Systems with Rational Dynamics: there is a distinct data volume threshold where the probability of finding a stabilizing controller transitions sharply from low to certain.

By scaling up the volume of historical data, the framework naturally compensates for measurement noise, narrowing down the structural uncertainty and allowing the synthesized control law to achieve optimal performance without sacrificing its mathematical safety guarantees.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 266 papers
  2. Version 1 · 2026-10-04 · 225 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_noise_sensitivity_in_data_driven_control_678efdda,
  title = {noise sensitivity in data-driven control},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/noise-sensitivity-in-data-driven-control/rdr_678efdda05f64314bae6d44f2ab149d4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 2, frozen text):

@misc{lacuna_noise_sensitivity_in_data_driven_control_678efdda_v2,
  title = {noise sensitivity in data-driven control},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/noise-sensitivity-in-data-driven-control/rdr_678efdda05f64314bae6d44f2ab149d4/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-22},
}

See something wrong? Report a mistake