Robust Data-Driven Receding Horizon Control

Jian Zheng, Sahand Kiani, Mario Sznaier, Constantino Lagoa

TL;DR

This paper introduces a robust, data-driven framework for controlling unknown systems under noise. By using "set-membership" constraints instead of traditional modeling, the controller identifies all possible systems consistent with observed data and optimizes for the worst-case scenario, while continuously refining its knowledge online to improve performance over time.

What Problem Does This Paper Address?

In modern engineering, creating an accurate mathematical model of a system (like a robot or a chemical plant) is often the hardest part of control design. Data-Driven Control (DDC) aims to bypass this by designing controllers directly from measurements.

However, most existing DDC methods rely on "Willems’ Fundamental Lemma," which can become computationally expensive or mathematically "messy" (non-convex) when the data is noisy. Furthermore, many of these methods are "static"—they use a batch of initial data but don't learn from the new data generated while the controller is actually running.

The authors tackle the challenge of designing a Robust Data-Driven Receding Horizon Control (DDRHC). They want a controller that:

  1. Handles bounded noise and disturbances.
  2. Respects strict state and input constraints (e.g., "don't let the motor overheat" or "don't exceed maximum voltage").
  3. Updates its internal "understanding" of the system in real-time to become less conservative as more data arrives.

Methods

The researchers move away from the standard "Fundamental Lemma" approach and instead use Set-Membership Constraints.

  • The Consistency Set: Instead of picking one "best-fit" model, the algorithm maintains a "consistency set"—a collection of all possible linear systems $(A, B)$ that could have produced the observed data given the known noise bounds.
  • Greedy Receding Horizon: At each time step, the controller solves a min-max optimization problem. It looks for a control input $u$ that minimizes the worst-case "contraction rate" ($\lambda$) across every single system in the consistency set. This ensures that even if the "worst" possible system is the real one, the state will still move toward the target.
  • The "Farkas" Trick: Robust optimization is usually computationally heavy because the number of possible systems can be infinite or the "corners" of the uncertainty set can grow exponentially. The authors use the Extended Farkas’ Lemma to transform this complex robust problem into a much simpler Linear Program (LP). This allows the controller to run efficiently without needing to explicitly list every possible system "vertex."
  • Online Refinement: Every time the system takes a step, the new state and input are added to the dataset. This shrinks the consistency set, reducing uncertainty and allowing the controller to be more aggressive and precise.

Datasets

As a theoretical control paper, the authors evaluate their method using simulated discrete-time Linear Time-Invariant (LTI) systems:

  • Example 1: A 2D system with a specific state-space matrix and process noise bounded by $\epsilon = 0.04$. It uses 10 initial noisy measurements for training.
  • Example 2: The same system but with significantly higher noise ($\epsilon = 0.1$) to test the limits of the robust framework.

Key Insights and Ideas

  • Data as a Shrinking Box: The core "aha" moment is treating system identification not as finding a point, but as shrinking a box. By incorporating execution data, the "box" of possible systems gets smaller every second the machine is running.
  • Duality for Speed: By using the Extended Farkas' Lemma, they prove that you don't need to know the "edges" of your uncertainty to protect against them. You can use mathematical duality to enforce robustness directly within the optimization constraints.
  • Guaranteed Stability: The paper provides a formal proof that this method guarantees Uniformly Ultimately Bounded (UUB) stability. This means that even with persistent noise, the system is guaranteed to stay within a specific "safe" region around the target.

Key Results

  • Faster Convergence: In simulations, the proposed DDRHC reached the target "safety set" significantly faster than standard data-driven controllers that don't use receding horizon updates.
  • Adaptability: As shown in the "Contraction $\lambda$" plots, the value of $\lambda$ (the measure of how fast the system converges) drops over time. This quantitatively proves that the controller is "learning" and becoming more efficient as it collects more data during execution.
  • Constraint Satisfaction: The method successfully kept the system within state and input boundaries (the "box" constraints) even when noise levels were increased by 250% (from 0.04 to 0.1).

Limitations and Future Work

  • Greedy Horizon: The current optimization looks only one step ahead (greedy). While this is computationally fast, traditional MPC often looks many steps into the future. The authors suggest extending the prediction horizon in future versions.
  • Linear Systems Only: The framework is currently designed for Linear Time-Invariant (LTI) systems. Real-world systems are often nonlinear, which would require more complex "consistency sets."
  • Noise Bounds: The method assumes we know the maximum possible noise ($\epsilon$). If the noise exceeds this bound, the "true" system might fall outside the consistency set, potentially breaking the stability guarantees. Future work may look into relaxing these assumptions or handling stochastic (probabilistic) noise.

Summary updated

Abstract

This paper presents a data-driven receding horizon control framework for discrete-time linear systems that guarantees robust performance in the presence of bounded disturbances. Unlike the majority of existing data-driven predictive control methods, which rely on Willem's fundamental lemma, the proposed method enforces set-membership constraints for data-driven control and utilizes execution data to iteratively refine a set of compatible systems online. Numerical results demonstrate that the proposed receding horizon framework achieves better contractivity for the unknown system compared with regular data-driven control approaches.
BibTeX
@inproceedings{zheng2025robust,
  author = {Jian Zheng and Sahand Kiani and Mario Sznaier and Constantino Lagoa},
  title = {{Robust Data-Driven Receding Horizon Control}},
  booktitle = {IFAC-PapersOnLine},
  year = {2025},
  doi = {10.1016/j.ifacol.2025.10.074},
}

Loading related content...

Figure