Predictive Control Using Raw Data Trajectories

This direction is merged into another direction: data-driven system representation. Its last version stays readable and citable here.

Version 1 of 124 papers25 conceptsfrozen 2026-10-04

In traditional control engineering, the path from a physical system to a working controller usually passes through a narrow bottleneck: system identification. To control a robot or a power grid, engineers first collect data to estimate a mathematical model—usually a set of differential equations or state-space matrices ($A, B, C, D$). Only then can they apply techniques like Model Predictive Control (MPC).

However, this intermediate step is often the weakest link. Modeling errors can lead to poor performance, and the process of fitting parameters to data can be computationally expensive and physically opaque. A growing body of research suggests we can bypass this bottleneck entirely. By leveraging Behavioral Systems Theory, researchers have developed methods to control complex systems using raw historical trajectories as the "model" itself.

The Fundamental Lemma of Data

The core intuition behind this shift is a result known as Willems’ Fundamental Lemma. It states that for a linear time-invariant system, any valid trajectory (a sequence of inputs and outputs) can be expressed as a linear combination of a sufficiently "rich" set of previously recorded trajectories.

In this context, "rich" means the data must be persistently exciting. If you have recorded enough diverse behavior from the system, those recordings span the entire subspace of what the system is physically capable of doing. As Jicheng Shi and Colin Jones demonstrate in Data-Driven Input Reconstruction and Experimental Validation, this allows us to reconstruct unknown inputs or predict future outputs by simply finding the right weights to combine our historical data points.

From Identification to DeePC

The most prominent application of this idea is Data-Enabled Predictive Control (DeePC). Introduced by Jeremy Coulson, John Lygeros, and Florian Dorfler in Data-Enabled Predictive Control: In the Shallows of the DeePC, DeePC replaces the explicit model in an MPC optimization problem with a data-driven constraint.

Instead of asking the optimizer to find an input $u$ that satisfies $x_{t+1} = Ax_t + Bu_t$, DeePC asks the optimizer to find a linear combination of past data trajectories that matches the current initial conditions and minimizes the cost function. This non-parametric approach allows for optimal control without ever defining what $A$ or $B$ are. Ivan Markovsky and Linbin Huang note in their tutorial that this is particularly powerful for power systems, where the underlying physics are complex and dominated by fast-switching power electronics.

Robustness and State Estimation

While the theory is elegant for perfect data, real-world sensors are noisy. Recent work has focused on making these data-driven methods robust. Yifei Li and colleagues in Data-driven Polytopic Output Synchronization of Heterogeneous Multi-agent Systems from Noisy Data use "polytopes" to bound uncertainties, ensuring that even with noisy data, a fleet of agents can stay synchronized.

The behavioral approach also extends to "seeing" the hidden states of a system. Matthias A. Müller has shown in Data-based Moving Horizon Estimation under Irregularly Measured Data that we can perform state estimation even when measurements are sparse or arrive at irregular intervals. By treating the history of the system as a library of possibilities, we can infer the current state by seeing which historical trajectories best match our recent, albeit imperfect, observations.

Hybridizing Physics and Data

We don't always have to choose between models and data. Some of the most effective applications use a hybrid approach. In Estimation of Road Friction Coefficient via the Data Enforced Unscented Kalman Filter, Jinheng Han and Chen Lv combine traditional vehicle dynamics models with a data-driven correction layer. This "Data Enforced" filter uses historical trajectories to correct the model's predictions in real-time, allowing a vehicle to sense a slippery road surface much faster than a standard filter could.

Verifying the Black Box

Finally, this framework provides a path toward safety guarantees for modern AI. As neural networks are increasingly used as controllers, engineers worry about their lack of formal guarantees. Han Wang and Antonis Papachristodoulou demonstrate in Model-Free Verification for Neural Network Controlled Systems that we can use data-driven representations to formally verify the stability of a neural network controller. By proving that the controller plays nicely with the subspace of behaviors observed in the data, we can gain confidence in "black-box" systems without needing a perfect physical model.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_predictive_control_using_raw_data_trajec_da97c1e9,
  title = {Predictive Control Using Raw Data Trajectories},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/predictive-control-using-raw-data-trajectories/rdr_da97c1e91fe6412f942fbb73bba1053b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_predictive_control_using_raw_data_trajec_da97c1e9_v1,
  title = {Predictive Control Using Raw Data Trajectories},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/predictive-control-using-raw-data-trajectories/rdr_da97c1e91fe6412f942fbb73bba1053b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}