Behavioral Control via Data Trajectories

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

Version 1 of 137 papers37 conceptsfrozen 2026-10-04

Traditional control theory relies on a "model-first" paradigm: first, you write down differential equations that describe a system's physics, and then you design a controller based on those equations. But for complex systems like soft robotic arms, urban traffic grids, or chemical plants, the "true" physics are often too messy to capture in a simple $x_{t+1} = Ax_t + Bu_t$ format.

A growing body of research suggests we can skip the equations entirely. By treating a dynamical system not as a set of formulas, but as a collection of possible trajectories, we can control complex hardware using nothing but raw historical data.

The Behavioral Perspective

In the behavioral approach, a "system" is simply the set of all possible signals (inputs and outputs) it can produce. If you have seen enough of a system's past behavior, you should be able to predict its future behavior without ever knowing its internal state variables.

The mathematical engine behind this is the Hankel matrix. As Liliana Lo Presti and Marco La Cascia demonstrate in Using Hankel matrices for dynamics-based facial emotion recognition and pain detection, a Hankel matrix stacks time-shifted versions of a single data sequence. This structure embeds the "observability" of the system—it captures how the system's variables evolve and correlate over time.

Willems’ Lemma: Data as a Basis

The reason this works for control is a result known as Willems’ Fundamental Lemma. It states that if a system is linear and the historical data is "persistently exciting" (meaning the input signal was rich enough to kick the system into all its possible modes), then any valid future trajectory of that system can be expressed as a linear combination of the columns in that Hankel matrix.

Essentially, the historical data acts as a basis for the space of all possible behaviors. To predict the future, you don't solve a differential equation; you find the right weights to combine your past observations. Marko Nonhoff applies this in Online convex optimization for data-driven control of dynamical systems to navigate unknown systems with time-varying costs, showing that a single prerecorded trajectory is often all you need to achieve optimal performance.

Data-Enabled Predictive Control (DeePC)

The most prominent application of this idea is Data-Enabled Predictive Control (DeePC). In a standard Model Predictive Control (MPC) setup, you use a model to forecast the future and pick the best action. In DeePC, you replace that model with the Hankel matrix of raw data.

Jeremy Coulson, John Lygeros, and Florian Dorfler formalized this in Regularized and Distributionally Robust Data-Enabled Predictive Control. They showed that while using raw data might seem risky due to noise, adding regularization terms to the optimization problem is mathematically equivalent to Distributionally Robust Optimization (DRO). This means the controller is implicitly protecting itself against the worst-case noise distribution in the data.

Why Direct Data Control Wins

One might ask: why not just use the data to identify a model first, and then use standard MPC? This is the "indirect" approach.

However, Jeremy Coulson and Ivan Markovsky proved in Bridging direct & indirect data-driven control formulations via regularizations and relaxations that direct methods like DeePC are actually convex relaxations of the indirect approach. By skipping the explicit model-fitting step, these controllers can be more robust to modeling errors and computationally more efficient.

This efficiency is visible in practical applications:

Bridging to Safety and Learning

The behavioral framework also provides a safety net for modern AI. Deep Reinforcement Learning (RL) is powerful but lacks stability guarantees. Shuyuan Wang and Bhushan Gopaluni have developed frameworks in A modular framework for stabilizing deep reinforcement learning control that use Hankel matrices as an internal representation. This allows an RL agent to explore and learn while ensuring the resulting behavior stays within the "subspace" of stable trajectories defined by the data.

By shifting the focus from equations of motion to observed behaviors, this research direction provides a rigorous way to control the world's most complex systems using the very thing they produce most: data.

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_behavioral_control_via_data_trajectories_2166b3f3,
  title = {Behavioral Control via Data Trajectories},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/behavioral-control-via-data-trajectories/rdr_2166b3f3547e487493979f99261dec91},
  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_behavioral_control_via_data_trajectories_2166b3f3_v1,
  title = {Behavioral Control via Data Trajectories},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/behavioral-control-via-data-trajectories/rdr_2166b3f3547e487493979f99261dec91/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}