Data Enabled Predictive Control Without System Identification

42 concepts42 papers1 proposal

For decades, the standard recipe for controlling a complex physical system has followed a strict two-step paradigm. First, engineers collect experimental data to build an explicit mathematical model of the system—a process known as system identification. Second, they use this model (typically a state-space representation like $x_{t+1} = Ax_t + Bu_t$) to design a controller, such as a Model Predictive Controller (MPC) or a Linear Quadratic Regulator (LQR).

However, this two-step approach introduces a significant bottleneck. As Steffen W. R. Werner et al. demonstrate in On the sample complexity of stabilizing linear dynamical systems from data, identifying a unique, high-fidelity system model often imposes a substantially higher data burden than the actual control task requires. Furthermore, any estimation errors introduced during the identification phase propagate directly into the controller design, frequently leading to degraded performance or closed-loop instability (Julian Berberich et al.).

This friction has driven a paradigm shift toward direct data-driven control. Instead of compressing historical data into a parametric model ($A$ and $B$ matrices) and discarding the residuals, can we use raw input-output trajectories directly to predict and control the system's future?

The Core Intuition: Willems' Fundamental Lemma

To understand how we can predict the future without a model, we must look to behavioral systems theory, specifically a foundational result known as Willems' Fundamental Lemma.

Instead of describing a system via internal states, behavioral theory views a system purely through its valid trajectories—the sequences of inputs and outputs it can legally produce. Imagine recording a single, continuous historical trajectory of a linear time-invariant (LTI) system. If the input sequence used during this recording is "persistently exciting"—meaning it moves the system enough to activate all of its internal dynamics—then this single trajectory implicitly contains everything there is to know about the system.

Mathematically, we arrange this historical data into a Hankel matrix, which stacks time-shifted windows of the recorded inputs and outputs. Because the underlying system is linear, the column space of this Hankel matrix defines the entire behavior of the system. Any new, valid input-output trajectory the system could ever generate must live within the column space of this matrix. As Matthias A. Müller et al. show in On a Continuous-Time Version of Willems' Lemma, this property can even be extended to continuous-time systems using time-varying differential parameters.

In short: We do not need to identify $A$ and $B$ because a properly structured matrix of raw data already spans the exact same linear subspace as the true system dynamics.

From Data to Control: The DeePC Mechanism

This insight forms the foundation of Data-Enabled Predictive Control (DeePC), a framework formalized by Jeremy Coulson et al. and analyzed by Julian Berberich et al. in Data-Driven Model Predictive Control with Stability and Robustness Guarantees.

To see how DeePC works online, consider how traditional MPC operates. At each time step, MPC looks at the current state $x_t$ and solves an optimization problem over a future horizon, using $x_{t+1} = Ax_t + Bu_t$ to ensure the planned inputs lead to the desired outputs.

DeePC replaces the parametric model constraint entirely. It partitions the historical Hankel matrix into two main blocks: "past" data ($U_p, Y_p$) and "future" data ($U_f, Y_f$).

Historical Data Matrix (Hankel)
✏️ [  U_p  ]  <- Reconstructs the immediate past (u_past)
✏️ [  Y_p  ]  <- Reconstructs the immediate past (y_past)
───────────────────────────────────────────────────────
🔮 [  U_f  ]  <- Predicts the optimal future inputs (u_future)
🔮 [  Y_f  ]  <- Predicts the optimal future outputs (y_future)

At the current time step, the controller measures the system's immediate past inputs and outputs ($u_{past}, y_{past}$). It then searches for a non-parametric decision vector $g$ that linearly combines the columns of the past data matrices to perfectly match our current situation:

$$U_p g = u_{past}$$ $$Y_p g = y_{past}$$

Because $g$ successfully reconstructs our current state from historical trajectories, Willems' Lemma guarantees that multiplying this same vector $g$ by the future blocks of our data matrix yields an exact prediction of the system's future behavior:

$$u_{future} = U_f g$$ $$y_{future} = Y_f g$$

By optimizing over $g$ and $u_{future}$ simultaneously, DeePC computes optimal control inputs that respect physical constraints without ever invoking an explicit state-space model. In Data-driven model predictive control: closed-loop guarantees and experimental results, Julian Berberich proved that this direct approach provides rigorous closed-loop stability guarantees identical to traditional model-based MPC.

Taming Real-World Noise

The explanation above assumes perfect, noise-free data. In reality, sensor measurements are corrupted by noise. If we strictly enforce $Y_p g = y_{past}$ with noisy data, the optimization vector $g$ will overfit to the noise, causing the controller to output erratic, unstable inputs.

To protect against this, data-driven predictive controllers introduce regularization terms (such as penalizing the norm of $g$) into the objective function. While this initially looks like a heuristic hack to smooth out control actions, the authors of Regularized and Distributionally Robust Data-Enabled Predictive Control provided a profound mathematical justification.

By framing the problem through Distributionally Robust Optimization (DRO), Jeremy Coulson et al. proved that adding a regularizer to $g$ is mathematically equivalent to optimizing performance against the worst-case noise distribution within a bounded Wasserstein ambiguity set. Regularization is not just a smoothing trick; it is an implicit, optimal shield against uncertainty.

[ Noisy Data ] ──> [ Standard DeePC ] ──> Overfitting & Instability
[ Noisy Data ] ──> [ Regularized / DRO ] ──> Robust, Safe Control Guarantees

Alternative strategies handle noise by filtering data streams dynamically. For example, Daniele Alpago et al. designed an Extended Kalman Filter in An Extended Kalman Filter for Data-Enabled Predictive Control that leverages the controller's own optimization structure to clean data online. Meanwhile, Joscha Bongard et al. proved in Robust stability analysis of a simple data-driven model predictive control approach that simply extending the prediction horizon long enough provides inherent, structural robustness against measurement noise.

Scaling and Generalization

Direct data-driven control offers immense freedom from the system identification bottleneck, but it introduces a new challenge: computational complexity. As the historical dataset grows, the size of the Hankel matrix expands, increasing the number of decision variables in the online optimization problem.

To scale these methods to high-frequency applications, researchers are exploring several frontiers:

  1. Algorithmic Acceleration: In Learning-Based Efficient Approximation of Data-enabled Predictive Control, Yihan Zhou et al. utilize differentiable convex programming and algorithm unrolling to learn a fixed-size neural approximation of the DeePC step, achieving significant computational speedups.
  2. Hybrid Paradigms: Rather than choosing between pure models or pure data, Sebastian Zieglmeier et al. propose Semi-Data-Driven MPC (Semi-Data-Driven Model Predictive Control: A Physics-Informed Data-Driven Control Approach). This hybrid approach uses a simplified parametric model to capture core physical laws, while relying on a data-driven component to handle complex, unmodeled residuals.
  3. Stochastic Control: The framework has expanded beyond tracking deterministic setpoints. In Data-Driven Covariance Steering Control Design, Joshua Pilipovsky demonstrates how behavioral systems theory can be adapted to steer entire probability distributions of states safely across uncertain environments using raw data alone.

By treating raw data trajectories as the fundamental representation of a system's physics, direct data-driven control bypasses the structural assumptions and errors of traditional system identification, paving the way for highly adaptive, mathematically rigorous control of complex systems.

Key Foundational Work

Reinforcement Learning Applied to Linear Quadratic Regulation (1992) — contributed 1 concept elements to this direction.
An Ensemble of HMMs for Cognitive Fault Detection in Distributed Sensor Networks (2014) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_model_based_control_dependency_0365b32a,
  title = {model-based control dependency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/model-based-control-dependency/rdr_0365b32a14ae4dc595cc4770d9f2f279},
  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_model_based_control_dependency_0365b32a_v1,
  title = {model-based control dependency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/model-based-control-dependency/rdr_0365b32a14ae4dc595cc4770d9f2f279/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake