Traditional control engineering relies on a elegant paradigm: first, write down differential equations that describe a system's physics (system identification), and second, use those equations to calculate optimal control inputs (Model Predictive Control, or MPC).
However, modern complex systems—such as weak electrical grids, multi-stack fuel cells, or building energy hubs—are notoriously difficult to model accurately. Components degrade over time, and environmental dynamics shift unpredictably.
To bypass this bottleneck, a powerful alternative has emerged: Data-Driven Predictive Control (DDPC), exemplified by frameworks like Data-Enabled Predictive Control (DeePC). Instead of using an explicit mathematical model, DDPC uses a library of raw historical input-output trajectories as a non-parametric surrogate model to predict future behavior and optimize control actions.
But this approach introduces a fundamental vulnerability: real-world data is noisy. Without a physical model to anchor it, how can an optimization algorithm distinguish between true system dynamics and random sensor noise? The answer lies in regularization.
The Core Mechanism: Trajectory Recombination
To understand why noise is fatal to vanilla data-driven control, we must look at how these controllers predict the future.
Rooted in behavioral systems theory, frameworks like DeePC rely on Willems' Fundamental Lemma. As explained by Markovsky et al., this lemma proves that any valid future trajectory of a linear system can be represented exactly as a linear combination of its past observed trajectories.
Mathematically, the controller structures historical data into a large matrix (often a Hankel matrix) and solves an optimization problem over a data-weighting vector, typically called $g$. This vector $g$ acts as a set of instructions: it tells the controller how to recombine past pieces of data to perfectly reconstruct the current state and project the optimal future path.
Historical Data Matrix Weight Vector Predicted Trajectory
[ . . . . ] [ g_1 ] [ Future ]
[ . . . . ] × [ g_2 ] = [ Control ]
[ . . . . ] [ g_3 ] [ Outputs ]
If the historical data is perfectly clean, this formulation yields exact, optimal control. However, when the data matrix is corrupted by measurement noise, the optimization process breaks down.
The Problem: Overfitting to Noise
When sensor noise is present, the optimization algorithm encounters a phenomenon known as overfitting. Because the data matrix contains random stochastic fluctuations, the optimizer will attempt to exploit these high-frequency variations to minimize its tracking error.
To perfectly match the noisy historical data, the algorithm often selects extreme, rapidly oscillating values for the weight vector $g$. As highlighted by Huang et al. and Berberich et al., these inflated weights cause the controller to become hyper-sensitive to noise. The resulting control actions become erratic, degrading performance and threatening closed-loop stability. In essence, the controller mistakes sensor glitches for real physical capabilities of the system.
Why Regularization Works: Bounding the Weights
To prevent the optimizer from exploiting noise, we must restrict its freedom. This is achieved by introducing a regularization term directly into the optimization cost function.
Instead of simply minimizing tracking error and control effort, the objective function is modified to penalize the norm of the weight vector $g$:
$$\min_{u, g} \text{Control Cost}(u) + \lambda \|g\|_p$$
where $\lambda$ is a tuning parameter and $\|\cdot\|_p$ represents a norm (typically $L_1$ or $L_2$).
By penalizing large or complex combinations of weights, regularization forces the controller to find smooth, low-energy combinations of historical data. It prevents the algorithm from picking massive, opposing weights to cancel out high-frequency noise.
Different regularization choices yield distinct behaviors:
- Tikhonov ($L_2$) Regularization: As shown by Berberich et al., applying an $L_2$ penalty suppresses noise sensitivity and ensures practical stability in real-world environments by smoothing out the elements of $g$.
- $L_1$ Regularization: Discussed in Data-Enabled Predictive Control: In the Shallows of the DeePC, an $L_1$ penalty encourages sparsity, forcing the controller to focus only on the most relevant historical patterns while ignoring noisy, unhelpful trajectories.
The Distributional Robustness Equivalence
Why does this mathematical adjustment work so reliably? A profound insight from Coulson et al. reveals that norm-based regularization is not just a heuristic trick; it is mathematically equivalent to Distributionally Robust Optimization (DRO) under a Wasserstein metric.
When you penalize the norm of $g$, you are implicitly instructing the controller to optimize against the worst-case noise perturbation within a bounded trajectory space. As Allgöwer et al. frame it, the regularization parameter acts as a quantifiable safety margin, providing explicit noise rejection without requiring the user to model the exact noise distribution.
Maintaining Feasibility with Slack Variables
Noise does not just distort predictions; it can also make the optimization problem mathematically impossible to solve. If a sensor report is artificially inflated by noise, it might appear to violate physical system constraints (such as voltage or pressure limits), causing the controller to fail due to infeasibility.
To guarantee that the controller always finds a valid solution, regularization is paired with slack variables ($\sigma$). These variables act as shock absorbers that softly relax the optimization constraints when noise pushes the system near its boundaries.
Strict Constraint: System State ≤ Hard Limit (Can fail with noise)
Relaxed Constraint: System State ≤ Hard Limit + Slack (σ)
By penalizing the size of these slacks in the cost function, the controller keeps them as close to zero as possible. Bongard et al. and Liu et al. demonstrate that this combination of trajectory weight regularization and constraint relaxation ensures recursive feasibility—meaning the controller remains mathematically stable and operational over indefinite horizons, even under heavy noise or cyber-attacks.
Generalizing to Complex Dynamics
The intuition of regularized data-driven control extends far beyond simple linear systems. When dealing with nonlinearities or specialized operational constraints, regularization adapts to protect the controller from model mismatch:
- Nonlinear Approximations: When controlling feedback linearizable or flat nonlinear systems, researchers use dictionaries of basis functions to lift the data into a linear space. Alsalti et al. and Alsalti et al. show that regularization simultaneously mitigates sensor noise and errors stemming from imperfect basis functions. Similarly, Xiong et al. use regularization to insulate Koopman bilinear realizations against truncation errors caused by approximating infinite-dimensional operators.
- Closed-Loop Data: If historical data is collected while a system is already running under an existing feedback controller, the noise becomes correlated with the inputs, biasing standard data-driven models. To decouple this, Wang et al. introduced an instrumental variable (IV) inspired regularizer, balancing the tracking objective with an accurate fit of closed-loop data.
- Adaptive Disturbance Rejection: For classical engineering problems like canceling repeating, periodic noise without a system model, Shahsavari et al. utilize a ridge-style cost function to penalize large control gains, ensuring the adaptive signal does not overreact to unmodeled dynamics.
By embedding regularization directly into the data-driven optimization loop, modern predictive controllers achieve the best of both worlds: the flexibility of a model-free approach and the rigorous safety guarantees of classical control theory.
Key Foundational Work
Adaptive Rejection of Periodic Disturbances Acting on Linear Systems with Unknown Dynamics (2016) — contributed 1 concept elements to this direction.
DEMAND SHIFTING USING MODEL-ASSISTED CONTROL (2013) — contributed 1 concept elements to this direction.