In classical control theory, we first identify a mathematical model of a system and then design a controller for it. Modern "data-driven" control skips the identification step, using raw historical trajectories to predict future behavior directly. While powerful, this approach faces a fundamental "data curse": as you collect more data to improve accuracy, the optimization problems you must solve to control the system become increasingly massive and slow.
A growing body of research is focused on complexity decoupling—restructuring these data-driven frameworks so that the computational cost depends on the system's physical complexity (its "rank" or "order") rather than the total volume of data collected.
The Bottleneck of Data-Enabled Control
Most modern data-driven methods, such as Data-Enabled Predictive Control (DeePC), rely on Willems’ Fundamental Lemma. This lemma proves that for a linear system, any possible trajectory can be expressed as a linear combination of a sufficiently rich set of past trajectories.
In practice, this means the controller solves an optimization problem where the decision variables are the weights assigned to historical data points. As Mohammad Alsalti, Julian Berberich, and Matthias A. Müller demonstrate in Data-based Control of Feedback Linearizable Systems, this allows us to control complex nonlinear systems by "lifting" them into a space where the lemma applies.
However, this creates a scalability crisis. If you have $N$ data points, your optimization problem typically has $N$ decision variables. In real-time applications like autonomous flight or high-frequency industrial sensing, $N$ grows rapidly, leading to what Shihao Li and Dongmei Chen describe as a cubic scaling of computational cost in Datamodel-Based Data Selection for Nonlinear Data-Enabled Predictive Control.
Decoupling Problem Size from Data Volume
The core intuition behind the recent breakthroughs is that while we might have thousands of data points, the underlying physical system usually has a much lower dimensionality. We don't need a decision variable for every observation; we only need enough to span the system's behavior.
Researchers have proposed several mechanisms to achieve this decoupling:
- Kernel and Rank-Based Representations: In Sample- and computationally efficient data-driven predictive control, Mohammad Alsalti and colleagues introduce eDDPC. They replace large data matrices with compact kernel representations. This ensures the number of decision variables scales only with the system's order and the prediction horizon, allowing the controller to utilize massive datasets without any increase in per-step latency.
- Recursive SVD Updates: Jicheng Shi and Colin N. Jones take a different approach in Efficient Recursive Data-enabled Predictive Control. They use Singular Value Decomposition (SVD) to compress the data into its essential "rank" and then use fast rank-1 updates to incorporate new data online. This achieved a 60x speedup in simulations by ensuring the optimization complexity remains constant even as the history grows.
- Algebraic Optimization: For nonlinear systems, Robin Strässer and Frank Allgöwer show in Data-Driven Control of Nonlinear Systems: Beyond Polynomial Dynamics that formulating the problem through algebraic optimization allows the decision variables to remain independent of the training dataset size.
Robustness Without Complexity
A major challenge in data-driven control is "Error-in-Variables" (EIV)—the fact that both your historical data and your current measurements are noisy. Standard robust optimization often makes the scalability problem worse, as it requires checking "worst-case" scenarios for every data point.
Jared Miller, Mario Sznaier, and Roy S. Smith address this in Data-Driven Structured Robust Control of Linear Systems. They utilize a "set-membership" approach that collapses the uncertainty of the entire dataset into a single structured optimization problem. Similarly, Felix Brändle uses the Sherman-Morrison-Woodbury formula in A System Parametrization for Direct Data-Driven Analysis and Control with Error-in-Variables to transform noisy data into a Linear Fractional Transformation (LFT). This allows for guaranteed performance bounds using a single optimization that doesn't get harder to solve as more data is added.
Scaling to High-Dimensional Safety
The drive for efficiency extends to safety-critical and stochastic domains. In Scalable Reachset-Conformant Identification of Linear Systems, Laura Lützow demonstrates how switching from halfspace-based to generator-based representations of zonotopes can turn an exponentially difficult reachability problem into a polynomial one.
In the realm of stochastic control, George Rapakoulias shows in Discrete-time Optimal Covariance Steering via Semidefinite Programming that lossless convex relaxations can speed up "uncertainty steering" by an order of magnitude. This allows robots to manage their own state distributions (e.g., how "spread out" their possible positions are) in real-time.
Why This Matters
The transition from model-based to data-driven control is often compared to the transition from rule-based AI to machine learning. However, control systems operate in physical time. If a controller takes 100ms to solve an optimization problem but the system requires a 10ms response to stay stable, the controller fails regardless of its accuracy.
By decoupling optimization complexity from data volume, this research direction enables a new generation of "plug-and-play" controllers—like the gene expression regulators proposed by Liam Perreault and Idris Kempf in Host-Aware Control of Gene Expression using Data-Enabled Predictive Control—that can learn from massive amounts of data while remaining fast enough for the real world.