Data-driven predictive control replaces a hand-built physics model with recorded input-output data. At every step, the controller searches over combinations of past trajectories to plan the next few moves. When the plant drifts or is nonlinear, practitioners keep only a sliding window of recent data, so the controller's implied model changes every step. That is where stability proofs break down. Each new batch of data defines a slightly different predictor, and nothing prevents that predictor from wandering to a model under which the controller's built-in stability certificate is false. We propose a remedy borrowed from online learning, where iterates are projected back into a safe set after each update. The projection target here is a convex set of models defined by the certificate itself, and the projection uses a data-weighted distance. The controller keeps what the data actually says and changes only what the certificate needs, mostly in the directions the data never excited.
The opportunity is concrete. Bridging Direct and Indirect Data-Driven Control Formulations via Regularizations and Relaxations showed that a projector-based regularizer $(I-\Pi)g$ is far more robust to tuning than standard norm penalties. That regularizer penalizes predicted trajectories lying outside the subspace spanned by the least-squares fit to data, where $g$ is the vector of weights on past trajectories. The analysis, however, is open-loop, and $\Pi$ is fixed. Closed-loop proofs for data-driven MPC, such as Data-driven model predictive control: closed-loop guarantees and experimental results, use generic norm penalties instead. We join the two by making the model that defines $\Pi$ an online parameter estimate, projected at every step onto the set of models compatible with a fixed Lyapunov certificate.
Core Claim
Stabilizing parameter sets are normally non-convex, which the research file lists as a central obstacle to projection-based safety. The set of stabilizing feedback gains is one example. The set of models that a given certificate stabilizes, however, is convex. Fix a terminal cost matrix $P \succ 0$, a terminal feedback gain $K$, and a contraction rate $\rho<1$. Then the set of models $\theta=[A\;B]$ of the extended state (built from a window of past inputs and outputs) satisfying
$$\Theta_P=\Big\{\theta:\ \begin{bmatrix}\rho P & (A+BK)^\top P\\ P(A+BK) & P\end{bmatrix}\succeq 0\Big\}$$
is a linear matrix inequality in $(A,B)$, so it is convex. The online model update is then a constrained least-squares problem,
$$\hat\theta_t=\arg\min_{\theta\in\Theta_P}\ \|Y_t-\theta\Phi_t\|_F^2+\mu\|\theta-\hat\theta_{t-1}\|_F^2 .$$
Here $\Phi_t$ and $Y_t$ are the regressor and next-state data in the current window, and $\mu>0$ is a small proximal weight. The weight $\mu$ keeps the problem well-posed when the data stop exciting some directions.
The projected model $\hat\theta_t$ then defines the regularizer's projector $\Pi_{\hat\theta_t}$ and pairs with the terminal cost $x^\top P x$. The claim has three parts. First, because every model the controller ever uses satisfies the same certificate, the usual recursive-feasibility and Lyapunov-decrease arguments survive the online data updates. This yields practical exponential stability, and the size of the residual neighborhood scales with noise level and with how far the true local linearization sits from $\Theta_P$. Second, under loss of excitation the data-weighted norm is nearly blind to unexcited directions, so the projection sets those directions to certificate-consistent values. Iterates therefore do not freeze on an arbitrary box boundary, which is the failure mode flagged for the projected estimator in Adaptive Economic Model Predictive Control for linear systems with performance guarantees. Third, the size of the projection residual in excited directions is a measurable signal that the certificate itself is wrong and should be re-synthesized.
Research Program
The program has three parts: theory, algorithm, and diagnostics.
The theory extends the terminal-ingredient arguments summarized in An overview of systems-theoretic guarantees in data-driven model predictive control to a time-varying predictor constrained to $\Theta_P$. The key lemma bounds one-step prediction mismatch between $\hat\theta_t$ and the plant. That bound combines the least-squares error, the projection displacement weighted by the current data, and the drift term $\|\hat\theta_t-\hat\theta_{t-1}\|$, which the proximal weight $\mu$ controls. With the lemma, we would prove recursive feasibility under tightened constraints and practical exponential stability. We will also state the honest failure case: if the plant's local linearization leaves $\Theta_P$, the guarantee is lost. The residual diagnostic is there to detect exactly that.
The algorithm solves one small semidefinite constrained least-squares problem per sampling interval, and warm starting should make most steps cheap. This follows the few-iterations-per-sample spirit of Real-time Projected Gradient-based Nonlinear Model Predictive Control with an Application to Anesthesia Control, which uses projected gradient steps with a Lyapunov stopping rule on a known model. The second piece of the algorithm is a slow outer loop. When the excited-direction residual exceeds a threshold for a dwell time, it re-synthesizes $(P,K,\rho)$ from the current data. A hysteresis rule prevents rapid switching between certificates.
In role terms, the file's papers play three parts here. The Bridging paper supplies the regularizer. The data-driven MPC guarantees paper supplies the closed-loop proof template and the four-tank benchmark. The adaptive economic MPC paper motivates the work, since its box projection bounds parameters but does not keep them stabilizing and it assumes an open-loop-stable plant. Separately, Learning Over Contracting and Lipschitz Closed-Loops for Partially-Observed Nonlinear Systems argues that stability-by-construction removes the need for projection. That argument covers policy parameters and a fixed plant. Our problem is an online predictor, and there the projection is doing real work.
Baselines
The naive reference is sliding-window data-driven MPC with no regularizer beyond a small ridge term and a fixed terminal cost. It shows what happens when the implied model is left to drift freely.
The strongest existing method is the regularized data-driven MPC of Data-driven model predictive control: closed-loop guarantees and experimental results, with norm regularization on the trajectory weights $\alpha$, terminal equality constraints, and a sliding window for nonlinear plants. It is the strongest competitor because it already carries closed-loop guarantees and hardware validation on exactly our benchmark. That paper reports a four-tank simulation closed-loop tracking cost of $1.42\times10^5$ against $3.1\times10^4$ for model-based nonlinear MPC. It will run on identical data windows, horizons, constraints, and noise realizations. Its $\lambda_\alpha$ gets the same tuning budget as our $\lambda$ and $\mu$: a log-spaced sweep of equal size.
Beating it means lower closed-loop cost at matched constraint satisfaction, with no instability anywhere in the sweep. We will also report the fraction of the sweep that remains stable, since tuning robustness is the Bridging paper's main qualitative claim. For that claim the file gives no numeric value, so no reported result was found on Lacuna for the projector regularizer's stable range.
The second baseline is the Bridging paper's open-loop projector regularizer made receding-horizon in the obvious way, with $\Pi$ recomputed each step from raw least squares and no certificate projection. The third is indirect adaptive MPC with LMS estimates projected onto a box of plausible parameters, as in the adaptive economic MPC paper, which reports no explicit numbers on its building benchmark.
Two ablations isolate the key ingredient. The first replaces the certificate set $\Theta_P$ with a box of matched volume. The second replaces projection with accept/reject gating, which keeps the previous model whenever the new least-squares model fails the certificate.
Why not just…?
Why not just use norm-regularized data-driven MPC, which already has closed-loop proofs? Those proofs treat the data as fixed or tolerate window updates only through a noise-like bound. The Bridging paper reports that $\ell_1$ penalties destabilize when slightly too small and destroy performance when slightly too large. The lambda sweep will show whether our method's stable region is actually wider. If the region is not wider on the four-tank and Lotka-Volterra benchmarks, the proposal's practical case fails.
Why not just reject model updates that violate the certificate? This alternative is the real threat. It is simple, cheap, and resembles the verified-fallback logic of Approximate Nonlinear Model Predictive Control With Safety-Augmented Neural Networks. The problem is that rejection discards all new information whenever any direction fails, so adaptation freezes precisely when the plant is moving. The gating ablation settles the question. It uses an operating-point-change scenario on the four-tank system and on Lotka-Volterra, and we will measure post-change tracking cost and time to recover. If gating matches projection there, the projection machinery is not justified and we would report that.
Why not just identify a model offline and run robust tube MPC? The Bridging paper shows indirect LTI fits degrade on Lotka-Volterra through structural bias. A fixed offline model also cannot follow drift. We include this baseline on the drifting and nonlinear scenarios. We expect it to win on the stationary 5th-order LTI system, and we will say so if it does.
Why not just use a stability-by-design policy class? Youla-REN guarantees do not hold under bounded unknown additive disturbances, according to Learning Over Contracting and Lipschitz Closed-Loops for Partially-Observed Nonlinear Systems. The ICNN variant in A modular framework for stabilizing deep reinforcement learning control was shown only on a stable two-tank plant and still requires persistent excitation. Neither addresses a constrained predictive controller whose model updates online, so we treat them as related work rather than as direct baselines.
Proposed Experiments
The first experiment uses the nonlinear four-tank system in simulation, which is also the benchmark in the data-driven MPC guarantees paper. It includes setpoint changes that move the local linearization, and measurement noise at several levels. Metrics are closed-loop tracking cost, hard output-constraint violations, the fraction of steps on which the model in use satisfies the certificate, and solve time against the 1.5-second sampling interval. If hardware access allows, the same protocol runs on the physical Quanser rig.
The second experiment uses the Bridging paper's 5th-order LTI and Lotka-Volterra systems. These test high-noise behavior and structural nonlinearity, and they host the lambda-sweep comparison across several orders of magnitude.
Because the adaptive economic MPC work was limited to open-loop-stable plants, the third experiment takes open-loop-unstable plants from the adapted COMPleib library used in Controller Design via Experimental Exploration With Robustness Guarantees. These are run with slow parameter drift.
The fourth experiment tests loss of excitation. It holds each plant at a setpoint long enough for the data window to become rank-deficient, then commands a step. We compare three things: post-step overshoot and cost, the condition number of the data window, and how far box-projected, gated, and certificate-projected models sit from the true local linearization in the unexcited directions.
The decisive checks are three. The certificate-projected controller must stay stable across the full lambda sweep where the norm-regularized baseline does not. It must recover from the post-excitation-loss step with lower cost than gating. And its excited-direction residual must rise before, not after, instability in a deliberately induced case where the plant leaves $\Theta_P$. That third check validates the re-synthesis trigger.
References
- Bridging Direct and Indirect Data-Driven Control Formulations via Regularizations and Relaxations
- Data-driven model predictive control: closed-loop guarantees and experimental results
- Adaptive Economic Model Predictive Control for linear systems with performance guarantees
- An Overview of Systems-Theoretic Guarantees in Data-Driven Model Predictive Control
- Real-time Projected Gradient-based Nonlinear Model Predictive Control with an Application to Anesthesia Control
- Learning Over Contracting and Lipschitz Closed-Loops for Partially-Observed Nonlinear Systems
- Approximate Nonlinear Model Predictive Control With Safety-Augmented Neural Networks
- A modular framework for stabilizing deep reinforcement learning control
- Controller Design via Experimental Exploration With Robustness Guarantees