Projection-Regularized Indirect Data-Driven Predictive Control

What the math gives to ML

The paper offers a constructive projection-regularization mechanism for data-driven prediction under errors-in-variables, together with a finite-sample confidence radius that remains valid under temporally correlated closed-loop data. Its transferable asset is not the specific predictive controller, but the combination of projecting noisy empirical predictors onto a regularized data-consistent subspace and propagating the resulting covariance into conservative constraints. In neural-network training, this can become a curvature- or gradient-subspace projection that suppresses poorly identified update directions. A second transfer is uncertainty-aware trust-region training or inference, where parameter updates and outputs are tightened according to an online martingale-style uncertainty radius rather than using a fixed heuristic margin.