Certified Stochastic Control via Covariance Steering with Pick-to-Learn

What the math gives to ML

The paper couples covariance steering with Pick-to-Learn, an adaptive constraint-learning procedure that repeatedly evaluates policies in a high-fidelity stochastic simulator and adds the worst observed violations to a surrogate constraint set. Its transferable asset is not covariance steering itself, but the closed-loop mechanism of using adversarially selected rollout failures to tighten a tractable surrogate while retaining a compression-based probabilistic bound on true violation probability. For neural networks, this suggests a safety-aware training loop that actively mines rare or worst-case simulator trajectories, converts them into differentiable penalties or shield constraints, and reports a finite-sample violation certificate rather than relying on nominal rollout averages.