Scenario-based Data-Enabled Predictive Control: Robustification via the Scenario Approach

What the math gives to ML

The paper's transferable asset is a data-driven scenario construction: empirical prediction errors are replayed as finite disturbance trajectories, and every candidate prediction is required to satisfy constraints under every sampled scenario. This suggests a scenario-robust training objective for neural dynamics or policy models, particularly when the final decision layer is convex and can inherit scenario-optimization guarantees. The adaptive error buffer is useful because it avoids assuming a parametric noise distribution and can track operating-point-dependent model mismatch. The strongest initial experiment is to freeze a neural feature extractor and optimize a convex output head against residual scenarios, where safety-violation reduction can be measured directly.