Generalized Semi-Infinite Programming for Robust Optimal Control with Decision-Dependent Uncertainty

What the math gives to ML

The paper gives a constructive treatment of universal constraints whose uncertainty set changes with the decision variable, using adaptive discretization rather than repeatedly solving the full robust problem globally. The transferable asset is an active-set loop: optimize against a finite set of difficult scenarios, then search for a newly admissible scenario that violates the current solution and add it only when necessary. In neural networks this enables decision-dependent robustness, where perturbation or environment admissibility depends on the current representation, prediction, policy, or parameters. The most practical experiment is a cutting-plane adversarial-training method with differentiable penalties for the inner feasibility conditions and an independent constrained oracle for evaluation.