Semi-infinite programming using Gaussian process regression for robust design optimization
Julia Stecher, Lothar Kiltz, Knut Graichen · 2022 European Control Conference (ECC) · 2022
A design centering approach to robust optimization is presented that determines the largest tolerance box within the feasible region which is defined by performance functions. The associated semi-infinite optimization problem is solved with learning-based optimization algorithms. In every iteration of the tolerance maximization problem, the feasibility of the tolerance box is evaluated based on Gaussian process (GP) models of the performance functions. The GP models are adaptively learned during the optimization and a constrained acquisition optimization problem is solved to find promising boxes. This approach is compared to a straight-forward baseline method where separate GP models for the feasibility of a box and for the performance functions are maintained. Numerical evaluations of both approaches are presented and they are compared in terms of sample efficiency and computational effort.