Numerical method for approximately optimal solutions of two-stage distributionally robust optimization with marginal constraints

Abstract: We consider a general class of two-stage distributionally robust optimization (DRO) problems which includes prominent instances such as task scheduling, the assemble-to-order system, and supply chain network design. The ambiguity set is constrained by fixed marginal distributions that are not necessarily discrete. We develop a numerical algorithm for computing approximately optimal solutions of such problems. Through replacing the marginal constraints by a finite collection of linear constraints, we derive a relaxation of the DRO problem which serves as its upper bound. We can control the relaxation error to be arbitrarily close to 0. We develop duality results and transform the inf-sup problem into an inf-inf problem. This leads to a numerical algorithm for two-stage DRO problems with marginal constraints which solves a linear semi-infinite optimization problem. Besides an approximately optimal solution, the algorithm computes both an upper bound and a lower bound for the optimal value of the problem. The difference between the computed bounds provides a direct sub-optimality estimate of the computed solution. Most importantly, one can choose the inputs of the algorithm such that the sub-optimality is controlled to be arbitrarily small. In our numerical examples, we apply the proposed algorithm to task scheduling, the assemble-to-order system, and supply chain network design. The ambiguity sets in these problems involve a large number of marginals, which include both discrete and continuous distributions. The numerical results showcase that the proposed algorithm computes high-quality robust decisions along with their corresponding sub-optimality estimates with practically reasonable magnitudes that are not over-conservative.