Distributed robust optimization via Cutting-Plane Consensus

This paper addresses the problem of robust optimization in large-scale networks of identical processors. General convex optimization problems are considered, where uncertain constraints are distributed to the processors in the network. The processors have to compute a maximizer of a linear objective over the robustly feasible set, defined as the intersection of all locally known feasible sets. We propose a novel asynchronous algorithm, based on outer-approximations of the robustly feasible set, to solve such problems. Each processor stores a small set of linear constraints that form an outer-approximation of the robustly feasible set. Based on its locally available information and the data exchange with neighboring processors, each processor repeatedly updates its local approximation. A computational study for robust linear programming illustrates that the completion time of the algorithm depends primarily on the diameter of the communication graph and is independent of the network size.