Distributionally Robust Optimization is a Multi-Objective Problem

Abstract: Distributionally Robust Optimization (DRO) is a worst-case approach to decision making when there is model uncertainty. Though formulated as a single-objective problem, we show that it is intrinsically multi-objective in that DRO solutions map out a near-Pareto-optimal frontier between expected cost and a measure of robustness called worst-case sensitivity (WCS). We take this as the starting point and explore robust decision making through a multi-objective lens. We show that WCS is a measure of spread and derive WCS for a collection of uncertainty sets commonly used in DRO. These sensitivity measures identify the errors against which the nominal expected cost is most vulnerable and the uncertainty set for the worst-case problem that most effectively mitigates it. The associated mean-sensitivity frontier is used to select its size. The multi-objective perspective provides a quantitative measure of robustness and a sensitivity-based approach to addressing important conceptual gaps in DRO -- how to choose the family and size of uncertainty sets for a given cost distribution, and how this affects the solution.