Flat Chance! Using Stochastic Gradient Estimators to Assess Plausible Optimality for Convex Functions
David J. Eckman, Matthew Plumlee, Barry L. Nelson · 2021
This paper studies methods that identify plausibly near-optimal solutions based on simulation results obtained from only a small subset of feasible solutions. We do so by making use of both noisy estimates of performance and their gradients. Under a convexity assumption on the performance function, these inference methods involve checking only a system of inequalities. We find that these methods can yield more powerful inference at less computational expense compared to methodological predecessors that do not leverage stochastic gradient estimators.