Abstract:

Robust optimization provides a mathematical framework for modeling and computing solutions to decision-making problems under worst-case uncertainty.  In this talk I will present recent work in two-stage robust optimization (2RO) problems, wherein first-stage and second-stage decisions are made before and after uncertainty is realized.  This results in a nested min-max-min optimization problem, which generally means that we are dealing with computationally challenging problems, especially in case of integer decisions.  Together with my co-authors, we propose Neur2RO, an efficient machine learning-based algorithm. We learn to estimate the value function of the second-stage problem via a neural network architecture designed to construct an easy-to-solve surrogate optimization problem.  Our computational experiments on two 2RO benchmarks demonstrate that we can find near-optimal solutions among different sizes of instances, often within orders of magnitude less computing time.