Learning-Assisted Optimization for the Stochastic Service Network Design Problem

Service Network Design (SND) problems arise when consolidation carriers plan vehicle or vessel capacities for their transportation networks for an upcoming operating season. The two-stage stochastic SND (SSND) further explicitly considers uncertain demand: carriers commit to offering transportation services in advance (first stage) and may later adapt by procuring additional costlier spot-market capacity (second stage). While more realistic, SSNDs are computationally challenging. In this work, we study the efficacy of using artificial neural networks to assist in solving two-stage SSND problems. The aim is to learn a predictive model that estimates the recourse cost of candidate first-stage SSND solutions under uncertain demand. This trained predictive model can then be used in lieu of the recourse subproblem, with the goal of forming a more tractable hybrid or surrogate optimization model. Results show that effective learning can be achieved: the solutions produced by the surrogate model exhibit an average estimation gap of about 4% between their true expected cost and their surrogate-predicted value. We also observe that estimation quality decreases as instance size increases (for a fixed training set size).