Learn to Optimize: Self-Supervised Progressive Hedging and Adaptive Penalties for Multistage Stochastic Network Recovery Planning

This work explores using self-supervised machine learning to solve large and computationally difficult multi-stage stochastic programs (MSSP). Recent work on integrating machine learning into existing algorithms to solve MSSPs has shown promise in reducing computational time with minimal reduction in solution quality. Two broad frameworks have emerged that utilize machine learning for solving optimization problems: the Predict-then-Optimize framework, in which accurate predictions serve as inputs to stochastic optimization problems, and the Learn-to-Optimize framework, where ML models are used to either predict solutions or algorithm parameters directly, or to improve algorithm structure or model formulation via classification or clustering models. Typically, these methods have two primary drawbacks: they are supervised methods and so require generating training data, which is computationally extensive, and they learn solutions to single scenarios or algorithm parameters at a time.