COAML for contextual stochastic optimization: Structured Reinforcement Learning and Local Search MCMC Layers

Combinatorial Optimization Augmented Machine Learning (COAML) is a rapidly growing field that integrates machine learning and operations research methods to solve data-driven problems involving both uncertainty and combinatorial structures. These problems frequently arise in industrial settings where organizations leverage large, noisy datasets to optimize operations. COAML embeds combinatorial optimization layers into neural networks and trains them using decision-aware learning techniques. It excels in contextual and dynamic stochastic optimization problems, as demonstrated by its winning performance in the 2022 EURO-NeurIPS Dynamic Vehicle Routing Challenge. This talk will cover three recent contributions: a primal-dual empirical cost minimization algorithm, a structured reinforcement learning extension, and new regularizations that exploit connections between local search and Monte Carlo methods. These algorithms enhance performance, reduce computational costs, and lower data requirements, enabling new large-scale contextual stochastic optimization applications. Additionally, they provide convergence guarantees that support new statistical learning generalization bounds.