Orateur
Description
In this talk, we will bridge the gap between combinatorial optimization and machine learning to derive policies for contextual multi-stage decision-making problems that arise in various stochastic settings, including transportation, control, and supply chain management. We will discuss how to encode effective policies by embedding combinatorial optimization layers into neural networks and training them with decision-aware learning techniques. Specifically, I will provide an overview of the underlying algorithmic pipelines and foundations, and elucidate two paradigms - learning by experience and learning by imitation - to train the pipeline’s statistical models in an end-to-end fashion. I will demonstrate the efficacy of optimization-augmented machine learning pipelines for selected application cases, among others discussing its winning performance on the 2022 EUROMeetsNeurIPS dynamic vehicle routing challenge. Finally, we will put the presented paradigms into perspective and learn how they can also be used to approximate (static) equilibrium problems, focusing on traffic equilibria as a use case.
