Decision-Focused AI: From Predictions to Better Business Decisions

Information Systems and Operations Management 

Intervenant: Maximilian Schiffer (TUM)

Salle: Bernard Ramanantsoa

Abstract

In this talk, I demonstrate how decision-focused AI bridges the gap between prediction and action, enabling firms to improve complex multi-stage decisions in transportation, operations, and supply chain management. I show how embedding optimization layers into neural networks and training them with decision-aware techniques allows AI systems to directly learn policies that optimize business outcomes. I present two paradigms — learning from experience and learning from imitation — as foundations for training such models in an end-to-end fashion. I illustrate their impact through applications in inventory management, production scheduling, mobility-on-demand, and large-scale logistics, including their winning performance in the EURO Meets NeurIPS dynamic vehicle routing challenge. I conclude by demonstrating how the introduced decision-focused learning paradigm can also be applied to approximate equilibrium problems, with traffic equilibria serving as a compelling example.