Fusing Deep Learning and Optimization

Pascal Van Hentenryck, PhD
Georgia Institute of Technology

Abstract: The fusion of deep learning and optimization has the potential to deliver outcomes for engineering applications that the two technologies cannot achieve independently. This talk illustrates this potential with the concept of optimization proxy, a differentiable program that can produce, in milliseconds, feasible (or near-feasible) and near-optimal solutions to classes of optimization problems. The talk reviews some of the foundations underlying optimization proxies, including end-to-end learning, compact optimization learning, dual learning, and self-supervised learning. The benefits of optimization proxies are demonstrated on applications in power systems.

Biography: Pascal Van Hentenryck is the director of the NSF AI Institute for Advances in Optimization (AI4OPT) and the A. Russell Chandler III Chair and Professor at the Georgia Institute of Technology. Several of his optimization systems have been in commercial use for more than 20 years. His current research focuses on AI for Engineering, fusing machine learning and optimization for applications in energy systems, supply chains and manufacturing, and mobility. Van Hentenryck is a fellow of AAAI and INFORMS, and the recipient of numerous research and teaching awards. He was also a Ulam fellow at the Los Alamos National Laboratories.

Attendance is mandatory for in-person seminar students. For online and part-time students, seminars will be recorded and made available through Canvas.