NAVER LABS Europe virtual seminars are open to the public. Please register here for your participation (Zoom event).
Date: 18th March 2025, 11:00 am (CET)
Self-improvement methods in neural combinatorial optimization
About the speaker: Jonathan is a third-year PhD student at the GrimmLab Professorship of Bioinformatics at the Technical University of Munich. Before starting his PhD, Jonathan worked as an AI consultant at Fujitsu, mainly in the area of computer vision. He received his B.Sc. and M.Sc. in Mathematics from the University of Regensburg. His research focuses on building bridges between reinforcement learning and self-supervised learning to solve complex planning problems.
Abstract: Combinatorial optimization is central to many real-world challenges, from logistics and manufacturing to finance and genomics. Traditional approaches rely on expert-designed heuristics, yet the relatively young field of Neural Combinatorial Optimization lets neural networks learn these strategies from data. When solutions are built step by step within this framework, Reinforcement Learning is almost a natural fit – provided we can efficiently search for better and better solutions during training. However, as network architectures scale to solve more and more problems, search efforts during training must remain feasible. Starting from lessons in AlphaZero, this talk explores expert iteration methods that align with the scale of modern sequence models, highlighting both their potential and limitations. Along the way, we’ll make diversions into applications in process flowsheet design for chemical engineering and molecular design.