Dr. Yifan Hu, EPFL

Bin Gao, Associate Professor

Global Optimality for Structured Nonconvex Optimization in Operations and Causality

10:00-11:00 July 16(Wednesday), N202

Nonconvex optimization arises in various domains such as operations, learning, and causality. While nonconvexity poses significant challenges in achieving global optimality, we demonstrate how these barriers can be overcome in several critical applications. In this talk, we first discuss hidden convexity motivated by quantity-based network revenue management problems. Leveraging hidden convexity, i.e., a convex reformulation through a (possibly implicit) variable transformation, we propose methods that achieve global optimality without requiring explicit knowledge of the convex transformation. Moreover, the complexities of these methods match the lower bounds for stochastic convex optimization, implying that they are optimal. Extensive numerical experiments on airline revenue management showcase the superiority of our approach, achieving higher revenue and lower computational costs compared to state-of-the-art bid-price control policies. Beyond hidden convexity, we further explore benign nonconvexity in causal discovery problems and devise efficient algorithms for global optimality. Leveraging structural information from applications, we highlight principles that can drive scalable algorithmic design with practical impacts.

Bio: Yifan Hu is a postdoc researcher jointly advised by Prof. Daniel Kuhn from EPFL and Prof. Andreas Krause from ETH Zurich in Switzerland. Prior to that, he obtained PhD in Operations Research from the University of Illinois at Urbana-Champaign, jointly advised by Prof. Xin Chen and Prof. Niao He. His research interests focus on the foundations of optimization and development of easy-to-implement algorithms for operations and causal inference applications.