Skip to main content

Clone of Abstract: Creating Automated Workflows for Functional Materials Discovery

Abstract: An Introduction to Bilevel Optimization: Theory, Algorithms, and Applications

Bilevel optimization is an important class of hierarchical decision-making where one optimization problem is nested within another. This modeling framework finds extensive applications across diverse domains, including machine learning, energy systems, and economics. However, the nested structure makes these problems challenging: even linear bilevel problems are NP-hard, the feasible region is typically nonconvex and nonsmooth, and standard optimality conditions may fail. As the scale and complexity of these systems continue to increase, there is a critical need for a solid theoretical understanding and for scalable, reliable algorithms.

In this talk, we first review the basics of optimization, including optimality conditions, Lagrangian duality, and constraint qualifications, with an emphasis on what these conditions mean. We then introduce bilevel optimization and its applications in machine learning, such as hyperparameter learning, meta-learning, and adversarially robust learning, as well as in power systems. Next, we discuss single-level reformulations and their theoretical pitfalls. We also derive optimality conditions for two cases: when the lower-level problem is strongly convex and when it is not. Finally, we present two of our recent first-order methods: a projection-free inexact conditional gradient method for constrained bilevel problems, and a sequential QCQP method with line search for problems whose lower level may have multiple solutions. For each method, we sketch the key steps of the convergence analysis to show how such guarantees are established.

 

Bio:

Dr. Yazdandoost Hamedani is an Assistant Professor in the Systems and Industrial Engineering Department at the University of Arizona and a member of Applied Math and Statistics & Data Science GIDPs. He received a bachelor’s degree in mathematics and applications from the University of Tehran, Tehran, Iran, in 2015 and a Ph.D. degree in Industrial Engineering and Operation Research with minor in statistics from Pennsylvania State University in August 2020. His research focuses on developing and analyzing algorithms for solving various optimization problems including saddle point problems, distributed optimization, and bilevel optimization. His research has been supported by NSF and Arizona TRIF.