Talk: Structured Decision-Making under Uncertainty: Optimization and Generalization under Constraints

LIVE STREAM: https://uwmadison.zoom.us/j/92454245885?pwd=N3BTM1h6Wks5L0ZIamdCUTZFWlMwQT09 

Abstract: Problems in decision-making under uncertainty often need to be tackled under limitations such as lack of high quality data with good coverage, random and adversarial perturbations, incomplete knowledge of the underlying model, and limited resources. This talk will outline a framework to address these challenges based on advances in optimization theory and statistical learning. Drawing from examples in machine learning, cyber-physical systems, and operations research, we discuss two key innovations: (1) Efficient algorithms for constrained optimization problems that arise in optimal resource allocation and decision-making under a dynamic environment or worst-case perturbations; (2) Generalizable algorithms to guarantee that decision policies learned from imperfect historical data can generalize to unseen data. 

In (1), we discuss optimal first-order algorithms for solving large-scale constrained optimization and minimax problems in nonconvex settings. By utilizing structural properties, specifically error bounds/perturbation bounds, we develop simple algorithms to achieve optimal iteration complexity. In (2), we focus on designing algorithms for offline reinforcement learning that are both tractable and generalizable. Again, by deriving and leveraging the structure of error bounds, we develop a constrained optimization approach to help us find a nearly optimal policy from imperfect data with nearly optimal sample complexity. Finally, we briefly discuss progress in contextual optimization. Here, we leverage constrained optimization techniques and structural properties to learn good policies using contextual data, even under model misspecification. These developments set the stage for my future research in the theory, algorithms, and implementation of optimization problems under “complex constraints” that are necessary to represent a dynamic and uncertain environment.

Bio: Jiawei Zhang is currently a postdoctoral scholar at MIT’s Laboratory for Information & Decision Systems (LIDS), working with Prof. Asuman Ozdaglar and Prof. Saurabh Amin. He earned his Ph.D. degree in Computer and Information Engineering from the Chinese University of Hong Kong, Shenzhen, under the supervision of Prof. Zhi-Quan (Tom) Luo. Previously, he obtained a B.Sc. in Mathematics from the University of Science and Technology of China through the Hua Loo-Keng Talent Program. His research interests include nonlinear and convex optimization, robust and generalizable learning algorithms, data-driven decision-making under uncertainty, and computational models for AI-driven platforms, sustainable energy systems, and signal processing.