Asymptotic Optimality in Data-driven Decision Making
Abstract
We study statistically optimal decision-making for general stochastic optimization problems, focusing on data generated by abstract stochastic processes, including heterogeneous sources and evolving mechanisms. Our approach constructs decisions with probably vanishing shifted regret risk at a prescribed exponential rate under minimal model perturbations. This leads to a multi-objective optimization problem balancing statistical consistency with asymptotic risk decay, governed by a generalized rate function from large deviation theory. Our framework encompasses and extends classical methods such as distributionally robust optimization. We illustrate its applicability with two examples from operations research: the newsvendor problem and portfolio optimization under aleatoric uncertainty from heterogeneous data.
About the speaker
Tobias Sutter received his B.Sc. and M.Sc. degrees in Mechanical Engineering from ETH Zürich, Switzerland, in 2010 and 2012, respectively, and earned a Ph.D. in Electrical Engineering from the Automatic Control Laboratory at ETH Zürich in 2017. Since 2025, he has been an Associate Professor in the Department of Economics at the University of St.Gallen. From 2021 to 2025, he was an Assistant Professor in the Department of Computer Science at the University of Konstanz. He previously held research and teaching positions at EPFL with the Chair of Risk Analytics and Optimization, and at the Institute for Machine Learning at ETH Zürich. His research interests lie in data-driven robust optimization, reinforcement learning, and dynamic decision-making under uncertainty. Tobias Sutter received the George S. Axelby Outstanding Paper Award from the IEEE Control Systems Society in 2016 and the ETH Medal in 2018 for his outstanding Ph.D. thesis on approximate dynamic programming.
Language
English
This is a free seminar. Registration is mandatory.
Supported by the Fond National de la Recherche,
Luxembourg (19441346)