A General Framework for Optimal Data-Driven Optimization

Aug 10, 2021 03:00 PM Singapore (Registration will open at 02:50 PM.)

Join Zoom Meeting:
https://sutd-edu.zoom.us/j/91336956727?pwd=TjNSQjh6M0FkTndxS2pZakxTdDVpZz09

Meeting ID: 913 3695 6727
Passcode: G2XHFk^D

Abstract

In the setting of data-driven stochastic programming, the decision-maker cannot observe the distribution of its exogenous uncertainties but rather has access to a finite set of data generated by a stationary process. In this talk, we characterize optimal decisions as the least conservative decisions that guarantee an exponential decay of its corresponding out-of-sample disappointment, where the out-of-sample disappointment quantifies the probability that the actual expected cost of the proposed decision exceeds its predicted cost. We show that under certain mild technical assumptions optimal decisions can be constructed via a separation into an estimation and a subsequent robust optimization phase. Our framework is an attempt rating the zoo of ambiguity sets derived in the recent literature in terms of statistical power.

The paper is available at: https://arxiv.org/pdf/2010.06606.pdf

About the Speaker

Tobias Sutter received a B.Sc. and M.Sc. degree in 2010 and 2012 from ETH Zurich, and a Ph.D. degree in Electrical Engineering at the Automatic Control Laboratory, ETH Zurich in 2017. He will start a position as an Assistant Professor at the Computer Science Department in Konstanz, Germany in fall 2021. Before, he held a research appointment with EPFL at the Chair of Risk Analytics and Optimization and at the Institute of Machine Learning at ETH Zurich.

His research interests revolve around control and data-driven robust optimization. He was a recipient of the 2016 George S. Axelby Outstanding Paper Award from the IEEE Control Systems Society and received the ETH Medal for his outstanding Ph.D. thesis in 2018.

For more information about the ESD Seminar, please email esd_invite@sutd.edu.sg