Abstract:
Learning and Robustness with Decision-Dependent Distribution Shifts
Data-driven methods promise significant real-world impact in terms of what we are able to do. For example, learning methods can discover control policies for a small number of autonomous vehicles to influence an entire traffic network, reducing congestion and emissions. However, once we deploy these decision-making processes in the real world, we have to consider the closed-loop effects: how will the underlying data-generating processes change? For example, suppose your data-driven method discovers that people who do X are less likely to default on a loan, and you deploy a decision-making process based on this. People will then intentionally do X in order to receive a loan, and this observation loses its statistical value.
In this talk, I focus on analyzing decision-dependent distribution shifts in settings where it is difficult to explicitly model these changes. We provide a theoretical framework to show when our decision-making processes are robust enough to handle certain classes of distribution shift, inspired by perturbation analysis in control theory. First, I'll talk about the guarantees we can recover in the case of static learning, and then on the case where we are trying to safely control a dynamical system in the presence of unknown state- and input-dependent noise. Finally, I'll talk about how some of these ideas can be applied in improving congestion and emissions in transportation systems, where the current infrastructure does not allow us to directly actuate vehicles but only influences them via advisory messages.
Bio:
Roy Dong is an Assistant Professor in the Industrial & Enterprise Engineering department at the University of Illinois at Urbana-Champaign. He received a BS Honors in Computer Engineering and a BS Honors in Economics from Michigan State University in 2010. He received a PhD in Electrical Engineering and Computer Sciences at the University of California, Berkeley in 2017, where he was funded in part by the NSF Graduate Research Fellowship. Prior to his current position, he was a postdoctoral researcher in the Berkeley Energy & Climate Institute, a visiting lecturer in the Industrial Engineering and Operations Research department at UC Berkeley, and a Research Assistant Professor in the Electrical and Computer Engineering department at the University of Illinois at Urbana-Champaign. His research uses tools from control theory, economics, statistics, and optimization to understand the closed-loop effects of machine learning, with applications in cyber-physical systems such as the smart grid, modern transportation networks, and autonomous vehicles.
