ORIE Colloquium: Vidya Muthukumar (Cal Berkeley) - Statistical inference in strategic environments

Description

The last few decades have seen the advent of automated online platforms for auctions, advertising, matching markets and mobile healthcare. In these platforms, decisions need to be made sequentially and real-time as a function of data that is provided by people who use these platforms—and thus a central problem is the ability to predict future realizations of data from past history. When this data is generated by people, we cannot, a-priori, assume that it is generated by any statistical model—the people generating the data could have substantial strategic incentives. Related to this, a significant intellectual challenge presents itself in understanding the dynamic incentives of the data generator herself: that is, how should a person manipulate the data that she provides for maximal personal gain?

In this talk, I describe solutions to the above problems in the fundamental framework of online decision-making. I first show, by designing adaptive online learning equivalents of seminal approaches to statistical model selection, that it is possible to adaptively select the environment that best describes the data (including adversarial environments) without knowing the identity of this environment beforehand. Next, I consider a strategic, but non-adversarial agent generating the data that is being used for learning. I introduce a tractable frequentist framework to approximately express such an agent’s incentives and tradeoffs involved in reaching the Stackelberg equilibrium of the ensuing non-zero-sum game. Finally, I outline future intellectual challenges that need to be addressed for a more complete understanding of statistical inference in the presence of strategic behavior, and briefly touch upon how the fundamental theory developed here can be reconciled with the counterintuitive success of modern over-parameterized models.

Bio:
Vidya Muthukumar is a final year graduate student in the EECS department at the University of California, Berkeley, advised by Anant Sahai. Her broad interests are in game theory, online and statistical learning. Recently, she is particularly interested in designing learning algorithms that provably adapt in strategic environments, fundamental properties of over-parameterized models, and fairness, accountability and transparency in machine learning. Her honors include the IBM Research Science for Social Good Fellowship, SanDisk Fellowship and the UC Berkeley EECS Outstanding Course Development and Teaching Award. She served as co-president of UC Berkeley Women in Computer Science and Engineering (WICSE) in the academic year 2016-2017.