Machine Learning Thesis Defense
- Remote Access Enabled - Zoom
- Virtual Presentation
- RITESH NOOTHIATTU
- Ph.D. Student
- Machine Learning Department
- Carnegie Mellon University
Machine Learning and Multiagent Preferences
Multiagent preferences are the basic ingredient in some of the most well-studied areas of algorithmic game theory -- computational social choice and fair division. In classical social choice and voting, there are n agents and m candidates, and each of these agents has a preference ordering or ranking over these m alternatives. Given the rankings of all the n agents, the goal is to find a winning candidate (or a consensus ranking) that is the most "fair"' outcome. In this thesis, we look into several variants of this standard setting. For instance, the domain may have an uncountably infinite number of alternatives, each of them defined by a set of d features. In such a setting, we cannot elucidate the entire ranking over the alternatives for any given voter, and hence need to learn this ranking by collecting a few pairwise comparisons from them, followed by generalizing to the rest of the alternative space. We also look into whether it is meaningful to pool all the comparisons together and learn a single community wide ranking directly, instead of learning individual rankings for the voters, followed by aggregating them.
Other variants we consider are settings where agents' preferences are very different from the realm of rankings over candidates. For instance, we consider the setting of a markov decision process, where we have multiple agents, but each of them with a different reward function (representing their preferences for this problem). Our goal is then to find a single policy that everyone would be "happy" with. Yet another example is the setting of a conference peer review system, where the agents are the reviewers of the conference, and their preferences are given by the defining characteristics they use to choose which papers are to be accepted at the conference. Finally, we also consider the setting where agents' preferences are defined by utility functions over a given set of outcomes, and our goal is to learn a classifier that is fair with respect to these preferences.
Broadly speaking, this thesis tackles problems in three areas: (i) fairness in machine learning, (ii) voting and social choice, and (iii) reinforcement learning, each of them handling multiagent preferences with machine learning. My defense talk will primarily focus on the multiagent inverse reinforcement learning & inverse bandit setting (Chapter 9), and the study of axioms satisfied by learning on pooled pairwise comparisons (Chapter 6).
Ariel D. Procaccia (Chair) (Harvard University)
Nihar B. Shah
Milind Tambe (Harvard University)
Zoom Participation Enabled. See announcement.