Event Date
Speaker: Eric Mazumdar, Assistant Professor in Computing and Mathematical Sciences & Economics, CalTech
Title: "Rethinking Machine Learning for Strategic Environments"
Abstract: Machine learning algorithms are increasingly being deployed into environments in which they must interact with other strategic agents with potentially misaligned objectives. While the presence of these strategic interactions creates new challenges for learning algorithms, they also give rise to new opportunities for algorithm design.
In this talk, I will begin by highlighting a line of work showing the unintuitive behaviors that arise from the interplay between learning algorithms and strategic agents. First, I will show---both in theory and practice---how learning algorithms (even those built on top of large language models) are susceptible to the gaming of data by vanishingly small groups of people, even if individuals have no effect on them in isolation. I will also present recent work on how strategic interactions can break our basic intuition that larger models, more data, and more compute always improves performance. The resulting Braess paradox-like phenomenon suggests that---even when one has access to infinite data---strategic interactions can make smaller and less expressive models yield better equilibrium outcomes. I will conclude with some recent work on algorithm design in strategic environments in the context of multi-agent RL. In particular, I will show how to tweak deep Q learning to allow it to have strong convergence guarantees in competitive games.
Speaker's webpage (links to calTech): http://users.cms.caltech.edu/~mazumdar/
Seminar Date/Time: Thursday March 14, 4:10pm
Location: MSB 1147 (refreshments at 3:30pm, MSB courtyard)