Talk: Learning with Feedback from Strategic Stakeholders: Algorithms and Systems

LIVE STREAM: https://uwmadison.zoom.us/j/94584320956?pwd=ZzJ1WnlGM2ZjKzFiZHMwbml0bjk0Zz09

Abstract: Machine learning has made great strides in the recent past, achieving super-human performance in tasks requiring learning from data. However, intelligent algorithms should go beyond simply learning; they should be able to make decisions to collect data in unknown environments and work towards a desired outcome. Quite often, we need to rely on feedback from stakeholders for this data, who themselves have vested interest in the decisions taken by such algorithms. In this talk, I will argue that effective learning and decision-making in such problems requires that we combine ideas from economics with machine learning.

I will present some of our recent work on learning with feedback from strategic stakeholders in fair division and in auctions. We will strive to design algorithms that are efficient (finds optimal outcomes), fair (treats all stakeholders fairly), and strategy-proof (cannot be manipulated by selfish stakeholders) when agent preferences and environment characteristics are unknown. On the theoretical side, I will discuss algorithms, asymptotic upper bounds on the three criteria, and complementary hardness results. On the applied side, I will discuss Cilantro, a Kubernetes-based system for resource allocation in clusters while obtaining feedback from a job's performance. We implement our methods on Cilantro and show that they are able to quickly learn efficient resource allocations while being empirically fair and strategy-proof.

Bio: Kirthevasan is a postdoctoral scholar at the University of California, Berkeley. He completed his PhD in machine learning at Carnegie Mellon University, and his B.Sc at the University of Moratuwa, Sri Lanka. His current research interests are in the intersection of decision-making under uncertainty and game theory with applications in computer systems. He is a recipient of a Facebook fellowship, Siebel scholarship, and an IJCAI best paper award.