Speaker:

Peter Marbach
Professor
Department of Computer Science
University of Toronto

Talk Title:

Learning to Optimize at Scale: Reinforcement Learning for Adaptive Distributed Systems

Date and Location:

Friday, March 27, 2026

3–4 p.m.

BA 7231

There is no registration required to attend this event in person. However, seating is limited, so arriving early is recommended.

Abstract:

What if large-scale compute systems could learn to optimize themselves without constant human tuning? As reinforcement learning (RL) becomes increasingly powerful, it is tempting to replace traditional control and optimization algorithms with RL agents. But can RL be safely and predictably deployed in large-scale distributed environments? In this talk, we examine what it would mean to treat RL as a systems primitive. RL promises adaptive resource allocation, congestion control, scheduling, and coordination, yet multi-agent learning in distributed environments raises fundamental challenges: instability under feedback loops, non-stationarity, emergent coordination failures, and unpredictable global dynamics. To reason about these challenges, we turn to a surprising but rigorous source: the formal study of self-organizing systems. Large-scale social systems offer tractable models of decentralized adaptation under local feedback. These models highlight structural principles that matter for distributed RL systems: alignment of local reward with global performance, emergence of stable coordination norms, and the role of feedback topology in convergence. This talk is forward-looking and exploratory, aimed at identifying the systems research questions that must be solved to make large-scale self-optimizing infrastructure viable.

Biography:

Peter Marbach was born in Lucerne, Switzerland. He received the Eidg. Dipl. El.-Ing. (1993) from the ETH Zurich, Switzerland, the M.S. (1994) in electrical engineering from the Columbia University, NY, U.S.A, and the Ph.D. (1998) in electrical engineering from the Massachusetts Institute of Technology (MIT), Cambridge, Massachusetts, U.S.A. He was appointed as assistant professor in 2000, and associate professor in 2005, at the Department of Computer Science of the University of Toronto. He has also been a visiting professor at Microsoft Research, Cambridge, UK, at the Ecole Polytechnique Federal at Lausanne (EPFL), Switzerland, and at the Ecole Normale Superieure, Paris, France, and a post-doctoral fellow at Cambridge University, UK. Peter Marbach has received the IEEE INFOCOM 2002 Best Paper Award for his paper "Priority Service and Max-Min Fairness". He is on the editorial board of the ACM/IEEE Transactions of Networking. His research interests are in the fields of communication networks, in particular in the area of wireless networks, peer-to-peer networks, and online social networks. He is a member of the Systems and Networking Group and the Computer Networks Group in the Department of Computer Science at the University of Toronto.