*Presentation will be recorded.
Abstract:
In this talk, we revisit distributed optimization and learning problems where a group of agents is interested in optimizing a cost function that is decomposable as a sum of local cost functions while each agent is aware of its own local cost function. We first discuss a unified framework to distributed optimization solvers that unifies many existing approaches to solve these problems and show how it can be used to synthesize new algorithms. In the second part of the talk, we discuss an often neglected aspect of such algorithms, i.e., assuming unlimited bandwidth and neglecting the communication constraints between the agents. For this, we propose and study a two-time-scale decentralized gradient descent algorithm to reach a consensus on an optimizer of the objective function. One time scale fades out the imperfect incoming information from the neighboring agents, and the second one adjusts the local loss functions' gradients. We show that under certain conditions on the connectivity of the underlying time-varying network and the time-scale sequences, the algorithm converges to an optimal point of the cost function.
Bio:
Behrouz Touri is an Associate Professor of Electrical and Computer Engineering at the University of California San Diego. He received his B.Sc. degree in Electrical Engineering from Isfahan University of Technology, Isfahan, Iran in 2006, his M.Sc. degree in Coding Theory from Jacobs University, Bremen, Germany, in 2008, and his Ph.D. degree in Industrial Engineering from the University of Illinois at Urbana-Champaign in 2011. Between 2011 and 2014, he was a postdoctoral researcher with the University of Illinois and Georgia Institute of Technology, and from 2014 to 2017, he was an Assistant Professor of Electrical Engineering with the University of Colorado Boulder. His research interests include applied probability theory, distributed optimization, learning, and computation, and game theory. He is the recipient of the American Control Council's Donald P. Eckman Award in 2018.
