Event
CCSP Seminar: Amrit Singh Bedi, ARL, Distributed & Online Learning with Stochastic Gradient Methods
Thursday, April 11, 2019
5:00 p.m.-6:30 p.m.
AVW 2168
Ajaykrishnan Nageswaran
301 405 3661
ajayk@umd.edu
http://www.ece.umd.edu/seminars/ccsp/
Communication, Control and Signal Processing Seminar
Distributed and Online Learning with Stochastic Gradient Methods
Amrit Singh Bedi
Army Research Laboratory
Abstract
The advancements in sensing and communication capabilities of recent technologies such as wireless communication and networks have lead to unprecedented growth in the complexity and bandwidth requirements of the network services in multi-agent network systems. The resulting stress on the network infrastructure has motivated the network designers to move away from simpler or modular architectures to optimum ones. The problem is challenging for three reasons: first, data is persistently arriving in a sequential manner from an unknown environment, second, in multi-agent systems, the agents of the network may correspond to different classes of objects, and third, each agent in a network has a different timescale requirement. These challenges motivate the need for online decentralized optimization algorithms which can robustly handle network delays and packet losses. This work deals with the development of online, asynchronous, and distributed stochastic convex optimization algorithms to address various challenges in multi-agent network systems. Based on the basic framework of the proposed algorithm, we categorize the work into three parts.
The first part details the work done in classical online settings where the network state is sequentially revealed. We develop the first memory-efficient stochastic algorithms for compositional online learning with kernels (COLK). The convergence guarantees of COLK are established, and experiments with robust formulations of supervised learning are demonstrated. The second part focuses on the development of asynchronous and distributed stochastic optimization algorithms for multi-agent heterogeneous networks. The third part considers the problem of online learning in the time-varying adversarial environments. We formulate a target tracking problem as a time-varying optimization problem and puts forth an inexact online gradient descent algorithm for solving it sequentially. The performance of the proposed algorithm is studied by characterizing its dynamic regret. This part also considers a problem of designing the user trajectory in a device-to-device setting. The theoretical results are backed by detailed numerical tests that establish the efficacy of the proposed algorithms under various settings.
