TY - GEN
T1 - An approach to distributed parametric learning with streaming data
AU - Liu, Ji
AU - Liu, Yang
AU - Nedić, Angelia
AU - Başar, Tamer
N1 - Funding Information:
This research was supported in part by Office of Naval Research (ONR) MURI Grant N00014-16-1-2710, and in part by US Army Research Office (ARO) Grant W911NF-16-1-0485. J. Liu is with the Department of Electrical and Computer Engineering, Stony Brook University (ji.liu@stonybrook.edu). Y. Liu is with the School of Engineering and Applied Sciences, Harvard University (yangl@seas.harvard.edu). A. Nedić is with the School of Electrical, Computer, and Energy Engineering, Arizona State University (angelia.nedich@asu.edu). T. Bas¸ar is with the Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign (basar1@illinois.edu).
Publisher Copyright:
© 2017 IEEE.
PY - 2018/1/18
Y1 - 2018/1/18
N2 - This paper presents an approach to solve a class of distributed parametric learning problems in a multi-agent network. Each agent acquires its private streaming data to establish a local learning model. The goal is for each agent to converge to a common global learning model, defined as the average of all local ones, by communicating only with its neighbors. Neighbor relationships are described by a time-dependent undirected graph whose vertices correspond to agents and whose edges depict neighbor relationships. It is shown that for any sequence of repeatedly jointly connected graphs, the approach leads all agents to asymptotically converge to the common global learning model, and the worst-case convergence rate is determined by the speed of local learning. A distributed linear regression problem and a distributed belief averaging problem are presented as illustrative examples.
AB - This paper presents an approach to solve a class of distributed parametric learning problems in a multi-agent network. Each agent acquires its private streaming data to establish a local learning model. The goal is for each agent to converge to a common global learning model, defined as the average of all local ones, by communicating only with its neighbors. Neighbor relationships are described by a time-dependent undirected graph whose vertices correspond to agents and whose edges depict neighbor relationships. It is shown that for any sequence of repeatedly jointly connected graphs, the approach leads all agents to asymptotically converge to the common global learning model, and the worst-case convergence rate is determined by the speed of local learning. A distributed linear regression problem and a distributed belief averaging problem are presented as illustrative examples.
UR - http://www.scopus.com/inward/record.url?scp=85046117325&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85046117325&partnerID=8YFLogxK
U2 - 10.1109/CDC.2017.8264129
DO - 10.1109/CDC.2017.8264129
M3 - Conference contribution
AN - SCOPUS:85046117325
T3 - 2017 IEEE 56th Annual Conference on Decision and Control, CDC 2017
SP - 3206
EP - 3211
BT - 2017 IEEE 56th Annual Conference on Decision and Control, CDC 2017
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 56th IEEE Annual Conference on Decision and Control, CDC 2017
Y2 - 12 December 2017 through 15 December 2017
ER -