Decentralized Optimization for Multi-Agent Networks Qing Ling - PowerPoint PPT Presentation

Decentralized Optimization for Multi-Agent Networks Qing Ling Department of Automation, University of Science and Technology of China (USTC) Joint work with Wotao Yin (UCLA), Wei Shi and Kun Yuan (USTC) 2014 Workshop on Optimization for Modern Computation 2014/09/02 1

Outline 2

Multi-agent networks 3

Decentralized consensus optimization 4

Example: target localization 5

Decentralized versus distributed optimization 6

Related work 7

Assumptions 8

Decentralized gradient descent (DGD) 9

Mixing matrix 10

Existing convergence analysis 11

Can we reach consensus? 12

Essence of DGD 13

When gradients are bounded? 14

Where to converge and how fast? 15

Concluding DGD 16

EXact firsT-ordeR Algorithm (EXTRA) 17

Mixing matrices 18

Limit properties 19

Explanations of EXTRA 20

Sublinear convergence 21

Linear convergence 22

Simulation settings 23

Simulation of DGD and EXTRA 24

Concluding EXTRA 25

Future research directions 26

Thank you 27