Parallel alternating direction method of multipliers

被引：19

作者：

Yan, Jiaqi ^{[1
]}

Guo, Fanghong ^{[2
]}

Wen, Changyun ^{[1
]}

Li, Guoqi ^{[3
]}

机构：

[1] Nanyang Technol Univ, Singapore, Singapore

[2] Zhejiang Univ Technol, Hangzhou, Zhejiang, Peoples R China

[3] Tsinghua Univ, Beijing, Peoples R China

来源：

INFORMATION SCIENCES | 2020年 / 507卷

基金：

中国国家自然科学基金;

关键词：

Distributed optimization; ADMM; Parallel algorithm; DISTRIBUTED OPTIMIZATION; SENSOR NETWORKS; CONSENSUS; ALGORITHM; ADMM; DECOMPOSITION; CONVERGENCE;

D O I：

10.1016/j.ins.2019.08.039

中图分类号：

TP [自动化技术、计算机技术];

学科分类号：

0812 ;

摘要：

In this paper, we consider the distributed optimization problem, where the objective function is the sum of local cost functions. To solve this problem, a new parallel Alternating Direction Method of Multipliers (ADMM) algorithm is developed, which guarantees that the agents cooperatively reach an optimal agreement. Different from most of the existing ADMM approaches, our algorithm allows all the agents to update their local variables simultaneously in a parallel manner. It is theoretically proved that the local solutions of all the agents could reach a consensus, and converge to the optimal solution asymptotically with the rate of O(1/k). Numerical examples are finally provided to validate the effectiveness of the proposed method. (C) 2019 Elsevier Inc. All rights reserved.

引用

页码：185 / 196

页数：12

共 38 条

[1] A zero-gradient-sum algorithm for distributed cooperative learning using a feedforward neural network with random weights [J].