Distributed Stochastic Optimization: Variance Reduction and Edge-Based Method

Abstract

The abstraction of distributed optimization is to achieve optimal decision making or control by local manipulation with private data and diffusion of local information through a network of computational nodes. Due to the promising prospects in machine learning, statistical computation [1, 2], and extensive applications for power systems, sensor networks, and wireless communication networks [3, 4], distributed optimization has harvested many attentions over the years. Most of issues arisen in these fields are cast as distributed optimization problems, in which nodes of a network collaboratively optimize a global objective function through operating on their local objective functions and communicating with their neighbors only.