Scalable Distributed Optimization with Separable Variables in Multi-Agent Networks

Abstract

Robotics, signal processing, and other disciplines involve distributed data collection and storage for state estimation, control, and predictive modeling using optimization. We consider large-scale optimization problems in which multiple agents with limited resources communicate over a network to obtain the optimal variables of the centralized problem. In this work, we present the Separable Optimization Variable ADMM (SOVA) method where each agent optimizes only over a subset of the optimization variables relevant to its data or role, avoiding unnecessary optimization over all the problem variables. We demonstrate superior convergence rates of the SOVA method compared to previous distributed ADMM methods. Further, we show applications of the SOVA method to robotics and data modeling.