Fast Projection-Free Algorithm for Distributed Online Learning in Networks

Abstract

In order to speed up the convergence of distributed online optimization algorithms, a Fast Distributed Online Conditional Gradient Algorithm (F-DOCG) is proposed in this paper. The Erdos-Renyi (ER) stochastic model is firstly established and an Edge Addition (AE) algorithm is proposed. Secondly, the Edge Addition algorithm and Distributed Online Conditional Gradient Algorithm are combined to propose a F-DOCG. The F-DOCG algorithm not only avoids the high cost projection problem with a linear approximation, but also improves the Regret bound based on the relationship between the underlying topology and the algebraic connectivity, and thus results in a faster convergence rate. Finally, compared with the existing Distributed Online Conditional Gradient Algorithm (DOCG), numerical simulation experiments show that the proposed F-DOCG has better performance.