Logarithmic Regret for Distributed Online Subgradient Method over Unbalanced Directed Networks

Abstract

This paper investigates a constrained distributed online optimization problem over strongly connected communication networks, where a local cost function of each agent varies in time due to environmental factors. We propose a distributed online projected subgradient method over unbalanced directed networks. The performance of the proposed method is evaluated by a regret which is defined by the error between the cumulative cost over time and the cost of the optimal strategy in hindsight. We show that a logarithmic regret bound can be achieved for strongly convex cost functions. We also demonstrate the validity of the proposed method through a numerical example on distributed estimation over a diffusion field.