Online Learning Over Dynamic Graphs via Distributed Proximal Gradient Algorithm

We consider the problem of tracking minimum a time-varying convex optimization over dynamic graph. Motivated by target and parameter estimation problems in intermittently connected robotic sensor networks, goal is to design distributed algorithm capable handling nondifferentiable regularization penalties. The proposed proximal online gradient descent built run fully decentralized manner utilizes consensus updates possibly disconnected graphs. performance analyzed developing bounds on its regret terms cumulative path length optimum. It shown that as compared centralized case, incurred $T$ time slots worse factor notation="LaTeX">$\log (T)$ only, despite network topology. empirical tested sparse recovery problem, where it incur close algorithm.