Push–Pull with Device Sampling

We consider decentralized optimization problems in which a number of agents collaborate to minimize the average their local functions by exchanging over an underlying communication graph. Specifically, we place ourselves asynchronous model where only random portion nodes perform computation at each iteration, while information exchange can be conducted between all and asymmetric fashion. For this setting, propose algorithm that combines gradient tracking with network-level variance reduction (in contrast within node). This enables track gradients objective functions. Our theoretical analysis shows converges linearly, when are strongly convex, under mild connectivity conditions on expected mixing matrices. In particular, our result does not require matrices doubly stochastic. experiments, investigate broadcast mechanism transmits from computing neighbors, confirm linear convergence method both synthetic real-world datasets.