Distributed discrete-time optimization by exchanging one bit of informations

摘要

This paper proposes a distributed discrete-time algorithm to solve an additive cost optimization problem over undirected deterministic or time-varying graphs. Different from most previous methods that require to exchange exact states between nodes, each node in our algorithm needs only the sign of the relative state between its neighbors, which is clearly one bit of information. Our analysis is based on optimization theory rather than Lyapunov theory or algebraic graph theory. The latter is commonly used in existing literature, especially in the continuous-time algorithm design, and is difficult to apply in our case. Besides, an optimization-theory-based analysis may make our results more extendible. In particular, our convergence proofs are based on the convergences of the subgradient method and the stochastic subgradient method. Moreover, the convergence rate of our algorithm can vary from O(1/ln(k)) to O(1/√k), depending on the choice of the stepsize. A quantile regression problem is included to illustrate the performance of our algorithm using simulations.