Dual Coordinate Descent Algorithms for Multi-agent Optimization

Multi-agent optimization problems arise in a wide variety of networked systems, and are often required to be solved in an asynchronous and uncoordinated way. However, existing asynchronous algorithms for constrained multi-agent optimization do not have guaranteed convergence rates and, thus, lack performance guarantees in on-line applications. This paper addresses this shortcoming by developing randomized coordinate descent algorithms for solving the dual of a class of constrained multi-agent optimization problems. We show that the algorithms can be implemented asynchronously and distributively in multi-agent networks. Moreover, without relying on the standard assumption of boundedness of the dual optimal set, the proposed dual coordinate descent algorithms achieve sublinear convergence rates of both its primal and dual iterates in expectation. The competitive performance is demonstrated numerically on a constrained optimal rendezvous problem.