Emulation Alternating Direction Method of Multipliers

Generally, centralized version of an algorithm performs better as compared to its decentralised counter parts. So, decentralizing an algorithm, while imitating the centralized version, could preserve its certain convergence properties. In this paper, we propose a novel method to completely decentralize Consensus-ADMM (C-ADMM) algorithm and try to mimic its convergence properties, by emulating the functionality of the central coordinator. We show that our algorithm behaves similar to noise induced ADMM and converges sub-optimally, in practice. We also give the bound on sub-optimality and ways to achieve desired accuracy while using our algorithm.