Distributed Alternating Direction Multiplier Method Based on Optimized Topology and Nodes Selection Strategy

ADMM (The Alternating Direction Multiplier Method) has been widely used in large-scale distributed machine learning, however the classical distributed ADMM algorithm lacks of optimization for network characteristics, which makes the convergence speed not reach the optimal. In this paper, A distributed ADMM with optimized network topology and nodes selection strategy is proposed. The proposed algorithm improves the performance of distributed computing by optimizing the topology structure of the network and reasonably allocating the node identity to avoid the large delay edges in the ADMM iterations as much as possible. The experimental results shown that based on proposed algorithm the calculation time of ADMM is improved, while the accuracy is controlled within an acceptable range. Compared with the classical distributed ADMM algorithm, the convergence time of distributed ADMM algorithm can be reduced by about 60% to 70%. The results also shown that the proposed algorithm is more prominent in the scenario of high delay or dynamics networks.