EDGE CONSENSUS COMPUTING FOR HETEROGENEOUS DATA SETS

Edge consensus computing is a framework to optimize a cost function when distributed nodes have distinct data sets available to them. The primal-dual method of multipliers (PDMM) is an optimization algorithm that forms a consensus among nodes by exchanging latent variables rather than the data sets. PDMM often has a high rate of convergence. However, when the nodes see statistically heterogeneous data sets then the performance of PDMM degrades. To overcome this problem, we propose quadratic PDMM. In this method, the original cost functions are replaced by their quadratic majorization based on the L2 norm to ensure homogeneous convexity among nodes. We describe a method to set its parameters optimally for fast convergence. Our experiments confirm that the proposed quadratic PDMM provides good performance even when the data sets are heterogeneous.