MMS-PSO for distributed regression over sensor networks

Regression is one of the effective techniques for data analysis in a WSN. Besides distributed data, the limited power supply and bandwidth capacity of nodes makes doing regression difficult in WSNs. Conventional methods, which employ some numerical optimization techniques such as Nelder-Mead simplex and gradient descent, generally work in a pre-established Hamiltonian path among the nodes. Low estimation accuracy and high latency are common shortcomings appear in these approaches. In this paper, we propose a distributed approach based on PSO, denoted as MMS-PSO (Multi Master Slave PSO), for regression analysis over sensor networks. Accordingly, after clustering the network each cluster is initially dedicated a swarm. The swarm of cluster, which sponsors learning the regressor of cluster, is equally distributed amongst the member nodes and consequently optimized through optimization of the sub-swarms (slaves). To guarantee the convergence of the cluster's swarm, some sharing points are placed between the sub-swarms via designated cluster head (master). After completion of in-cluster optimizations, each cluster head sends its regressor to the fusion center. Finally, the fusion center uses weighted averaging combination rule to combine the received regressors for constructing the final model. Our evaluation and results show that the proposed approach has quite better performance in terms of the estimation accuracy, latency and energy efficiency compared to its counterparts.