D-pso for Distributed Regression over Wireless Sensor Networks

In this paper, a novel distributed approach based on particle swarm optimization (PSO) has been proposed for distributed regression over sensor networks. Besides distributed data, the limitations of sensor nodes make doing regression more difficult. Conventional methods employ numerical optimization techniques such as gradient descent or Nelder-Mead Simplex algorithms in which the sensor nodes collaborate through a pre-established Hamiltonian path. Although NM Simplex based approaches converge faster than the gradient counterparts, both of them suffer from low accuracy and high latency. In the proposed approach, denoted as D-PSO (Distributed PSO), a swarm of particles is dedicated to each cluster and the cluster regressor is learned. Afterwards, the clusters regressors are sent to the fusion center in order to build the global model. To do this, weighted averaging combination rule, which comes from MCS (Multiple Classifier Systems) concept, is applied on the received clusters regressors. The experimental results show that the proposed approach has a superior performance in terms of the prediction accuracy, latency, and energy efficiency compared to its counterparts.