Optimized Regression in Sensor Networks by Integration of Harmony Search and Particle Swarm Optimization

Regression modeling in sensor networks is a difficult task due to (i) the network data is distributed among the nodes and (ii) the restricted capabilities of the sensor nodes, such as limited power supply and bandwidth capacity. Recently, some distributed approaches have been proposed based on gradient descent and Nelder-Mead simplex methods. Although in these methods, the energy consumption is low, but the accuracy is still far from the centralized approach. Also, they suffer from a high latency. In this paper, a two-fold distributed approach has been proposed for doing regression analysis in wireless sensor networks. After clustering the network, the regressor of each cluster is learned by the integration of particle swarm optimization and harmony search. Afterwards, cluster heads collaborate to construct the global network regressor using a weighted averaging combination rule. The experimental results show the proposed approach improves the accuracy and latency significantly while its energy consumption is considerably acceptable in comparison with its popular counterparts.