Outliers detection and correction for cooperative distributed online learning in Wireless sensor network

The recent distributed online convex optimization framework has developed in Wireless sensor networks (WSN) provide the promising approach for solving approximately stochastic optimization problem over network of sensors follows distributed manner. In practice, most of real environmental sensing activities are highly dynamic where noisy sensory information often appears and affects to the learning performance. However, the original distributed saddle point (DSPA) algorithm is lack of considering about the consequence of falsification in online learning. Based on the simulation observations conducted in this paper, we figure out the fluctuation and the slow convergence rate leads to overall prediction performance reduction of distributed online least square problem. Therefore, we propose an integrated outliers detection, correction mechanism in order to stabilize prediction and improve convergence rate.