Multi-output gaussian processes applied to modeling and reducing dimensionality of data transmitted in a wireless sensor network

This work presents data modelling from Wireless Sensor Networks (WSN) applying Multiple Output Gaussian Processes (MOGP). The objective, in addition to describing the dynamics of sensed magnitudes (temperature, relative humidity, atmospheric pressure, and soil humidity), is to exploit the probabilistic learning nature of MOGPs and use the variance (uncertainty) provided by this method to reduce dimensionality in the network data, meaning the elimination of redundant data. It is compared with other machine learning models such as simple Gaussian Processes (GP), Support Vector Regressor (SVR), Neural Networks (NN), and Random Forest (RF), in terms of Root Mean Squared Error (RMSE) for modeling real data sensed by the WSN located at Universidad del Quindío. Results demonstrate that MOGPs are highly accurate supervised learning algorithms, flexible in modeling any physical magnitude, and capable of detecting redundant data, in some cases achieving a reduction greater than 50 %.