A parallel methodology using radial basis functions versus machine learning approaches applied to environmental modelling

Abstract

Parallel nonlinear models using radial kernels on local mesh support have been designed and implemented for application to real-world problems. Although this recently developed approach reduces the memory requirements compared with other methodologies suggested over the last few years, its computational cost makes parallelisation necessary, especially for big datasets with many instances or attributes. In this work, several strategies for the parallelisation of this methodology are proposed and compared. The MPI communication protocol and the OpenMP application programming interface are used to implement the algorithm. The performance of this methodology is compared with various machine learning methods, with particular consideration of techniques using radial basis functions (RBF). Different methods are applied to model the daily maximum air temperature from real meteorological data collected from the Agroclimatic Station Network of the Phytosanitary Alert and Information Network of Andalusia, an autonomous community of southern Spain. The obtained goodness-of-fit measures illustrate the effectiveness of this nonlinear methodology, and its training process is shown to be simpler than those of other powerful machine learning methods.