Selecting the best artificial neural network model from a multi-objective Differential Evolution Pareto front

The objective of this work is to select artificial neural network models (ANN) automatically with sigmoid basis units for multiclassification tasks. These models are designed using a Memetic Pareto Differential Evolution Neural Network algorithm (MPDENN) based on the Pareto dominance concept. We propose different methodologies to obtain the best model from the Pareto front obtained with the MPDENN algorithm. These methodologies are based on choosing the best models for training in both objectives, the Correct Classification Rate and Minimum Sensitivity, and the two models closest to the centroids of two clusters formed with the models of the first and second Pareto fronts. These methodologies are compared with three standard ensembles methodologies with very competitive results.