Optimization of Neural Network Training with ELM Based on the Iterative Hybridization of Differential Evolution with Local Search and Restarts

摘要

An Extreme Learning Machine (ELM) performs the training of a single-layer feedforward neural network (SLFN) in less time than the back-propagation algorithm. An ELM defines the input weights and biases of the hidden layer with random values, and then analytically calculates the output weights. The use of random values causes SLFN performance to decrease significantly. The present work carries out the adaptation of three continuous optimization algorithms of high dimensionality (IHDELS, DECC-G and MOS) and compares their performance to each other and with the state-of-the-art method, a memetic algorithm based on differential evolution called M-ELM. The results of the comparison show that IHDELS using a validation model based on retention (Training/Testing) obtains the best results, followed by DECC-G and MOS. All three algorithms obtain better results than M-ELM. The experimentation was carried out on 38 classification problems recognized by the scientific community, while Friedman and Wilcoxon nonparametric statistical tests support the results.