An alternative approach to solve convergence problems in the backpropagation algorithm

Abstract

The multilayer perceptron network has become one of the most used in the solution of a wide variety of problems. The training process is based on the supervised method where the inputs are presented to the neural network and the output is compared with a desired value. However, the algorithm presents convergence problems when the desired output of the network has small slope in the discrete time samples or the output is a quasi-constant value. The proposal of this paper is presenting an alternative approach to solve this convergence problem with a pre-conditioning method of the desired output data set before the training process and a post-conditioning when the generalization results are obtained. Simulations results are presented in order to validate the proposed approach.