Research Article
Najam Ul Qadir*
Abstract
Artificial Neural Networks (ANNs) have generally been observed to learn with a relatively higher rate of convergence resulting in an improved training performance if the input variables are preprocessed before being used to train the network. The foremost objectives of data preprocessing include size reduction of the input space, smoother relationship, data normalization, noise reduction, and feature extraction. The most commonly used technique for input space reduction is Principal Component Analysis (PCA) while two of the most commonly used data normalization approaches include the min-max normalization or rescaling, and the z-score normalization also known as standardization. However, the selection of the most appropriate preprocessing method for a given dataset is not a trivial task especially if the dataset contains an unusually large number of training patterns. This study presents a first attempt of combining PCA with each of the two aforementioned normalization approaches for analyzing the network performance based on the Levenberg-Marquardt (LM) training algorithm utilizing exact formulations of both the gradient vector and the Hessian matrix. The network weights have been initialized using a linear least squares method. The training procedure has been conducted for each of the proposed modifications of the LM algorithm for four different types of datasets and the training performance in terms of the average convergence rate and a proposed performance metric has been compared with the Neural Network Toolbox in MATLAB® (R2017a).