Publication: Investigation on neural network input selection methods
Date
2009-04-01
Authors
Lee, Hooi Khee
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The purpose of this project is about the investigation of neural network input selection methods and its objectives is to eliminate the noise, irrelevant and redundant information that may improve the predictive power of an algorithm and to reduce the amount of data to process and the training time for effectiveness. Initially, a few input selection methods have been investigated and the best input selection method is chosen. The few set of methods include Principle Component Analysis (PCA) and stepwise selection strategies that assess usefulness of the inputs in the model are investigated. From the investigation, PCA is not suitable to be used in this project as this algorithm alter the original representation of the variables which may loss some of the important information. Although PCA is a powerful tool used for dimensionality reduction and feature extraction, this method is not optimal for dimensional reduction in target detection and classification applications. On the other hand, stepwise selection strategy is found out to be an efficient strategy to perform input selection task. Once the appropriate input variables have been selected, NN (neural network) will then be used to test this new training set for classification purpose. Then, the training and testing accuracy are evaluated. The best neural network has to be chosen correctly and efficiently in order to give the best training and testing performance to diagnosis results. Multilayer feedforward neural network (MLP) is chosen in this thesis to perform the clustering purpose due to the fact that MLPs are flexible, non-parametric modelling techniques and allowing us to perform any complex function mapping with arbitrarily desired accuracy. The MatlabR2008a, version 7.6, is the software that was used for the implementation in this project.