Use this URL to cite or link to this record in EThOS:
Title: Simplified neural networks algorithms for function approximation and regression boosting on discrete input spaces
Author: Haider, Syed Shabbir
ISNI:       0000 0004 2718 0558
Awarding Body: University of Manchester
Current Institution: University of Manchester
Date of Award: 2011
Availability of Full Text:
Access from EThOS:
Access from Institution:
Function approximation capabilities of feedforward Neural Networks have been widely investigated over the past couple of decades. There has been quite a lot of work carried out in order to prove 'Universal Approximation Property' of these Networks. Most of the work in application of Neural Networks for function approximation has concentrated on problems where the input variables are continuous. However, there are many real world examples around us in which input variables constitute only discrete values, or a significant number of these input variables are discrete. Most of the learning algorithms proposed so far do not distinguish between different features of continuous and discrete input spaces and treat them in more or less the same way. Due to this reason, corresponding learning algorithms becomes unnecessarily complex and time consuming, especially when dealing with inputs mainly consisting of discrete variables. More recently, it has been shown that by focusing on special features of discrete input spaces, more simplified and robust algorithms can be developed. The main objective of this work is to address the function approximation capabilities of Artificial Neural Networks. There is particular emphasis on development, implementation, testing and analysis of new learning algorithms for the Simplified Neural Network approximation scheme for functions defined on discrete input spaces. By developing the corresponding learning algorithms, and testing with different benchmarking data sets, it is shown that comparing conventional multilayer neural networks for approximating functions on discrete input spaces, the proposed simplified neural network architecture and algorithms can achieve similar or better approximation accuracy. This is particularly the case when dealing with high dimensional-low sample cases, but with a much simpler architecture and less parameters. In order to investigate wider implications of simplified Neural Networks, their application has been extended to the Regression Boosting frame work. By developing, implementing and testing with empirical data it has been shown that these simplified Neural Network based algorithms also performs well in other Neural Network based ensembles.
Supervisor: Zeng, Xiaojun Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available