Modified clustering algorithms for radial basis function networks
Loading...
Date
2006-06
Authors
Eng Aik, Lim
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Some training algorithms for training radial basis function networks are presented
in this dissertation. These training algorithms are based on clustering methods
which have a number of advantages such as faster training time and reduced
network architecture compared to the standard radial basis function network.
These proposed training algorithms were implemented in the radial basis function
network created by the newrb function from MATLAB. These modified version
of clustering radial basis function (RBF) networks were tested against the standard
radial basis function networks and standard clustering RBF networks in
forecasting. Since the use of neural network techniques as a predictor is
considered as a reliable yet cost-effective method for the purpose, the
experimental models were tested on three real world application problems,
particularly in Air pollutant problem, biochemical oxygen demand (BOD)
problem and phytoplankton problem. The average improvement of each proposed
modified clustering-REF networks over these standard clustering-RBF networks
has been found to be 40.{)3%, 41.64%, 42.32% and 34.99%. Clearly, the results
·show that the performance of the RBF networks using the proposed modified
clustering algorithms has been improved and this proved the proposed model
features a more satisfactory prediction performance and yet effective in
forecasting purposes.
Description
Keywords
Clustering algorithms , Function networks