RANGKAIAN NEURAL PERSEPTRON BERBILANG LAPISAN HffiRID BERKELOMPOK UNTUK PENGELASAN CORAK YANG LEBIH BAlK
Loading...
Date
2009-11
Authors
WAN MAMAT, WAN MOHD FAHMI
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Perseptron based and Radial Basis Function (RBF) neural networks are commonly used
for pattern classification. However, their performances are limited to several weaknesses.
For the Perseptron based neural networks, their training procedures are often trapped at a
local optimum with slow convergence rate and sensitive to initial parameter values.
Whereas, three,' typical problems for the RBF network are dead centers, centers
redundancy arid trapped centers in local minima. Thus, this study introduces a new neural
network architecture called Clustered Hybrid Multilayered Perceptron or ClusteredHMLP.
In this work, the Hybrid Multilayered Perceptron network architecture has been
modified by introducing an additional layer called cluster layer to form the proposed
neural network. The cluster layer concept is adopted from the RBF network architecture.
The proposed Clustered-HMLP is then trained using a new modified training algorithm
called Clustered Modified Recursive Prediction Error or Clustered-MRPE. The
Clustered-MRPE algorithm is a combination of the modified version of the conventional
MRPE and the Moving k-mean clustering algorithms. The capability of the ClusteredHMLP
network is demonstrated using 14 VCI benchmark datasets. The results indicate
that the introduction of pre-classification process at additional cluster layer favors the
proposed Clustered-HMLP network by producing better performance as compared to
other 12 conventional intelligent classifiers. Furthermore, the proposed Clustered-HMLP
network has successfully been applied to develop a Transformer Fault Diagnosis System
and Drug-like and Non Drug-like Classification System.
Description
Keywords
NEURAL PERSEPTRON , LAPISAN HffiRID