A Voting Technique Of Multilayer Perceptron Ensemble For Classification Application
Loading...
Date
2014-02
Authors
Talib, Hafizah
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
MLP is a model of artificial neural network, which is simple yet successfully
applied in various applications. The instability of MLP performance where small
changes in training parameter could produce different models that inhibiting
attainment of high accuracy in classification applications. In this research, an
integrated system of Multi-Layer Perceptron Ensemble (MLPE) consisting of an
MLPE and a new voting algorithm has been developed to increase classification
accuracy and reduce the number of reject class cases. MLPE is produced from
singular MLPs that are diverse in term of training algorithm and their initial weights.
Three training algorithms used are Levenberg-Marquardt (LM), Resilient
Backpropagation (RP) and Bayesian Regularization (BR). In order to choose the
final output of MLPE, a new voting algorithm named Trust-Sum Voting (TSV) is
proposed. The effectiveness of MLPE with TSV (MLPE-TSV) has been tested on
four classification case studies which are Electrical Capacitance Tomography (ECT),
Landsat Satellite Image (LSI), German Credit (GC) and Pima Indian Diabetes (PID).
The performance of MLPE-TSV has been compared with the performance of MLPE
which employs existing voting algorithms which are Majority Voting (MLPE-MV)
and Trust Voting (MLPE-TV). The obtained results have shown that the proposed
MLPE-TSV is capable of increasing the accuracy of classification as compared to
singular MLPs, MLPE-MV and MLPE-TV. MLPE-TSV has also managed to reduce
the number of cases in reject class.
Description
Keywords
Voting Technique , Multilayer Perceptron Ensemble