Novel Art-Based Neural Network Models For Pattern Classification, Rule Extraction And Data Regression
Loading...
Date
2010-05
Authors
Yap, Keem Siah
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
This thesis is concerned with the development of novel neural network models for
tackling pattern classification, rule extraction, and data regression problems. The
research focuses on one of the advanced features of neural networks, i.e., the
incremental learning ability. This ability relates to continuous learning of new
knowledge without disturbing the existing knowledge base and without re-iterating
through the training samples. The Adaptive Resonance Theory (ART) and Generalized
Regression Neural Network (GRNN) models are employed as the backbone in this
research. The resulting hybrid neural network model (denoted as GART) is capable of
handling pattern classification and data regression problems. The capability of GART is
further enhanced (denoted as EGART) with a number of features, which include the
used of Laplacian loss and likelihood functions, a new definition of vigilance function, a
match tracking mechanism. In addition, a pre-processing technique, i.e., the ordering
algorithm, for determining the presentation sequence of training samples is applied
(denoted as O-EGART). The O-EGART model is equipped with a series of postprocessing
procedures (denoted as O-EGART-PR), i.e., network pruning and rule
extraction. Network pruning requires computation of the confidence factor of each
protoptye node in O-EGART-PR based on a set of validation samples. A quantization
process is also applied to convert the prototype weights into a set of IF-THEN rules. In
addition, a standard Fuzzy Inference System (FIS) is constructed (denoted as OEGART-
PR-FIS) in order to evaluate the quality of the extracted rules. The
performances of the proposed ART-based models are compared with those from other
approaches using benchmark data sets, and the bootstrap method is used to quantify the
results. To evaluate the practical applicability of the proposed ART-based models,
empirical experiments based on seven benchmark and real-world data sets, i.e., three
from power systems, three from fire safety engineering, and one from medical
application, are conducted. These results show good performances, e.g., accuracy rates
are 98.92% and 97.20% for classification of harmonic currents in distribution network
and diagnosis of circulating water systems in power generation plant, respectively, hence
justified the usefulness of the proposed ART-based models in undertaking pattern
classification and data regression problems.
Description
Keywords
Novel neural network models , for tackling pattern classification