Cover Image

Improved Genetic Algorithm Multilayer Perceptron Network For Data Classification

Secara umumnya, algoritma genetik (GA) konvensional mempunyai beberapa kelemahan seperti penumpuan pramatang, kecenderungan terperangkap pada penyelesaian optima setempat dan ketidakupayaan penalaan di sekitar kawasan berpotensi. Oleh itu, GA ditambahbaik dengan strategi pencarian, penghasilan s...

Full description

Saved in:
Bibliographic Details
Main Author: Ahmad, Fadzil
Format: Thesis
Language:English
Published: 2017
Subjects:
T Technology
TK1-9971 Electrical engineering. Electronics. Nuclear engineering
Online Access:http://eprints.usm.my/47400/1/Improved%20Genetic%20Algorithm%20Multilayer%20Perceptron%20Network%20For%20Data%20Classification.pdf
http://eprints.usm.my/47400/
Tags: Add Tag
No Tags, Be the first to tag this record!
id my.usm.eprints.47400
record_format eprints
institution Universiti Sains Malaysia
building Hamzah Sendut Library
collection Institutional Repository
continent Asia
country Malaysia
content_provider Universiti Sains Malaysia
content_source USM Institutional Repository
url_provider http://eprints.usm.my/
language English
topic T Technology
TK1-9971 Electrical engineering. Electronics. Nuclear engineering
spellingShingle T Technology
TK1-9971 Electrical engineering. Electronics. Nuclear engineering
Ahmad, Fadzil
Improved Genetic Algorithm Multilayer Perceptron Network For Data Classification
description Secara umumnya, algoritma genetik (GA) konvensional mempunyai beberapa kelemahan seperti penumpuan pramatang, kecenderungan terperangkap pada penyelesaian optima setempat dan ketidakupayaan penalaan di sekitar kawasan berpotensi. Oleh itu, GA ditambahbaik dengan strategi pencarian, penghasilan semula dan elitisma baharu dicadangkan dalam kajian ini. Pernambahbaikan pertama melibatkan perubahan kepada struktur operasi GA yang mana ia menumpu pencarian di sekitar kawasan berpotensi tinggi. Kedua, teknik baharu penghasilan semula yang dinamakan Segmented Multi-Chromosome Crossover (SMCC) telah diperkenalkan. Teknik tersebut mengelak kemusnahan maklumat hampir optima yang terkandung dalam segmen genetik dan membolehkan generasi baharu mewarisi maklumat penting daripada berbilang induk. Ketiganya, tiga jenis variasi elitism dinamakan sebagai Best Among Normal and Improved Population (BANI), Best Between Similar Rank (BBSR) dan Equally Contributed (EQ) telah dibangunkan. Ia melibatkan pertandingan di kalangan individu terbaik daripada populasi normal dan ditambahbaik untuk kelangsungan pada generasi selepasnya. GA yang ditambahbaik kemudiannya digunakan untuk mengoptimasi dan merekabentuk rangkaian perseptron berbilang lapisan (MLP) secara automatik bagi penyelesaian masalah pengkelasan corak. Bilangan nod terlindung, nilai pemberat sambungan awalan dan pemilihan ciri MLP yang memainkan peranan penting dalam menentukan prestasi pengkelasan dipilih untuk dioptimasi secara automatik oleh GA ditambahbaik. Prestasi GA ditambahbaik telah dinilai menggunakan fungsi ujian penanda aras yangv rumit serta berbilang mod dan dibandingkan dengan GA piawai. Berdasarkan kekerapan sesuatu algoritma menghasilkan keputusan terbaik terhadap fungsi ujian yang berbeza; ianya telah terbukti bahawa prestasi teknik yang dicadangkan mengatasi GA piawai. BANI, BBSR dan EQ mencatatkan 30, 18 dan 17 kekerapan keputusan terbaik masing-masing berbanding GA piawai yang hanya mencatatkan 3 keputusan terbaik. Manakala, prestasi pengkelasan GA-MLP yang ditambahbaik telah dinilai menggunakan set-set data yang berbeza dari segi saiz ciri kemasukan dan bilangan kelas keluaran. Keputusan menunjukkan keberkesanan algoritma baharu daripada segi peratusan ujian kejituan. Peratus peningkatan keseluruhan sebanyak 0.6%, 0.1% dan 0.3% bagi ujian kejituan dicatatkan oleh BANI, BBSR dan EQ berbanding dengan GA-MLP piawai. ____________________________________________________________ In general, conventional genetic algorithm (GA) has several drawbacks such as premature convergence, high tendency to get trapped in local optima solution and incapable of fine tuning around potential region. Thus, new improved GA that focuses on new search, reproduction and elitism strategy is proposed in this study. The first improvement involves changes in the operational structure of GA in which it concentrates the search in highly potential area in the search region. Secondly, a novel reproduction technique called Segmented Multi-Chromosome Crossover (SMCC) is introduced. The proposed technique avoids the destruction of nearly optimal information contained in the gene segment and allows offspring to inherit highly important information among multiple parents. Thirdly, three new variations of elitism scheme namely Best Among Normal and Improved Population (BANI), Best Between Similar Rank (BBSR) and Equally Contributed (EQ) are developed. It involves competition among best individuals from normal and improved population to ensure survival in the next generation. The improved GA is then applied for optimization and automatic design of multilayer perceptron (MLP) neural network in solving pattern classification problem. Hidden node size, initial weights and feature selection of the MLP that play significant role in the classification performance are selected to be automatically optimized by the improved GA. The performance of improved GA has been evaluated using highly complicated and multimodal benchmark test functions and compared with the standard GA. Based on the occurrences of the best result obtained by an algorithm across different test functions; it is proven that the proposed method outperforms standard GA. BANI, BBSR and EQ scores 30, 18 and 17 occurrences respectively compared to the standard GA that only scores 3 occurrences. Meanwhile, the improved GA-MLP classification performance has been evaluated using datasets that vary in input features and output sizes. The results demonstrate the effectiveness of the new algorithms in term of test accuracy percentage. There is an overall improvement of 0.6%, 0.1% and 0.3% in test accuracy of BANI, BBSR and EQ compared to the standard GA-MLP.
format Thesis
author Ahmad, Fadzil
author_facet Ahmad, Fadzil
author_sort Ahmad, Fadzil
title Improved Genetic Algorithm Multilayer Perceptron Network For Data Classification
title_short Improved Genetic Algorithm Multilayer Perceptron Network For Data Classification
title_full Improved Genetic Algorithm Multilayer Perceptron Network For Data Classification
title_fullStr Improved Genetic Algorithm Multilayer Perceptron Network For Data Classification
title_full_unstemmed Improved Genetic Algorithm Multilayer Perceptron Network For Data Classification
title_sort improved genetic algorithm multilayer perceptron network for data classification
publishDate 2017
url http://eprints.usm.my/47400/1/Improved%20Genetic%20Algorithm%20Multilayer%20Perceptron%20Network%20For%20Data%20Classification.pdf
http://eprints.usm.my/47400/
_version_ 1717094490212139008
spelling my.usm.eprints.47400 http://eprints.usm.my/47400/ Improved Genetic Algorithm Multilayer Perceptron Network For Data Classification Ahmad, Fadzil T Technology TK1-9971 Electrical engineering. Electronics. Nuclear engineering Secara umumnya, algoritma genetik (GA) konvensional mempunyai beberapa kelemahan seperti penumpuan pramatang, kecenderungan terperangkap pada penyelesaian optima setempat dan ketidakupayaan penalaan di sekitar kawasan berpotensi. Oleh itu, GA ditambahbaik dengan strategi pencarian, penghasilan semula dan elitisma baharu dicadangkan dalam kajian ini. Pernambahbaikan pertama melibatkan perubahan kepada struktur operasi GA yang mana ia menumpu pencarian di sekitar kawasan berpotensi tinggi. Kedua, teknik baharu penghasilan semula yang dinamakan Segmented Multi-Chromosome Crossover (SMCC) telah diperkenalkan. Teknik tersebut mengelak kemusnahan maklumat hampir optima yang terkandung dalam segmen genetik dan membolehkan generasi baharu mewarisi maklumat penting daripada berbilang induk. Ketiganya, tiga jenis variasi elitism dinamakan sebagai Best Among Normal and Improved Population (BANI), Best Between Similar Rank (BBSR) dan Equally Contributed (EQ) telah dibangunkan. Ia melibatkan pertandingan di kalangan individu terbaik daripada populasi normal dan ditambahbaik untuk kelangsungan pada generasi selepasnya. GA yang ditambahbaik kemudiannya digunakan untuk mengoptimasi dan merekabentuk rangkaian perseptron berbilang lapisan (MLP) secara automatik bagi penyelesaian masalah pengkelasan corak. Bilangan nod terlindung, nilai pemberat sambungan awalan dan pemilihan ciri MLP yang memainkan peranan penting dalam menentukan prestasi pengkelasan dipilih untuk dioptimasi secara automatik oleh GA ditambahbaik. Prestasi GA ditambahbaik telah dinilai menggunakan fungsi ujian penanda aras yangv rumit serta berbilang mod dan dibandingkan dengan GA piawai. Berdasarkan kekerapan sesuatu algoritma menghasilkan keputusan terbaik terhadap fungsi ujian yang berbeza; ianya telah terbukti bahawa prestasi teknik yang dicadangkan mengatasi GA piawai. BANI, BBSR dan EQ mencatatkan 30, 18 dan 17 kekerapan keputusan terbaik masing-masing berbanding GA piawai yang hanya mencatatkan 3 keputusan terbaik. Manakala, prestasi pengkelasan GA-MLP yang ditambahbaik telah dinilai menggunakan set-set data yang berbeza dari segi saiz ciri kemasukan dan bilangan kelas keluaran. Keputusan menunjukkan keberkesanan algoritma baharu daripada segi peratusan ujian kejituan. Peratus peningkatan keseluruhan sebanyak 0.6%, 0.1% dan 0.3% bagi ujian kejituan dicatatkan oleh BANI, BBSR dan EQ berbanding dengan GA-MLP piawai. ____________________________________________________________ In general, conventional genetic algorithm (GA) has several drawbacks such as premature convergence, high tendency to get trapped in local optima solution and incapable of fine tuning around potential region. Thus, new improved GA that focuses on new search, reproduction and elitism strategy is proposed in this study. The first improvement involves changes in the operational structure of GA in which it concentrates the search in highly potential area in the search region. Secondly, a novel reproduction technique called Segmented Multi-Chromosome Crossover (SMCC) is introduced. The proposed technique avoids the destruction of nearly optimal information contained in the gene segment and allows offspring to inherit highly important information among multiple parents. Thirdly, three new variations of elitism scheme namely Best Among Normal and Improved Population (BANI), Best Between Similar Rank (BBSR) and Equally Contributed (EQ) are developed. It involves competition among best individuals from normal and improved population to ensure survival in the next generation. The improved GA is then applied for optimization and automatic design of multilayer perceptron (MLP) neural network in solving pattern classification problem. Hidden node size, initial weights and feature selection of the MLP that play significant role in the classification performance are selected to be automatically optimized by the improved GA. The performance of improved GA has been evaluated using highly complicated and multimodal benchmark test functions and compared with the standard GA. Based on the occurrences of the best result obtained by an algorithm across different test functions; it is proven that the proposed method outperforms standard GA. BANI, BBSR and EQ scores 30, 18 and 17 occurrences respectively compared to the standard GA that only scores 3 occurrences. Meanwhile, the improved GA-MLP classification performance has been evaluated using datasets that vary in input features and output sizes. The results demonstrate the effectiveness of the new algorithms in term of test accuracy percentage. There is an overall improvement of 0.6%, 0.1% and 0.3% in test accuracy of BANI, BBSR and EQ compared to the standard GA-MLP. 2017-03-01 Thesis NonPeerReviewed application/pdf en http://eprints.usm.my/47400/1/Improved%20Genetic%20Algorithm%20Multilayer%20Perceptron%20Network%20For%20Data%20Classification.pdf Ahmad, Fadzil (2017) Improved Genetic Algorithm Multilayer Perceptron Network For Data Classification. PhD thesis, Universiti Sains Malaysia.
score 13.211869

Similar Items