2018_Batch Back Propagation Algorithm Using Dynamic Parameters And Heuristic Approach

Bibliographic Details
Format: General Document
_version_ 1860798148961632256
building INTELEK Repository
collection Online Access
collectionurl https://intelek.unisza.edu.my/intelek/pages/search.php?search=!collection3
copyright Copyright©PWB2025
country Malaysia
date 2018-09-06
format General Document
id 16183
institution UniSZA
originalfilename BATCH BACK PROPAGATION ALGORITHM USING DYNAMIC PARAMETERS AND HEURISTIC APPROACH.pdf
person Mohammed Sarhan Ghalib Al Duais
recordtype oai_dc
resourceurl https://intelek.unisza.edu.my/intelek/pages/view.php?ref=16183
sourcemedia Server storage
Scanned document
spelling 16183 https://intelek.unisza.edu.my/intelek/pages/view.php?ref=16183 https://intelek.unisza.edu.my/intelek/pages/search.php?search=!collection3 General Document Malaysia Library Staff (Top Management) Library Staff (Management) Library Staff (Support) Terengganu Faculty of Informatics & Computing English application/pdf 1.5 Server storage Scanned document Universiti Sultan Zainal Abidin UniSZA Private Access UNIVERSITI SULTAN ZAINAL ABIDIN SAMBox 2.4.24; modified using iTextSharp™ 5.5.10 ©2000-2016 iText Group NV (AGPL-version) 279 Copyright©PWB2025 2018-09-06 BATCH BACK PROPAGATION ALGORITHM USING DYNAMIC PARAMETERS AND HEURISTIC APPROACH.pdf Mohammed Sarhan Ghalib Al Duais Backpropagation (Artificial intelligence) 2018_Batch Back Propagation Algorithm Using Dynamic Parameters And Heuristic Approach The Batch back-propagation (BBP) algorithm is widely considered one of the most efflcient training algorithms for multi-layer perception. However, in the BBP algorithm, the weight updates after each cycle and slow training because ofthe inherent local minimum, requiring the manual setting of several pararneters. Training the BBP algorithm for adjusted weight is largely influenced by factors such as leaming rate and momentum. To achieve the objectives of this study, two techniques were performed to escape the local minimum through several dynamic leaming rates created with boundaries. To avoid extreme cumulative weight updates during training, a dynamic leaming rate and momentum factor were integrated as an implicit function with boundaries. These techniques were implemented using different structures of ANNs, with the sigmoid function used as the activation function. Several datasets obtained {iom the UCI Standard Machine Leaming Repository were used to benchmark the effrciency of the improved algorithm. A total of 853 experiments were performed in Matlab with the goal error set at 104. The effectiveness of the improved aigorithm was tested using several criteria, including average training time, number oftraining epochs or iterations, standard deviation (SD), and training accuracy. This study found that the improved algorithm, dynamic learning rate, and dynamic momentum factor with boundaries significantly affected training time and enhanced accuracy rate. The dynamic parameters helped the improved or dynamic algorithms avoid the local minimum and eliminated training saturation. [n the first contribution. the accuracy rates of the structures were 98.702 and 99.1%, and the processing times of the improved algorithm were 3936 and 4755 times faster, respectively, than the BBp algorithm. In the second contribution, the accuracy rate was 99.3%o, and the processing time of the improved algorithm was I 8 times faster than that of the BBP algorithm. In the third contribution, the accuracy rates of the structures were 98.57o and 99.1%, and the processing times of the improved algorithm were 1 8 l 6 and 869 times faster, respectively, than the BBP algorithm. In the fourth contribution, the accuracy rates of the structures were 98.5% and 99.loh, and the processing times of the improved algorithm were 861 and 638 times faster, respectively, than the BBp algorithm. Furthermore, the improved algorithm provided faster training than the existing BBp algorithm. Therefore, the improved or dynamic algorithm provides superior performance compared to the existing studies. Dissertations, Academic Backpropagation Algorithm Batch Learning Artificial Neural Networks Thesis
spellingShingle 2018_Batch Back Propagation Algorithm Using Dynamic Parameters And Heuristic Approach
state Terengganu
subject Backpropagation (Artificial intelligence)
Dissertations, Academic
summary The Batch back-propagation (BBP) algorithm is widely considered one of the most efflcient training algorithms for multi-layer perception. However, in the BBP algorithm, the weight updates after each cycle and slow training because ofthe inherent local minimum, requiring the manual setting of several pararneters. Training the BBP algorithm for adjusted weight is largely influenced by factors such as leaming rate and momentum. To achieve the objectives of this study, two techniques were performed to escape the local minimum through several dynamic leaming rates created with boundaries. To avoid extreme cumulative weight updates during training, a dynamic leaming rate and momentum factor were integrated as an implicit function with boundaries. These techniques were implemented using different structures of ANNs, with the sigmoid function used as the activation function. Several datasets obtained {iom the UCI Standard Machine Leaming Repository were used to benchmark the effrciency of the improved algorithm. A total of 853 experiments were performed in Matlab with the goal error set at 104. The effectiveness of the improved aigorithm was tested using several criteria, including average training time, number oftraining epochs or iterations, standard deviation (SD), and training accuracy. This study found that the improved algorithm, dynamic learning rate, and dynamic momentum factor with boundaries significantly affected training time and enhanced accuracy rate. The dynamic parameters helped the improved or dynamic algorithms avoid the local minimum and eliminated training saturation. [n the first contribution. the accuracy rates of the structures were 98.702 and 99.1%, and the processing times of the improved algorithm were 3936 and 4755 times faster, respectively, than the BBp algorithm. In the second contribution, the accuracy rate was 99.3%o, and the processing time of the improved algorithm was I 8 times faster than that of the BBP algorithm. In the third contribution, the accuracy rates of the structures were 98.57o and 99.1%, and the processing times of the improved algorithm were 1 8 l 6 and 869 times faster, respectively, than the BBP algorithm. In the fourth contribution, the accuracy rates of the structures were 98.5% and 99.loh, and the processing times of the improved algorithm were 861 and 638 times faster, respectively, than the BBp algorithm. Furthermore, the improved algorithm provided faster training than the existing BBp algorithm. Therefore, the improved or dynamic algorithm provides superior performance compared to the existing studies.
title 2018_Batch Back Propagation Algorithm Using Dynamic Parameters And Heuristic Approach
title_full 2018_Batch Back Propagation Algorithm Using Dynamic Parameters And Heuristic Approach
title_fullStr 2018_Batch Back Propagation Algorithm Using Dynamic Parameters And Heuristic Approach
title_full_unstemmed 2018_Batch Back Propagation Algorithm Using Dynamic Parameters And Heuristic Approach
title_short 2018_Batch Back Propagation Algorithm Using Dynamic Parameters And Heuristic Approach
title_sort 2018_batch back propagation algorithm using dynamic parameters and heuristic approach