Authors: Mohamed H. Gadallah, Khaled Abdel Hamid El-Sayed, Keith Hekman
Addresses: Department of Operations Research, Institute of Statistical Studies and Research, Cairo University, 5 Tharwat Street, Orman, Dokki, Giza, 12613, Egypt. ' Department of Mechanical Engineering, The American University, P.O. Box 2511, Cairo, Egypt. ' Department of Mechanical Engineering, The American University, P.O. Box 2511, Cairo, Egypt
Abstract: Engineering systems are difficult to model and expensive to experiment with. A supervised neural network (NN) using radial basis network (RBN) is developed. The RBN uses error back-propagation algorithm (EBP) as predictive tools for the modelling process. Statistical techniques and orthogonal arrays (OAs) have been employed to offset this expense. A comparison between several experimental based models on predictive capability and number of training patterns is given. Sometimes information is not available, and the modeller can compromise accuracy information for the experimental cost. Several two-level, three-level, four-level, and five-level OAs are used. These are L8 OA, L9 OA, L27 OA, L32 OA, and L25 OA respectively. Results show that each individual model has a potential for approximation. Besides an attempt to combine the models in a sequence and the resulting composed models are compared for approximation. Results indicate that a certain sequence leads to a better model with faster convergence and less predictive error.
Keywords: radial basis networks; RBN; neural networks; process modelling; orthogonal arrays.
International Journal of Machining and Machinability of Materials, 2010 Vol.8 No.1/2, pp.177 - 197
Published online: 05 Aug 2010 *Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article