Hacia un diseño óptimo de la arquitectura Multilayer Feedforward
- Fernández Redondo, Mercedes
- Carlos Hernández Espinosa Director
Defence university: Universitat Jaume I
Fecha de defensa: 21 February 2008
- Pedro Gómez Vilda Chair
- Ángel Pascual del Pobil Ferré Secretary
- Francisco Javier López Aligué Committee member
- Alberto Prieto Espinosa Committee member
- Manuel Graña Romay Committee member
Type: Thesis
Abstract
The objective of this Doctoral Thesis was to carry a comparative study on several existent methods in order to solve different aspects of the design of Multilayer Feedforward architecture, in neural networks classification problems. The aspects of design studied were: handling unknown input information, input selection, selection of the number of hidden units, influence in the generalization capability of the number of hidden layers and weight initialization. For each one of these aspects, we carried out a comparative study of several existent methods in order to solve the problem. We recommend the use of the best methods in order to develop a concrete application.