University of the Aegean
Ανάπτυξη μεθόδων ασαφούς συσταδοποίησης για τη μοντελοποίηση νευρωνικών δικτύων συναρτήσεων ακτινικής βάσης
Abstract
dc:descriptionIn the present doctoral thesis takes place inquiring effort in the training of Neural Networks and concretely the Radial Basis Function Neural Networks – RBFNN. For the training of radial basis function neural networks were used algorithms of fuzzy clustering. Five new methods for the training of radial basis function neural networks were developed and have been published in international scientific conferences and journals. The contribution of this doctoral dissertation is that systematic research was conducted with a view to exploring the use of fuzzy clustering in training of radial basis functions neural networks. Within the framework of the research we focused on the analysis of advantages and disadvantages and the effective impact of fuzzy clustering in the process of training of this type of neural networks.The result of this research is to develop in the first level of three simple methods for clustering fuzzy neural networks for the training of radial basis functions. Concretely in the first method, was used the very known method of fuzzy c-means (FCM) for the preprocessing of the training data and then the method of weighted fuzzy c-means (FCM) for the calculation of the core parameters of the neural network hidden nodes. In the second method, was used the optimal clustering for the determination of the number of cores of radial basis functions as well as the parameters of neural network. Finally, we proposed a new method that combines input-output fuzzy clustering and optimal fuzzy clustering for the efficient design of radial basis function neural networks.In second level, were developed two innovative methods for the training of radial basis function neural networks. Concretely in first method is proposed a new innovator hybrid method of clustering, which combines in a uniform fashion crisp and fuzzy clustering for the calculation of the neural network core parameters. The second proposed method uses information granulation for the calculation of the core centers and a new metric distance for the calculation of widths of the radial basis functions. In all the proposed methods became practical application in real models and functions approximation, the results of which showed that the proposed methods produce reliable, precise and compact radial basis function neural networks. The comparison with the international bibliography showed that RBF networks that was created have better results with regard to the mean square error – MSE as well as the normalized mean square error – NRMSE so much for training data as long as for the test data.
Degree
thesis:*- Grantor dc:publisher
- University of the Aegean
- Year dc:date
- 2011
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Niros, Antonios
- Νείρος, Αντώνιος
Subjects
dc:subject × 14- Νευρωνικά δίκτυα συναρτήσεων ακτινικής βάσης
- Διακριτή συσταδοποίηση
- Ασαφής συσταδοποίηση
- Βέλτιστη ασαφής συσταδοποίηση
- Κοκκοποίηση πληροφορίας
- Radial basis function neural networks
- Crisp clustering
- Fuzzy clustering
- Optimal fuzzy clustering
- Information granulation
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Natural Sciences
- Computer and Information Sciences
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/26067