University of the Aegean
Ανάπτυξη πολυωνυμικών νευρωνικών δικτύων με χρήση ορθογωνίων πολυωνύμων
Abstract
dc:descriptionThis PhD thesis presents the research involving the creation, training and use of Polynomial Artificial Neural Networks (pNN's) using orthogonal polynomials. This research focuses on the effect of these polynomials into the structure of Artificial Neural Networks.Three original structures of Polynomial Neural Networks have been developed and they have been published in peer-reviewed journals and in proceedings of scientific conferences. The purpose of all three methods is their usage on high non-linear data sets in order to increase the accuracy of their results. The datasets involve: (a) experimental data collected from the research programs "THALES (ISLA) Physical and Economic Impacts of Sea Level Changes on the beaches of the Aegean Archipelago" and "Beachtour: Synergy for sustainable development and safe use of the Greek beaches" (see Appendix B), and ( b) data selected by the UCI Repository database. The first proposed network (Chapter 2) uses the orthogonal Chebychev polynomials into the structure of an ANN. The network consists of the following three parts: (i) A polynomial neural network which uses polynomials Chebychev as activation functions. (ii) A part consisting of an ANN which uses radial basis functions (RBF) and uses an original clustering method of the input-output data space. This method clusters input data according to the distribution of the output data space and thus provides better approximation capabilities to the ANN. And (iii) The inference part which combines the two previous parts of the network in order to calculate the network's output. The network was trained using the Gradient Descent method with the Armijo rule, in order to minimize the square error of the network. Experiments (numerical simulations) of the proposed ANN were applied on UCI Repository datasets and on a high non-linear data set which defines the position of a shoreline from digital images that show the standard deviation of a set of consecutive images.The second method (Chapter 3) presents a neuro-fuzzy network which consists of (i) a fuzzy system (ii) a polynomial ANN that uses the orthogonal Hermite polynomials and (iii) a phase that uses the previous two parts in order to extract the desired output/result. The network was trained using the Gradient Descent method with the Armijo rule, in order to minimize the square error of the network. The network was tested on the UCI Repository data sets and on a non-linear data set that models/predicts the coastal erosion through data generated from spatial and meteorological measurements. The third method (Chapter 4) consists of a polynomial ANN using the Legendre polynomials. In order to maintain the orthogonality of the polynomials which operates into the interval, a scaling procedure was applied before the data enter the network's layer that uses the polynomials. The proposed ANN was trained using a swarm intelligence algorithm, the GABC algorithm which is a modification of the Artificial Bee Colony algorithm (ABC) that uses the G-best position of the method Particle Swarm Optimization (PSO). Finally the network was tested on the UCI Repository data sets and on a data set that predicts the beach rotation phenomenon. In all the above methods the results showed that expressing the activation functions of the nodes of a neural network using orthogonal polynomials are capable of improving the network performance in regression problems. .
Degree
thesis:*- Grantor dc:publisher
- University of the Aegean
- Year dc:date
- 2017
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Rigos, Anastasios
- Ρήγος, Αναστάσιος
Subjects
dc:subject × 8Rights
dc:rights- Statement dc:rights
-
- BY_NC_ND
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/40256