National Technical University of Athens (NTUA)
Ανάπτυξη αλγορίθμων υπολογιστικής νοημοσύνης και εφαρμογές στην εξέλιξη του παγκόσμιου ιστού
Abstract
dc:descriptionComputational intelligence (CI) is a branch of artificial intelligence, which is based on imitating elements of intelligence exhibited by living organisms and offers techniques that are capable of solving complex problems that could not be solved by using conventional methodologies. The objective of this doctoral dissertation is the development of novel algorithms belonging to the field of computational intelligence, which are capable of handling numerical, categorical and mixed attributes, while providing improved predictive abilities when compared to conventional methodologies. The computational intelligence techniques which constitute the basic tools used in this dissertation are the radial basis function (RBF) neural networks, the evolutionary simulated annealing (ESA) technique and the particle swarm optimization (PSO) algorithm. As far as numerical attributes are concerned, 3 novel RBF network training algorithms where developed in this dissertation, with the ability to perform full network optimization by concurrently determining all the RBF parameters. The first algorithm, namely PSO-NSFM, is based on the non symmetric fuzzy means (NSFM) algorithm, which is enhanced by a PSO optimization engine, in order to optimize the fuzzy partition of the input space. The PSO-NSFM algorithm produces RBF models with simpler structures, while at the same time exhibiting higher prediction accuracy in shorter computational times, when compared to standard machine learning techniques. The second algorithm which was developed for handling numerical inputs, namely COOP-VW, optimizes the construction of RBF networks trained with the NSFM algorithm, based on a cooperative PSO (CPSO) framework. The COOP-VW algorithm allows for using variable-width basis functions, which increase the flexibility of the produced models. The result is an additional improvement in prediction accuracy, albeit at the cost of a relatively small increase of the computational cost, when compared to the PSO-NSFM technique. The third algorithm which was developed, namely ESA-NSFM, is based on a modification of the ESA technique, in order to optimize the RBF models produced by the NSFM algorithm. The ESA-NSFM algorithm produces RBF network models of increased prediction accuracy compared to standard machine learning techniques, and at the same time offers fast operational parameter customization. Using algorithms similar to those mentioned above, the basic RBF network structure can handle numeric inputs with great success; however, it is unable to deal with categorical, or mixed data. In order to solve this problem, two novel neural network architectures were developed within this doctoral dissertation. The first one, namely CRBF networks, is an expansion of RBF networks, providing them with the ability of handling categorical data. This is accomplished by using categorical tuples as RBF centers, while the categorical center selection (CCS) algorithm is also proposed for training these networks. The second neural network architecture that was developed, namely MRBF networks, is capable of handling mixed data. In fact, MRBF networks are essentially a synthesis of RBF and CRBF networks, and are trained with a novel methodology which combines the fuzzy means algorithm and the CCS technique. Due to their ability to handle in a different manner the nature of numerical and categorical data, MRBF networks can provide solutions with increased prediction accuracy, in contrast to the methodologies that are inherently designed for a single type of data. Moreover, because of the simple and fast training algorithm that they use, MRBF networks are suitable to handle big data. All the aforementioned methods were evaluated on a large number of benchmark datasets, derived from a variety of scientific fields, including function approximation and classification problems. Results and comparisons with a wide range of machine learning methods, including different neural network methods, support vector machines (SVMs), naïve Bayes classifiers, decision trees, random forest trees, etc., highlight the effectiveness of the proposed approaches. In addition, at the level of application, a novel intelligent system was developed with the ability to automatically classify web pages based on their content. The latter is a generic tool for web page classification which offers satisfactory precision and is able to classify a webpage depending on functional, genre, subject classification, etc. This system was called MRBF-WPC and adopts the MRBF neural network architecture and web mining tools. Applying this system to a widely used dataset of webpages and comparing it with other methods, has shown that the system provides increased accuracy in classifying web pages. Thus, the MRBF-WPC system is an integrated tool that can be used as the basis for enrichment of the web content with metadata, and thus, for the World Wide Web growth into the Semantic Web.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2017
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Chondrodima, Evangelia
- Χονδροδήμα, Ευαγγελία
Subjects
dc:subject × 32- Νευρωνικά δίκτυα
- Νευρωνικά δίκτυα ακτινικής συνάρτησης βάσης
- Τεχνητή νοημοσύνη
- Υπολογιστική νοημοσύνη
- Βελτιστοποίηση σμήνους σωματιδίων
- Μηχανική μάθηση
- Αλγόριθμοι εκπαίδευσης
- Αλγόριθμος προσομοιωμένης ανόπτησης
- Αλγόριθμοι βελτιστοποίησης
- Παγκόσμιος ιστός
- Κατηγορικά δεδομένα
- Μικτά δεδομένα
- Neural networks
- Radial basis function networks
- Artificial intelligence
- Computational intelligence
- Particle swarm optimization
- Machine learning
- Training algorithms
- Simulated annealing algorithm
- Optimization algorithms
- Word Wide Web
- Categorical data
- Mixed data
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Natural Sciences
- Computer and Information Sciences
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/41612