Aristotle University Of Thessaloniki (AUTH)
Μέθοδοι σταδιακής μηχανικής μάθησης σε χρονικώς μεταβαλλόμενα προβλήματα: εφαρμογές αναγνώρισης προτύπων, χρονοσειρών και συστημάτων για λήψη αποφάσεων σε πραγματικό χρόνο
Abstract
dc:descriptionThe thesis proposes pattern recognition techniques based on machine learning methods and automated data mining; the application domain is systems that evolve over time. It describes algorithms and resolution strategies that combine and exploit features from established methodologies developed in the framework of computational and statistical learning. The attributes of the proposed identification and recognition techniques are: 1) low time-space complexity, 2) incremental and fast learning, 3) real-time response, 4) application independence, 5) simplicity, 6) automated operation, 7) scalability, and 8) distributed computing. After the application of these techniques, the produced solutions are not proved to be the optimal ones; nevertheless they are sufficient and functional. The empirical verification is implemented in real-world problems of different complexity. These problems can be represented by time evolving systems and emerge from the fields of environmental parameters forecasting, voice/speech and image recognition, and video-based content extraction such as human pose recovery and human action recognition. It is known that time evolution occurs in a variety of problems encountered in physics, chemistry, biology, economy and more; thus, the proposed recognition schemes present a wide range of applications. Additionally, the main trait of incremental learning algorithms is the integration of new information into the knowledge base without the need for the model’s retraining. This attribute is applicable to domains where the data flow is continuous such as autonomous robots learning, information retrieval on the Web, human-machine interaction, medical monitoring devices, error diagnosis, security systems, predictive models for natural phenomena and more. The difficulty in defining the analytical form of functions that describe the aforementioned problems, the high dimensionality of data sets, the intervals of intense discontinuity and the presence of high noise make impossible the application of the differential calculus and classical linear or nonlinear analysis. This research work demonstrates experimentally that the adoption of techniques which combine the regression with continuous clustering and classification, allows the handling of non- well defined problems, especially in those cases where the conditions that lead to the convergence to the optimal solution cannot be verified in real world.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2010
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Syrris, Vasileios
- Συρρής, Βασίλης
Subjects
dc:subject × 16- Μηχανική μάθηση
- Στατιστική μάθηση
- Εξόρυξη δεδομένων
- Τεχνητά νευρωνικά δίκτυα
- Θεωρία δικτυωμάτων
- Μηχανές πυρήνα
- Machine learning
- Statistical learning
- Data mining
- Artificial neural networks
- Lattice theory
- Kernel machines
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/24658