Back to search

University of West Attica

Ανάπτυξη μεθόδων υπολογιστικής νοημοσύνης για την εκπαίδευση νευρωνικών δικτύων ακτινικής συνάρτησης βάσης με έμφαση σε προσαρμοστικές τεχνικές

Abstract

dc:description

The purpose of this PhD thesis is to develop new computational intelligence methods, focusing specifically on the development of new training algorithms for radial basis function (RBF) networks. The primary goal of the research is to address the challenges associated with modeling real-world systems, which present complex computational and statistical difficulties. By employing both conventional and non-conventional methods and incorporating new adaptive training algorithms, this research seeks to further enhance the existing body of knowledge around RBF network training with the goal of modeling complex systems. To begin with, three new RBF network training algorithms were developed in the context of modeling time-invariant systems, offering distinct features and advantages. The first methodology is based on the use of conventional optimization techniques, in particular the Levenberg-Marquardt (LM) algorithm combined with the variable projection (VP) method, while the fuzzy means (FM) algorithm formes the basis of the method. Extensive experimentation on benchmark data sets demonstrated the ability of the algorithm to achieve high accuracy in short computational time compared to known models and methods of literature. The other two training algorithms are based on metaheuristic optimization methods, namely the tabu search (TS) and differential evolution (DE) in two novel combinatorial frameworks with the non symmetric fuzzy means (NSFM) algorithm. The proposed training schemes are able to handle large datasets in short computational time and take into account many input variables for efficient system modeling. The performance of the algorithms was evaluated on real-world systems, more specifically on the modeling of wind turbine power curves and the modeling of biogas production in an industrial wastewater treatment plant, respectively. The comparative analyses with other existing methods showed the superior performance of the proposed approaches in terms of modelling capabilities, also allowing the identification of potential plant malfunctions. In the field of adaptive algorithms, two new algorithms for modeling data streams were developed through this PhD thesis. The first method integrates the adaptive FM algorithm with Givens rotations, forming a new combinatorial FM-Givens framework. This approach facilitates efficient continuous training on high-volume and high-dimensional data streams, as demonstrated by extensive experimental work on real and artificial data stream sets. This research has spurred subsequent studies in the field of adaptive training methods to overcome some problems implied by the use of standard least squares. In particular, the second algorithm developed involves the incorporation of a modified recursive least squares (RLS) problem. The results highlighted the superiority of the algorithm compared to known methods in the literature, and compared to the FM-Givens algorithm, experimental results showed that the FM-RLS algorithm offers several advantages, including low computational cost and efficiency, making it suitable for large datasets and data streams. Another challenge of this PhD thesis is to overcome the difficulties associated with modelling non-linear time series using RBF networks. Thus, a new hybrid model is proposed that combines the characteristics of an autoregressive moving average (ARMA) model with RBF networks. For the new ARMA-RBF hybrid model, a new two-phase training algorithm is developed that allows automatic determination of the network structure and parameters. The research demonstrates that the proposed approach ensures the convergence of the centers and, moreover, formulates the necessary conditions to achieve convergence of the weights. Extensive evaluations conducted on benchmark time series datasets show that the proposed model outperforms its competitors in terms of modeling accuracy, while achieving faster computational times compared to known methods in the literature. The findings of this research contribute to the advancement of nonlinear time series modeling techniques using RBF networks, providing a promising approach to address real-world prediction problems. Finally, this PhD thesis also focuses on the development of new reinforcement learning (RL) methods based on fully adaptive RBF networks with the primary goals of improving accuracy and reducing complexity. The evaluation of the new reinforcement learning techniques is carried out using two reference control systems, namely a simple pendulum and a rotating inverted pendulum (RIP) system. These are two complex and nonlinear systems with scalar difficulty for which traditional control methods face challenges in the absence of accurate models. The goal in both systems is to guide the pendulum to the upper vertical position through full control by a novel reinforcement learning algorithm based on adaptive RBF networks. The broad findings of this research highlight the effectiveness of the proposed algorithm in challenging control systems.

Degree

thesis:*
Grantor dc:publisher
University of West Attica
Year dc:date
2023

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Karamichailidou, Despoina
  • Καραμιχαηλίδου, Δέσποινα

Subjects

dc:subject × 20

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/54933

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Karamichailidou, Despoina; Καραμιχαηλίδου, Δέσποινα. Ανάπτυξη μεθόδων υπολογιστικής νοημοσύνης για την εκπαίδευση νευρωνικών δικτύων ακτινικής συνάρτησης βάσης με έμφαση σε προσαρμοστικές τεχνικές. University of West Attica, 2023. http://hdl.handle.net/10442/hedi/54933