International Hellenic University
Development of intelligent computational methods for solving time-varying problems
Abstract
dc:descriptionThis PhD thesis presents the development of novel recurrent neural network (RNN) architectures, introducing new activation functions and discrete-time formulations to address time-varying problems in numerical linear algebra. While nonlinear activation functions have been shown to enhance the convergence speed of Zeroing Neural Network (ZNN) models, existing ZNN frameworks typically impose strict constraints—requiring activation functions to be both strictly odd and monotonically increasing. These limitations reduce the flexibility and expressive capacity of the models, underscoring the need for innovation in this area. In response, this thesis proposes enhanced models that significantly improve the accuracy and convergence performance of traditional approaches. The thesis is structured into four chapters. Chapter One introduces foundational concepts from mathematical analysis and matrix theory that support the structure of RNN models. It also reviews key definitions related to RNNs, common recurrent architectures, and widely used activation functions. Essential theorems from dynamical systems theory and core programming techniques employed throughout the research are also discussed. The chapter concludes with a survey of real-world applications where RNNs have proven effective in solving practical problems. Chapter Two extends the Zhang Neural Network (ZNN) framework to handle cases where the time derivative of the Zhangian function does not exist at certain points. To address such scenarios—where the indefinite error function E(t), whether matrix-, vector-, or scalar-valued, contains non-differentiable elements—two novel approaches are proposed. In addition, a new family of activation functions, specifically designed for integration within the ZNN framework, is introduced. These are used to construct advanced models for solving time-varying matrix equations, including the Stein, Sylvester, and Generalized Sylvester equations. Compared to earlier nonlinear RNN models with conventional activation functions, the proposed models exhibit significantly improved convergence rates. Simulink-based simulations validate the effectiveness of these models, demonstrating both faster convergence and improved solution accuracy. In Chapter Three, the thesis explores the application of discretization techniques to the ZNN framework, resulting in the formulation of novel discrete-time RNN models. These models are employed to solve a range of mathematical problems, such as matrix inversion and nonlinear optimization. Additionally, the discrete-time models are applied to robotics tasks, including robotic manipulator control and tracking control for the Stewart platform. Experimental results show that these models outperform existing discrete-time methods in terms of accuracy. Finally, Appendices A and B include Simulink diagrams of the baseline RNN models and activation functions used in comparative analyses, alongside those of the novel formulations proposed in this work.
Degree
thesis:*- Grantor dc:publisher
- International Hellenic University
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Gerontitis, Dimitrios
- Γεροντίτης, Δημήτριος
Subjects
dc:subject × 16- Νευρωνικά δίκτυα
- Χρονικά μεταβαλλόμενα συστήματα
- Συναρτήσεις ενεργοποίησης
- Αριθμητική Γραμμική Άλγεβρα
- Neural networks
- Time-varying systems
- Activation functions
- Numerical linear algebra
- Μαθηματικά
- Φυσικές Επιστήμες
- Υπολογιστικά μαθηματικά
- Εφαρμοσμένα μαθηματικά
- Mathematics
- Natural Sciences
- Computational Mathematics
- Applied Mathematics
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/59856