University of Patras
Μοντελοποίηση μη γραμμικών δυναμικών συστημάτων με νευρωνικά δίκτυα και συνήθεις διαφορικές εξισώσεις: εκτίμηση παραμέτρων και εφαρμογές σε συστήματα Lotka-Volterra
Abstract
dc:descriptionThis thesis develops and evaluates a hybrid data-driven dynamical framework for modeling and forecasting the evolution of competitive systems observed through three interacting time series. The methodological core combines (i) neural networks for learning smooth trajectories from historical data, (ii) identification of a three-dimensional generalized Lotka--Volterra (LV) model through estimation of its interaction parameters from neural-network-implied derivatives, and (iii) out-of-sample prediction by numerically integrating the calibrated LV system with a classical fourth-order Runge--Kutta (RK4) scheme. The approach is designed to balance predictive performance with interpretability: the LV coefficients provide a compact and meaningful description of self-effects and cross-effects among the three components, while the neural network stage offers a flexible approximation that stabilizes estimation in the presence of noise and mild nonlinearities. Within the framework of this thesis, four studies were developed, each addressing a distinct aspect of the proposed hybrid dynamical modeling approach. The first study provides the foundational background by examining the use of feed-forward neural networks for solving ordinary differential equations, establishing key tools for approximating solutions and derivatives. Building on this foundation, three application-focused studies formulate and test the unified NN→LV→RK4 pipeline on real-world competitive settings, including (i) telecommunications market-share dynamics, (ii) cryptocurrency dominance/market-share dynamics, and (iii) the competition of fossil-fuel shares in the energy mix. In each case, the neural network is trained on the earliest 70% of the sample (chronological split), the LV interaction parameters are estimated from the learned trajectories, and the resulting dynamical system is propagated forward over the remaining 30% to produce genuine holdout forecasts. Empirical results across the applications show that the hybrid model can achieve accurate outof-sample predictions while simultaneously yielding interpretable interaction structures that are consistent with domain intuition (e.g., substitution, reinforcement, and asymmetric competition). Beyond forecasting, the framework supports diagnostic analysis through parameter interpretation, stability considerations, and cross-case comparisons. The thesis concludes by synthesizing methodological lessons across domains, discussing limitations such as regime shifts and unmodeled exogenous drivers, and outlining extensions including time-varying parameters, uncertainty quantification, and richer hybrid dynamical specifications.
Degree
thesis:*- Grantor dc:publisher
- University of Patras
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kastoris, Dimitrios
- Καστώρης, Δημήτριος
Subjects
dc:subject × 12Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62450