Heterogeneous model of neural networks using the learning algorithms based on dynamic modification of network topology

L_AvdiyenkoLiliya Avdiyenko graduated from Kharkiv National University of Radio Electronics, Ukraine and has recently passed her MPhil viva at WIT with a thesis on “Heterogeneous model of neural networks using the learning algorithms based on dynamic modification of network topology”. Her external examiner was Prof Leslie Smith from the University of Stirling, Scotland, and the internal was Prof Alex Galybin.

The thesis presents research into heterogeneous adaptive neural networks for problems of handling non-stationary processes. The aim of the research was the investigation of a hybrid neural network model consisting of the adaptive radial basis functions networks (RBFNs) which use growing and pruning procedures for determining an optimal network topology.

Most artificial neural networks are homogeneous systems because they search for a solution using one-type elements in the entire feature space. In such cases ensembles of the homogeneous systems, which are called heterogeneous or hybrid neural networks, are often used for enhancement and stabilization of the results.

Along with an advantage of the neural networks to adapt weights of their synaptic connections during learning they have a significant drawback. There is no formalized method to determine an optimal network topology which has a great influence on a network performance. So-called adaptive or ontogenic neural networks try to solve this problem. During learning they adapt both weights of synaptic connections and network structure. There are two basic classes of ontogenic algorithms such as growing and pruning algorithms.

The thesis contains an analysis of different hybrid neural networks and considers various adaptive algorithms depending on learning types and methods of generating an optimal network topology. On the basis of this analysis a hybrid neural network model was implemented which consisted of growing and pruning a radial basis function network for handling processes with characteristics that are changeable in time.

As a result of her research both examiners recommended that Liliya be awarded the degree of Master of Philosophy.

Grateful acknowledgement is given to the Foreign and Commonwealth Office for their support.