Time Series Prediction Using Evolving Polynomial Neural Networks

i DECLARATION No portion of the work referred to in this dissertation has been submitted in support of an application for another degree or qualification of this or any other university or other institution of learning. ii ABSTRACT Real world problems are described by non-linear and chaotic processes which makes them hard to model and predict. The aim of this dissertation is to determine the structure and weights of a polynomial neural network, using evolutionary computing methods, and apply it to the non-linear time series prediction problem. This dissertation first develops a general framework of evolutionary computing methods. Genetic Algorithms, Niched Genetic Algorithms and Evolutionary Algorithms are introduced and their applicability to neural networks optimisation is examined. Following, the problem of time series prediction is formulated. The time series prediction problem is formulated as a system identification problem, where the input to the system is the past values of a time series, and its desired output is the future values of a time series. Then, the Group Method of Data Handling (GMDH) algorithms are examined in detail. These algorithms use simple partial descriptions, usually polynomials, to gradually build complex models. The hybrid method of GMDH and GAs, Genetics-Based Self-Organising Network (GBSON), is also examined. The method implemented for the time series prediction problem is based on the GBSON method. It uses a niched generic algorithm to determine the partial descriptions of the final model, as well as the structure of the neural network used to model the time series to be predicted. Finally, the results obtained with this method are compared with the results obtained by the GMDH algorithm. iii ACKNOWLEDMENTS I would like to express my sincere gratitude to my supervisor Mr. P. Liatsis, for his interest, help and invaluable guidance given throughout the work for this project. I would like to thank all my friends for being there for me and making Manchester a nice place to live in. Also, I would like to thank Thodoros and Anthoula for using their laptop to run the simulations that enabled me to save invaluable time. This dissertation is dedicated to my parents, Fotis Fokas and Lambrini Foka, for believing in me and making it possible for me to get here. This dissertation is also dedicated to my sister, Elpiniki, and my brother, Chrysanthos, for contributing in their own way to support my studies at UMIST. iv CONTENTS