An Investigation of Different Methods for Building Robust Multiple Neural Network Models (2001)

Student name: Zainal Ahmad

Degree and date obtained: MSc in Applied Process Control (Distinction), October 2001

Career after study: PhD student at University of Newcastle

Thesis Abstract: Artificial neural networks have been shown to be able to approximate any continuous non-linear functions and have been used to build data base empirical models for non-linear processes. In this thesis two cases of empirical data was present which is for tank level prediction and also for neutralisation prediction. The main advantage of neural networks based processes a model is they are easy to build. This feature is particularly useful when modelling complicated processes where detailed mechanistic models are difficult to develop. However a critical shortcoming of neural network is that they often lack robustness unless a proper networks training and validation procedure is used.

Here model robustness means that a model can still give satisfactory predictions when applied to new unseen data. Model robustness is a basic requirement in advanced process monitoring and control. Several techniques have been recently develop to improve neural networks model robustness, such as training with regularisation, cross validation based ‘early stopping’, staked neural networks, bootstrap aggregation neural networks, committees of networks and Bayesian learning.

Among them, multiple neural networks appear to be the most promising approach. However there is a different method for developing the individual networks and for combining them. Therefore this thesis investigates and compares all the differences in developing multiple neural networks and try to find out how they can be combined.

Z. Ahmad and J. Zhang, "A Comparison of Different Methods for Combining Multiple Neural Networks Models", Proceedings of 2002 World Congress on Computational Intelligence, Honolulu, Hawaii, U.S.A., 12 - 117 May, 2002.