Use this URL to cite or link to this record in EThOS:
Title: Computational issues in process optimisation using historical data
Author: Nawi, Nazri Mohd
Awarding Body: Swansea University
Current Institution: Swansea University
Date of Award: 2007
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
This thesis presents a new generic approach to improve the computational efficiency of neural-network-training algorithms and investigates the applicability of its 'learning from examples'' featured in improving the performance of a current intelligent diagnostic system. The contribution of this thesis is summarised in the following two points: For the first time in the literature, it has been shown that significant improvements in the computational efficiency of neural-network algorithms can be achieved using the proposed methodology based on using adaptive-gain variation. The capabilities of the current Knowledge Hyper-surface method (Meghana R. Ransing, 2002) are enhanced to overcome its existing limitations in modelling an exponential increase in the shape of the hyper-surface. Neural-network techniques, particularly back-propagation algorithms, have been widely used as a tool for discovering a mapping function between a known set of input and output examples. Neural networks learn from the known example set by adjusting its internal parameters, referred to as weights, using an optimisation procedure based on the 'least square fit principle'. The optimisation procedure normally involves thousands of iterations to converge to an acceptable solution. Hence, improving the computational efficiency of a neural-network algorithm is an active area of research. Various options for improving the computational efficiency of neural networks have been reviewed in this thesis. It has been shown in the existing literature that the variation of the gain parameter improves the learning efficiency of the gradient-descent method. However, it can be concluded from previous researchers' claims that the adaptive-gain variation improved the learning rate and hence the efficiency. It was discovered in this thesis that the gain variation has no influence on the learning rate; however, it actually influences the search direction. This made it possible to develop a novel approach that modifies the gradient-search direction by introducing the adaptive-gain variation. The proposed method is robust and has been shown that it can easily be implemented in all commonly used gradient- based optimisation algorithms. It has also been shown that it significantly improves the computational efficiency as compared to existing neural-network training algorithms. Computer simulations on a number of benchmark problems are used throughout to illustrate the improvement proposed in this thesis. In a foundry a large amount of data is generated within the foundry every time a casting is poured. Furthermore, with the increased number of computing tools and power there is a need to develop an efficient, intelligent diagnostic tool that can learn from the historical data to gain further insight into cause and effect relationships. In this study the performance of the current Knowledge Hyper-surface method was reviewed and the mathematical formulation of the current Knowledge Hyper-surface method was analysed to identify its limitations. An enhancement is proposed by introducing mid-points in the existing shape formulation. It is shown that the midpoints' shape function can successfully constrain the shape of decision hyper-surface to become more realistic with an acceptable result in a multi-dimensional case. This is a novel and original approach and is of direct relevance to the foundry industry.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available