Use this URL to cite or link to this record in EThOS:
Title: Issues in learning cause and effect relationships from examples : with particular emphasis on casting processes
Author: Ransing, M. R.
Awarding Body: University of Wales Swansea
Current Institution: Swansea University
Date of Award: 2003
Availability of Full Text:
Access from EThOS:
The aim of this research is to provide a self-learning, computer based, decision-making tool for industry. The tool will have a knowledge of current/past rejection levels within the manufacturing set up, the diagnosis done by experts and will use this information to automatically learn a cause and effect relationship. The work presented in this thesis constitutes an algorithm for associating belief values in the occurrence of a cause (or a combination of causes) with corresponding belief values in the occurrence of an effect (or a combination of effects). For this research, the neural network approach was first chosen because of its potential advantages over the traditional rule based approach. The influence of network parameters such as weights, biases, learning rate, momentum term and mode of training, has been analysed on network training. The analysis of the influence of the variation of a gain value during the training and testing phase showed that gain is not an independent parameter as perceived before, but depends on the initial weight values and the learning rate value. It was discovered that the variation in the gain value also influences the learning speed. A coupled algorithm has been proposed in this thesis to change the gain value adaptively. It has been found during this research that neural networks are probably not the best available techniques for learning cause and effect relationships form examples, particularly due to their poor extrapolation abilities on incomplete and noisy training data sets. A novel and efficient method has been proposed, which overcomes the major limitation of the poor extrapolation ability of neural networks. This was achieved by effectively storing prior knowledge about the cause and effect relationship within the network. Enhancements have also been introduced to make this algorithm efficient. The belief value in the occurrence of the likely causes of one or more given effects is determined using this method. The algorithm developed is generic and is applicable to all manufacturing processes and possibly in all situations where the cause and effect relationship is complex and a data set associating belief values in causes and effects is available.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available