Transparent neural network data modelling

Roadknight, CM, 2000. Transparent neural network data modelling. PhD, Nottingham Trent University.

Preview

Text 10290128.pdf - Published version Download (35MB) | Preview

Abstract

The research set out in this thesis was carried out with the aim of making the adoption of Neural Networks for real world problem solving more likely. It attempts to guide the reader in methods of application and provide novel tools for successful adoption. The testing ground for this thesis is a biological problem, but the findings of the research are applicable to any real world problem where the number and complexity of causative agents make deciding their actions complex.

The use of ANNs as predictors of natural phenomena is an important application but equally important is any resulting explanation of the heuristics a network uses to achieve this prediction. The novel methods of equation synthesis and correlated activation pruning (CAPing) are introduced and used to extract meaning from a trained ANN. Equation synthesis involves the incremental increase in the number of connections of the trained ANN used until satisfactory prediction is achieved. CAPing involves the identification of nodes that have similar effects on the desired output. Comparison of the inputs to these nodes can lead to useful dependancy relationships. Several useful generalisations have been made in this project by using these methods. Generalisations have been made using ANNs, equation synthesis and CAPing. For example, the temporal dynamics of an ozone exposure are evidenced as being more important than the quantity of the ozone exposure and light levels are shown to be the most important modifying factor in the injury process.

When the concentration of ground level ozone reaches significant levels, severe and economically important damage can occur to agricultural crop plants. Environmental modifying factors affect the expression of this injury. The use of artificial neural networks (ANNs) as tools for ozone damage prediction is investigated in this project and novel approaches to extracting meaning from these networks are examined.

Two sets of biological results have been investigated which have allowed injury development in semi-natural and natural environments to be modelled with satisfactory predictive success. In the first instance, it was possible, given seven days of data for ozone and climatic conditions, to predict if further injury will develop on the leaves during these 7 days and, with less accuracy, the amount of leaf area effected. For the second model, a neural network can predict accurately if leaf injury will be expressed on the following day when given detailed data for the levels of ozone and light on the 5 preceding days.

ANNs have been further used to create a set of rules by which the onset of injury can be discriminated, the performance of these rules is only slightly inferior to the ANNs and their explicit nature makes them valuable.

Biological experimentation has been carried out to confirm any generalisations and add to the clarification process. This final step in the research sequence completes a cycle that ends with more data being generated that can be used for further network training.

The effect of Ozone, and its modifying factors, on yield is modelled using ANNs. The structures of these yield predicting models are explained. The importance of the nature of the ozone episode and modifying factors is analysed.

The methods discussed in this thesis could be applied to any suitably complex and multivariate data set. A degree of transparency, not generally expected from neural network approaches, would be apparent. Therefore the theories would be suitable for both neural network practitioners looking to add more transparency to their modelling and for the traditional data modeller looking for alternative techniques for their complex data set but fearful of ANNs traditional 'black box' downfall.

Item Type:	Thesis
Creators:	Roadknight, C.M.
Date:	2000
ISBN:	9781369323771
Identifiers:	Number Type PQ10290128 Other
Divisions:	Schools > School of Science and Technology
Record created by:	Linda Sullivan
Date Added:	02 Oct 2020 13:45
Last Modified:	03 Oct 2023 15:52
URI:	https://irep.ntu.ac.uk/id/eprint/41120

Actions (login required)

Edit View

Statistics

Views

Downloads