Over the last decade the multilayer perceptron (MLP) artificial neural network (ANN) has been applied increasingly to nonlinear modelling problems in fields such as process control and machine vision. Nonlinear modelling is also a problem which has been studied extensively by statisticians for several decades, and in recent years several people have pointed out that standard MLP and statistical regression methods are in fact very closely related. This is a useful observation because MLP modelling has traditionally been a somewhat trial and error empirical process. Identifying the similarity between MLP and regression methods thus offers the possibility that the large body of existing statistical theory and practice may be used to improve our understanding and use of the MLP. This thesis adopts this approach to examining two important practical problems in MLP regression. These are: the use of robust estimators to improve the fit, particularly when the training data contains outliers, and prediction error estimation for MLP model complexity selection. The investigations into robust MLP regression discovered that only simple robust estimators are likely to be useful in most MLP regression problems. Though more sophisticated estimators have previously been suggested for this task, it is shown why these are in fact unsuited to this. Estimating prediction error is a particularly important problem in MLP regression. The investigations into estimating prediction error yielded a fast method for estimating prediction error by cross-validation and also examined its limitations. This method is particularly useful when the amount of training data is limited. The primary motivation for investigating these two issues was the desire to use the MLP to model a phenomenon known as curl in papermaking, and to use this model to improve the yield of a papercoating process. Only a limited amount of data was available for this task, and it was suspected that the data included several gross errors. Since these are general problems in MLP regression, the techniques devised here have wide applicability and importance.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:659839 |
| Date | January 1997 |
| Creators | Myles, Andrew J. |
| Publisher | University of Edinburgh |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://hdl.handle.net/1842/15471 |
Page generated in 0.002 seconds