| Popis: |
One of the obstacles to applying neural networks in industry is a lack of confidence in model predictions. Although neural networks are very powerful and flexible for mapping non-linear relationships between the input and output, it is often the lack of assurance in their prediction which prevents their use in engineering applications. Recently, there has been increased interest in tackling this lack of confidence by adding some confidence bounds on predictions. In this paper a methodology of establishing the confidence bound using an ensemble of neural network models will be presented. There are several other advantages of using an ensemble neural network model instead of a single neural network, among these being the improvement of the overall prediction accuracy, the smoothness of the response surface and generalisation ability. The ensemble modelling approach will be demonstrated by an example application to mechanical property prediction for steel products. Initial results are promising, with the confidence bound consistent with metallurgical experience. |
