Improving robustness of artificial neural networks model using genetic algorithm

Abstract

Artificial Neural Networks (ANN) has been widely accepted as process estimators due its ability to capture complex relationships. However, experiences in implementing ANN estimators in research and industry have exposed some weakness that can be detrimental to the overall performance of plant operations. Among these, the issue of robustness is of particular importance. This paper proposes adaptation of networks weight as means to improve robustness. Comparisons between GA approach and conventional backpropagation in adaptation of weights are in on-line estimation and control of fatty acid composition in a distillation column. Significant improvements were obtained by the adaptive model especially model generalization perspective.