Optimization of the neural structure based on domain expert previous knowledge via GA and sensitivity factors

In this work, the optimization of a neural structure by means of genetic algorithms based on the sensitivity factors, as criterion of the best representatives of a generation selection is presented. As all optimization procedure has the objective to find a neural network structure capable to represent quantitative and qualitatively the process, the sensitivity factors, calculated directly of the neural networks during the training process, are considered. These factors, when compared with the knowledge a priori of the process, represented through symbolic rules, confirm not only the quantitative aspect as well as the qualitative aspect of the process being represented through a specific structure. The results obtained and the time (epochs) to reach the neural network target, applied for the cold rolling process, show that this structure is promising.