Hybrid structure based on previous knowledge and GA to search the ideal neurons quantity for the hidden layer of MLP-Application in the cold rolling process

The neural representation of a physical process has the objective of explaining the cause-effect relationship among the parameters involved in the process. The representation is normally evaluated through the error reached during the training and validation processes. As the neural representation is not based on the physical principles, its mathematical representation can be correct in the quantitative aspect but not in the qualitative one. In this work, it is shown that a neural representation can fail when its qualitative aspect is evaluated. The search of the ideal neurons quantity for the hidden layer of the MLP neural network, by means of Genetic Algorithms and the sensitivity factors calculated directly from the neural networks during the training process, is presented. The new optimization structure has the objective to find a neural network structure capable to represent the process quantitatively and qualitatively. The sensitivity factors, when compared with the expert knowledge of the human agent, represented through symbolic rules, can evaluate not only the quantitative but also the qualitative aspect of the process being represented through a specific neural structure. The results obtained, and the time (epochs) necessaries to reach the neural network target show that this combination is promising. As a case study, the new structure is applied for the cold rolling process.