Numerical study of the speed’s response of the various intelligent models using the tansig, logsig and purelin activation functions in different layers of artificial neural network

Abstract

<p><span lang="EN-US">Today's world is no longer that of yesterday, the pace with which we live and also the speed is enormous and rapid, that overnight we discover the appearance of new technologies and solutions in all the fields, in particular, that of scientific research. Artificial intelligence plays the main role. Predicting the behavior of new materials using artificial neural networks has become a frequently adopted solution by researchers today. The performance of neural networks depends mainly on the activation functions used. This work was designed to mainly study the impact of these functions on the response speed of an artificial neural network in general, and particularly on the model we are working on to predict the thermomechanical behavior of innovative materials. By using TANSIG, PURELIN and LOGSIG in a feed forward back propagation training by Levenberg-Marquardt algorithm, we were able to generate 9 models. For each of these models, we were interested in analyzing the speed’s response of the network and studying its regression. Thus, this work was able to show us that choosing the right neuron activation function from one layer to another can clearly influence the performance of the results. Depending on the problem studied, the desired objective and the chosen architecture, the activation function can radically change the result and provide us with the expected efficiency. </span></p>