Analysis of a functional “approacher” utilizing the artificial neural networks MLP trained by the “back propagation” algorithm.

Abstract

The neural networks are mathematical models which have analogical features like those which are of the biological neural networks. How it works is based on the structure of the real neuron where each one receives data (electric signals) from the innumerous dendrites, and these signals can be reduced or amplified, depending on the dendrite to which it is associated. Each dendrite has a mass which stores all the data of a neural network and its updating is what one knows at the learning process. The aim of this assignment is to carry out a study on the characteristics of the neural networks and the algorithm “back propagation” to join the continuous functions with more precision. In this sense a simulator was developed which carries out the training and the graphic follow-up of the same. The neural network used showed satisfactory results during these test sessions, demonstrating an excellent capacity of generalization. Corroborating with the idea of modeling of functions with no need of prior knowledge of the same, the applicability of this system can be extended to the control area.

Key word

Neural network, functional approacher, back propagation, mathematical models, generalization.