Subconjunto de treinamento e critério de confiabilidade para Redes Neurais Artificiais de Domínio Real, Complexo e de Clifford

AUTOR(ES)
DATA DE PUBLICAÇÃO

2009

RESUMO

The use of Artificial Neural Networks (ANN) Multilayer Perceptrons (MLP) Backpropagation (BP) for regression is done in three phases: training, validation and testing. This research proposes a methodology to approach two aspects of these phases: 1st) to estimate a parameter to confront the accuracy of the ANN output classifying those which can be considered reliable and those which are not, therefore in the testing phase of the ANN it will not be possible to evaluate the function of interest for 100% of the data; and 2nd) to establish the minimum number of patterns which allow the convergence and the generalization of those networks to be used in the training phase. The method proposes the use of two networks: Direct ANN (DANN), used to approximate the function of interest, and Inverse ANN (IANN), used to approximate the Inverse Function (IF) of DANN. After the joint training of the two networks, the difference between the input of the DANN and the output of the IANN, should be the same computerwise in the convergence of the networks, it will define the parameters presented. In case the function to be approximated does not have a defined IF, the domain is restricted for where there is one. The method proposed will be demonstrated using synthetic data starting from the quadratic function f(x) = x2, with the objective of controlling the input and the output to demonstrate the validity of the method that will be applied for the ANN MLP in the Real, Complex and Clifford domain. The Clifford (multidimensional) domain is restricted so that it is isomorphous to the complex domain, allowing the graphic visualization of the results and the comparison to the complex domain.

ASSUNTO(S)

artificial neural networks rna engenharia eletrica mlp clifford redes neurais artificiais

ACESSO AO ARTIGO

Documentos Relacionados