Summary:
In this article, we present a new method for the selection of the most pertinent variables to set at the input of a MLP (Multi Layer Perceptron) or RBF (Radial Basis Function) neural network. This method relies on the statistical analysis of the derivatives of the outputs of the network with respect to the inputs. When all the derivates of the outputs with respect to a given input are statistically zeros, the influence of this input is neglectable ans this input may be eliminated. Illustrations will be provided on several simulated or real temporal series prediction models.
Keywords: Neural networks, time series prediction, variable selection, multi-layer perceptron, radial basis function
Published on paper: July 2001.
Citation:
T. Czernichow, B. Dorizzi, A. Muñoz, Sélection de variables et séries temporelles par analyse statistique des sensibilités. Revue d'Intelligence Artificielle. Vol. 15, nº. 3-4, pp. 411 - 427, July 2001.
