La Prédiction des Séries Temporelles utilisant les Paradigmes de Soft Computing

Abstract

In this work, the problem of design of neural predictor is studied. In a first part, we present the time series used for the conception of the predicteurs. Than we presente the stochastic methodes used for the prediction of time series and the methodologie of Box and Jinkis. Then a art state of the soft computing techniques: genetic algorithms, the neural networks, fuzzy logic. Taking support on this art state we propose at first the method of optimization of a neuronal predicor based on a real genetic algorithm. This method consists of the simultaneous optimization of the topology of neural networks the control parameters, and the initial intervals of the weights synaptiques. We suggested at first a method of representation of all optimizing neuronal predicteur parameters. This representation used the a coding in real number. Then we described the stage of chromosomes initialization. Finally to test the efficiency of the method, the simulations in 3 domaines of economy , ecology, meteorology