Participation à une conférence de Mathématiques (appliquées à la physique) : l'IC-Msquare 2013 de Prague

Time series modeling with pruned multi-layer perceptron and 2-stage damped least-squares method

Voyant C, Tamas W, Paoli C, Muselli M, Nivet M, Notton G


Participation à une conférence de Mathématiques (appliquées à la physique) : l'IC-Msquare 2013 de Prague

final_paper.doc final_paper.doc  (276.5 Ko)

Cette conférence ("International Conference on Mathematical Modeling in Physical Sciences) sera à l'occasion pour nous de présenter nos résultats concernant un nouvel algorithme d'apprentissage des MLP.
Ce dernier baptisé "2-stage damped least-squares method" permet d'obtenir des réseaux "prunés" automatiquement.
Cet outil permet de faciliter le temps d'optimisation et d'améliorer le recherche des valeurs de poids et de biais lors de modélisation de séries temporelles météorologiques et de prédiction à horizon. L'abstract de cette présentation est:

"A Multi-Layer Perceptron (MLP) defines a family of artificial neural networks often used in TS modeling. Because of the “black boxes” aspect, many researchers refuse to use it. Moreover, the optimization (often based on the exhaustive approach where “all” configurations are tested) and learning phases of this artificial intelligence tool (often based on the Levenberg-Marquardt algorithm; LMA) are weaknesses of this approach. These two tasks must be repeated depending on the knowledge of each new problem studied, making the process, long, laborious and not robust. In this paper a pruning process is proposed. This method allows, during the training phase, to carry out an inputs selecting method activating (or not) inter-nodes connections and to verify both if optimization phase is shorten and forecasting are improved. We propose to use iteratively the popular damped least-squares method to activate inputs and neurons. A first pass is applied to 10% of the learning sample to determine weights significantly different from 0 and delete other. Then a classical batch process based on LMA is used with the new MLP. We validated our method on 10 measured meteorological TS with different behaviors (stationarity and periodicity). We used 8 years for the training and 2 years for the test, results of the classical LMA and the 2-stage LMA are cross-compared."


Rédigé par Cyril VOYANT le Mardi 7 Mai 2013 à 19:20 | Lu 522 fois