A Nonlinear Black-Box Modelling Methodology for Neural Networks

Guy-Michel Cloarec (1) , John Ringwood

1 Technocentre Renault [Guyancourt]

The aim of this paper is to present a nonlinear black-box modelling methodology that, in a general context, introduces the use of various statistical techniques to improve the modelling performance. The improvements are obtained by addressing the main problem that any black-box technique is confronted with: the optimal choice of a model topology and the related parameters with respect to the complexity of the problem. The main idea is to efficiently combine statistical resampling and analysis with preprocessing, optimal topology determination, supervised parameter optimization and ensemble techniques.

Mots clés: Nonlinear black-box modelling; statistical theory; preprocessing; curse of dimensionality; parameter optimization; ensemble techniques; statistical bootstrat; chaotic time series

Domaines: Machine Learning [stat.ML]; Intelligence artificielle [cs.AI]; Optimisation et contrôle [math.OC]; Statistiques [math.ST]; Dynamique Chaotique [nlin.CD]; Calcul parallèle, distribué et partagé [cs.DC]; Algorithme et structure de données [cs.DS]; Modélisation et simulation; Réseau de neurones [cs.NE]; Génie logiciel [cs.SE]; Traitement du signal et de l'image [eess.SP]; Systèmes et contrôle [cs.SY]; Théorie de l'information et codage [math.IT]

Guy-Michel Cloarec, John Ringwood. A Nonlinear Black-Box Modelling Methodology for Neural Networks. 2021. ⟨hal-03235806⟩