Thèse soutenue

Sélection de modèles à l’aide des chemins de régularisation pour l’objectivation mono et multi-prestations : application à l’agrément de conduite.
FR  |  
EN
Accès à la thèse
Auteur / Autrice : Jean-François Germain
Direction : François Roueff
Type : Thèse de doctorat
Discipline(s) : Signal et images
Date : Soutenance en 2008
Etablissement(s) : Paris, ENST

Mots clés

FR

Mots clés contrôlés

Résumé

FR  |  
EN

The thesis, in cooperation with the french car maker RENAULT, is about prestations’ objectivation. In the automotive field, a “prestation” is an item on which the customer has a subjective evaluation. The goal of our methodology is the following: define which physical variables explain this subjective answer. In this context of variable selection, we propose a method inspired by recent works about the well-known and widespread Lasso method. The Lasso method is the name given to the L1-penalized version of the least squares estimation problem (addition of constant times of coefficients L1 norm). We develop an algorithm for the calculation of this path in the L1-penalized logistic regression case. Instead of answering this (difficult and unsolved) problem of selecting the regularization constant, we propose to build from the regularization path a growing collection of models and to select the best model thanks to the Bayesian Information Criterion (BIC). We notice that engineers take a great interest in the reading of the regularization path itself. That’s why we also propose to study the asymptotic behavior of the path as a function of the regularization “constant”. The combination of building the path and using it, through the growing models collection, for selecting the model maximizing the BIC has been implemented and is used as an internal software by RENAULT engineers. Another RENAULT expectation involves the processing of multiple subjective simultaneous answers: we call it “the multi-prestations case”. We also present an adaptation of the tool we developed for handling this case.