Skip to Main content Skip to Navigation
Theses

Classification Dynamique de données non-stationnaires :
Apprentissage et Suivi de Classes évolutives

Abstract : Most of natural or artificial processes have evolutionary behaviours described by non-stationary data. This thesis studies the problem of dynamical clustering of non-stationary data. We propose a generic description of dynamical classifiers by using a neural network with evolutionary architecture. It is composed of four learning procedures: creation, adaptation, fusion, and evaluation. From this generic description, we develop two algorithms. The first one is a new version of AUDyC (AUto-adaptive and Dynamical Clustering). AUDyC uses a mixture model described following the multimodal approach. The second one, called SAKM (Self-Adaptive Kernel Machine), is based on SVM & kernel methods. Both are created with some recursive update rules that allow the adaptive modelling and the pursuit (tracking) of evolving clusters. They have self-adaptive abilities in non-stationary environment and good performances of convergence and complexity. These latter are theoretically proved and also illustrated by simulation.
Complete list of metadatas

Cited literature [157 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00106968
Contributor : Habiboulaye Amadou Boubacar <>
Submitted on : Monday, October 16, 2006 - 6:02:00 PM
Last modification on : Tuesday, November 24, 2020 - 2:18:22 PM
Long-term archiving on: : Tuesday, April 6, 2010 - 7:55:38 PM

Identifiers

  • HAL Id : tel-00106968, version 1

Collections

Citation

Habiboulaye Amadou Boubacar. Classification Dynamique de données non-stationnaires :
Apprentissage et Suivi de Classes évolutives. Automatique / Robotique. Université des Sciences et Technologie de Lille - Lille I, 2006. Français. ⟨tel-00106968⟩

Share

Metrics

Record views

693

Files downloads

8844