AbstractIn this thesis, we propose a unified framework for a class of algorithms based on approximate solutions of proximal point subproblems. The purpose is to describe a new general approximate solution criterion. This enhances the constructive approximation approach of the recently proposed hybrid projection-priximal and extra-gradient-proximal methods. Our method, with this more flexible criterion, possesses global convergence and local linear rate of convergence under standard assumptions. We apply this method to splitting formard-backward methods and to bundle methods.
|Date of Award||24 Jun 2005|
|Supervisor||Jean-Jacques STRODIOT (Supervisor), Van Hien Nguyen (Jury) & Frank CALLIER (Jury)|
Un cadre unifié pour les méthodes de point proximal inexactes
Harvanek, C. (Author). 24 Jun 2005
Student thesis: Master types › Master in Mathematics