Les méthodes de filtre en optimisation non-linéaire : une comparaison des variantes avec recherche linéaire et région de confiance

Abstract

This work deals with a theoretical analysis of the filter methods for nonlinear constrained optimization problems, which aim at avoiding the use of merit functions. The main difference of the studied algorithms is an approach based on trust region or on line search. For the two trust region methods, the global convergence to first-order critical points is proved. The third method, considering a line search, has the advantage to extend to the local convergence to points that satisfy the second order optimality conditions thanks to second order correction steps. In the three approaches, we decompose the step to the next iterate into its normal and tangential components. A comparison of these methods is proposed on a theoretical plan in a last chapter.

Date of Award	Jun 2002
Original language	French
Supervisor	Philippe TOINT (Supervisor), Annick Sartenaer (Co-Supervisor), Benoît Colson (Jury) & Jean-Jacques STRODIOT (Jury)