Exact Penalties for Variational Inequalities

Author(s)

Thiago Afonso de Andre

• Document type: Master's Dissertation
• Press: São Paulo.
• Institution: Universidade de São Paulo (USP). Instituto de Matemática e Estatística (IME/SBI)
• Defense date: 2007-02-01

Examining board members

Paulo José da Silva e Silva; Roberto Andreani; Marcelo Gomes de Queiroz

• Advisor: Paulo José da Silva e Silva

Abstract

This work intends to build upon differentiable exact penalty methods for nonlinear programming, using them to solve variational inequality problems. Such problems have been given a lot of attention in the literature lately and have applications to diverse areas of knowledge such as Engineering, Physics and Economics. Differentiable exact penalty methods were developed during the 70s and 80s to solve constrained optimization problems by means of the solution of unconstrained problems. Those problems are such that, with an appropriate choice of the penalty parameter, one finds a solution of the original constrained problem by solving only one unconstrained problem. The function which is minimized is similar to the classic augmented lagrangian, but an estimate of the multiplier is automatically calculated from the primal point. In this thesis we show how to couple Glad and Polak?s multiplier estimate, with the classic augmented lagrangian of a variational inequality developed by Auslender and Teboulle. This allowed us to obtain an exact penalty function for variational inequality problems. The best exactness results were obtained in the particular case of nonlinear complementarity problems. An important characteristic of the proposed penalty is that it doesn?t involve second order information of any of the functions which compose the variational inequality. In addition to those results, which are the core of this work, we also present a brief review of inexact differentiable penalties, exact nondifferentiable penalties and differentiable exact penalties in optimization. (AU)