Optimization and Systems Theory Seminar
Friday, November 28, 1997, 11.00-12.00, Room 3721, Lindstedtsvägen 25

Michael Patriksson
Department of Mathematics
Chalmers University of Technology
Göteborg, Sweden

Adaptive sequential quadratic programming (SQP) algorithms for non-strictly monotone variational inequality problems

Sequential quadratic programming (SQP) algorithms are classical Newton-type methods for the solution of nonlinear optimization problems with nonlinear constraints. In existing analyses, the convergence of these methods rely on an exact penalty function for the problem to be strictly decreasing. Inherent in the analysis of this property is the requirement that the value of the penalty parameter chosen must be larger than a problem-specific parameter, which is hard to estimate a priori. In the talk, we address the (more general) variational inequality problem, and show how the problem of estimating the proper value of this penalty parameter can be eliminated entirely through the introduction of an adaptive scaling parameter; the new method allows for the penalty parameter to take on any non-negative value. The new analysis also reveals for the first time that SQP methods are globally convergent for variational inequality problems with non-strictly monotone problem mappings.
Calendar of seminars
Last update: November 17, 1997 by Anders Forsgren, andersf@math.kth.se.