Abstract


"On the modified absolute-value factorization norm for trust-region minimization"
N.I.M. Gould, J. Nocedal.
High Performance Algorithms and Software in Nonlinear Optimization, pp. 225-241,
eds. R. De Leone, A. Murli, P. M. Pardalos, G. Toraldo (1998) Kluwer

A trust-region method for unconstrained minimization, using a trust-region norm based upon a modified absolute-value factorization of the model Hessian, is proposed. It is shown that the resulting trust-region subproblem may be solved using a single factorization. In the convex case, the method reduces to a backtracking Newton linesearch procedure. The resulting software package is available as HSL_VF06 within the Harwell Subroutine Library. Numerical evidence shows that the approach is effective in the nonconvex case.
Download (pdf)