A quasi-Newton trust region method with a new conic model for the unconstrained optimization (Q2378932)

The authors consider a trust region method for the solution of unconstrained nonlinear optimization problems. A quasi-Newton update is used to approximate the second order information and a conic model is employed for the trust region subproblem. Compared to earlier papers, the choice of a parameter for the conic model is modified such that it gives more flexibility during the solution procedure and allows a better approximation of the objective function. Additionally, a dogleg method is applied for the step determination. The corresponding algorithm is discussed in detail and a convergence proof is given. Numerical results are given for academic test problems with up to 2000 variables.