A comparison of methods for traversing non-convex regions in optimization problems

This paper considers again the well-known problem of dealing with non-convex regions during the minimization of a nonlinear function F(x) by Newton-like methods. The proposal made here involves a curvilinear search along an approximation to the continuous steepest descent path defined by the solution of the ODE dx/dt = -grad F(x). The algorithm we develop and describe has some features in common with trust region methods; and we present some numerical experiments in which its performance is compared with some other ODE-based and trust region methods.