A new globalization technique for nonlinear conjugate gradient methods for nonconvex minimization (Q555476)

The authors analyze a new multiplier selection in the context of nonlinear conjugate gradient methods for non-convex minimization. For this purpose, they propose a so-called truncated descent (TD) approach to limit the multiplier used to modify the steepest descent direction in a certain sense. This new multiplier is then combined with the Polak-Ribiere-Polyak (PRP) and the Liu-Storey (LS) method to get a TDPRP and a TDLS method. For the TDPRP approach convergence is shown in the strongly convex but also in the non-convex case. Numerical results for some of the CUTEr test problems are shown. This includes a comparison of the TDPRP and the TDLS method with the CG-DESCENT algorithm.