A modified CG-DESCENT method for unconstrained optimization (Q535462)

This paper presents a modification of the CG-DESCENT proposed by Hager and Zhang yielding the estimate \(g_k^Td_k = -\|g_k\|^2\) instead of \(g_k^Td_k \leq -7\|g_k\|^2/8\) as in the original paper. The authors prove global convergence of their algorithm for strongly convex functions when using a line search parameter such that the Wolfe conditions are fulfilled. Furthermore, global convergence is shown for a general nonlinear function when the line search parameter fulfils the strong Wolfe conditions. The section on numerical results compares the performance of the modified CG-DESCENT method with the original one and two variants of the Polak-Ribiere-Polyak approach for 73 test cases.