A conjugate direction gradient method with reconnaissance steps for unconstrained minimization (Q1113622)

A new conjugate direction type gradient method is proposed for solving an unconstrained minimization problem in \(R^ n\). At each iterate the algorithm generates n-1 ``reconnaissance points'' to gather more information about the behaviour of the minimized function and then executes n-1 line searches along directions generated which are shown to be conjugate in the quadratic case. A global convergence theorem is proved under assumption that all the reconnaissance points are contained in the initial level set and an additional condition is satisfied. A comparison with the Fletcher-Reeves method for the Rosenbrock test function is given.