DEVELOPING A NEW CONJUGATE GRADIENT ALGORITHM WITH THE BENEFIT OF SOME DESIRABLE PROPERTIES OF THE NEWTON ALGORITHM FOR UNCONSTRAINED OPTIMIZATION

The conjugate gradient method and the Newton method are both numerical optimization techniques. In this paper, we aim to combine some desirable characteristics of these two methods while avoiding their drawbacks, more specifically, we aim to develop a new optimization algorithm that preserves some essential features of the conjugate gradient algorithm, including the simplicity, the low memory requirements, the ability to solve large scale problems and the convergence to the solution regardless of the starting vector (global convergence). At the same time, this new algorithm approches the quadratic convergence behavior of the Newton method in the numerical sense while avoiding the computational cost of evaluating the Hessian matrix directly and the sensitivity of the selected starting vector. To do this, we propose a new hybrid conjugate gradient method by linking (CD) and (WYL) methods in a convex blend, the hybridization paramater is computed so that the new search direction accords with the Newton direction, but avoids the computational cost of evaluating the Hessian matrix directly by using the secant equation. This makes the proposed algorithm useful for solving large scale optimization problems. The sufficient descent condition is verified, also the global convergence is proved under a strong Wolfe Powel line search. The numerical tests show that, the proposed algorithm provides the quadratic convergence behavior and confirm its efficiency as it outperformed both the (WYL) and (CD) algorithms.