Improved Diagonal Hessian Approximations for Large-Scale Unconstrained Optimization

Abstract

We consider some diagonal quasi-Newton methods for solving large-scale unconstrained optimization problems. A simple and effective approach for diagonal quasi-Newton algorithms is presented by proposing new updates of diagonal entries of the Hessian. Moreover, we suggest employing an extra BFGS update of the diagonal updating matrix and use its diagonal again. Numerical experiments on a collection of standard test problems show, in particular, that the proposed diagonal quasi-Newton methods perform substantially better than certain available diagonal methods.