Modifications of the Limited Memory BFGS Algorithm for Large-scale Nonlinear Optimization

In this paper we present two new numerical methods for unconstrained large-scale optimization. These methods apply update formulae, which are derived by considering different techniques of approximating the objective function. Theoretical analysis is given to show the advantages of using these update formulae. It is observed that these update formulae can be employed within the framework of limited memory strategy with only a modest increase in the linear algebra cost. Comparative results with limited memory BFGS (L-BFGS) method are presented.