An Approximate Inverse Preconditioner in Truncated Newton Methods for Large Scale Optimization

Resumen

This work considers the use of truncated Newton methods for the solution of large scale unconstrained optimization problems. Two key aspects of truncated Newton methods may be still considered open questions: how to handle the case with indefinite Hessian and how to formulate a general effective preconditioning strategy. We propose the use of Conjugate Gradient-based schemes as a tool for facing up to both the questions. These schemes can be successfully used for computing an efficient Newton-type direction whenever the Hessian is indefinite. Furthermore they enable to define a suitable approximate inverse preconditioning technique for reducing the overall inner iterations. [DOI: 10.1685 / CSC06076] About DOI