2nd European Workshop on Automatic Differentiation

Thursday November 17- Friday November 18, 2005

Whitworth Conference Centre

Cranfield University (Shrivenham Campus)

Defence Academy of the UK

Shrivenham, Swindon

Nick Gould (RAL)

Requirements for AD in Numerical Optimization

Abstract: Few people would disagree that if an optimization problem involves differentiable functions, then every effort should be made to use their derivatives when developing or using algorithms for its solution. Automatic differentiation has opened the door to "error-free" evaluation of complicated function expressions, and has had a large impact in the solution of large-scale optimization problems. Nevertheless, algorithm developers still have a number of hopes for AD that have, to date, not been fully addressed. These include the automatic identification of (partial) separability, the efficient evaluation of Hessian sub-structures (such as the diagonal or a band around the diagonal) for preconditioning purposes, and the calculation of (Lipschitz) bounds on derivative terms for use in global optimization. In this talk, we will review AD requirements from an optimization algorithm developers viewpoint.

Slides: GouldNov05.pdf