Computational Differentiation

CD Home

Welcome to the Computational Differentiation (CD) homepage providing course and additional online information.

CD is a technique for augmenting computer codes in a rule-based fashion so that the resulting codes can efficiently compute derivatives that are needed, for example, in optimization or parameter identification applications. The associativity of the underlying chain rule of differential calculus allows for a variety of algorithmic choices, resulting in codes that deliver the same results, but at potentially greatly different cost. This lecture describes the principles behind CD and its ramfications in applying CD to large simulation packages.

Preliminaries:

• Since course announcements will be given here students are encouraged to check this page regularly.  Also make sure that you sign up for this course in Campus-Office .
• Lectures take place on Thursdays from 12.00-13.30 in seminar room 5056, beginning on April 13th, 2006.
• There will be no lecture on June 1, 2006.

Materials:

• A short tour covering some of the issues we will be dealing with in the lecture.
• Recipes for Adjoint Generation by Ralf Giering and Thomas Kaminski: Summary, transparencies, paper.
• The Minpack driven cavity problem as an example of a hierarchical view of AD [pdf]