An Application for Semi-Automatic Differentiation

Abstract

A large software project involving the optimization of cellular phone systems required evaluating a complicated function and its derivatives. Due to stringent limitations on run-time and memory consumption, all the derivatives were hand-coded. We discuss the time and memory costs for the hand-coded derivatives and compare them with the costs of automatic differentiation. We also consider the software development costs for hand-coded derivatives and suggest some approaches for adding as much automation as possible subject to the constraint that overall run time and memory consumption is similar to that of the hand-coded derivatives. We also consider methods for partial automation while preserving the efficiency of hand-coded derivatives, and we discuss some preliminary efforts in this direction. There need to be more tools for partial automation so that applications such as this one can get at least some of the benefits of automatic differentiation.

1 Figure or Table

Cite this paper

@inproceedings{Hobby2004AnAF, title={An Application for Semi-Automatic Differentiation}, author={John D. Hobby}, year={2004} }