The Trajectory Problem in Automatic Differentiation
by
Christele Faure
INRIA Sophia Antipolis
Coauthors: Uwe Naumann (University of Hertfordshire)

The problem of efficiently computing a compact information allowing to reverse a program using the reverse mode of AD is examined. We distinguish between two components of the trajectory - control and Jacobian information. Problems with the version 1.7 of Odyssee will serve as motivation for our approach.

The proposed method is built on a boolean status flag attached to all program variables which is propagated through the computational graph. We will propose methods for dealing with branches, loops, and subroutines - constructs within programs which usually complicate analyses.

Numerical results illustrating the efficiency of our approach will be presented by comparing the sizes of the trajectory generated by the standard algorithm used in Odyssee version 1.7 with the results obtained using the proposed analysis.

http://homepages.feis.herts.ac.uk/~comqun/AD2000/taad2_abstr.ps

Date received: February 11, 2000