Automatic- versus Manual- differentiation for non-linear inverse modeling

Abstract : Emerging tools for automatic differentiation (AD) of computer programs should be of great benefit for the implementation of many derivative-based numerical methods such as those used for inverse modeling. The Odyssée software, one such tool for Fortran 77 codes, has been tested on a sample model that solves a 2D non-linear diffusion-type equation. Odyssée offers both the forward and the reverse differentiation modes, that produce the tangent and the cotangent models, respectively. The two modes have been implemented on the sample application. A comparison is made with a manually-pr- oduced differentiated code for this model (MD), obtained by solving the adjoint equations associated with the model's discrete state equations. Following a presentation of the methods and tools and of their relative advantages and drawbacks, the performances of the codes produced by the manual and automatic methods are compared, in terms of accuracy and of computing efficiency (CPU and memory needs). The perturbation method (finite-difference approximation of derivatives) is also used as a reference. Based on the test of Taylor, the accuracy of the two AD modes proves to be excellent and as high as machine precision permits, a strong indication of Odyssée's capability to produce error-free codes. Comparatively, the manually-produced derivatives (MD) sometimes appear to be slightly biased, which is likely due to the fact that a theoretical model (state equations) and a practical model (computer program) do not exactly coincide, while the accuracy of the perturbation method is very uncertain. The MD code largely outperforms all other methods in computing efficiency, a matter of current research for the improvement of AD tools. It is reckoned though that such tools can already be of considerable help for the computer implementation of many numerical methods, avoiding the tedious task of hand-coding the different- iation of complex algorithms.
Type de document :
Rapport
[Research Report] RR-3981, INRIA. 2000, pp.41
Liste complète des métadonnées

https://hal.inria.fr/inria-00072666
Contributeur : Rapport de Recherche Inria <>
Soumis le : mercredi 24 mai 2006 - 10:32:20
Dernière modification le : samedi 27 janvier 2018 - 01:32:14
Document(s) archivé(s) le : dimanche 4 avril 2010 - 23:17:27

Fichiers

Identifiants

Citation

David Elizondo, Christèle Faure, Bernard Cappelaere. Automatic- versus Manual- differentiation for non-linear inverse modeling. [Research Report] RR-3981, INRIA. 2000, pp.41. 〈inria-00072666〉

Partager

Métriques

Consultations de la notice

326

Téléchargements de fichiers

481