An information-theoretic approach to assess practical identifiability of parametric dynamical systems

Sanjay Pant 1 Damiano Lombardi 1
1 REO - Numerical simulation of biological flows
LJLL - Laboratoire Jacques-Louis Lions, Inria Paris-Rocquencourt, UPMC - Université Pierre et Marie Curie - Paris 6
Abstract : A new approach for assessing parameter identifiability of dynamical systems in a Bayesian setting is presented. The concept of Shannon entropy is employed to measure the inherent uncertainty in the parameters. The expected reduction in this uncertainty is seen as the amount of information one expects to gain about the parameters due to the availability of noisy measurements of the dynamical system. Such expected information gain is interpreted in terms of the variance of a hypothetical measurement device that can measure the parameters directly, and is related to practical identifiability of the parameters. If the individual parameters are unidentifiable, correlation between parameter combinations is assessed through conditional mutual information to determine which sets of parameters can be identified together. The information theoretic quantities of entropy and information are evaluated numerically through a combination of Monte Carlo and k-nearest neighbour methods in a non-parametric fashion. Unlike many methods to evaluate identifiability proposed in the literature, the proposed approach takes the measurement-noise into account and is not restricted to any particular noise-structure. Whilst computationally intensive for large dynamical systems, it is easily parallelisable and is non-intrusive as it does not necessitate re-writing of the numerical solvers of the dynamical system. The application of such an approach is presented for a variety of dynamical systems – ranging from systems governed by ordinary differential equations to partial differential equations – and, where possible, validated against results previously published in the literature.
Liste complète des métadonnées

Littérature citée [22 références]  Voir  Masquer  Télécharger

https://hal.inria.fr/hal-01099901
Contributeur : Damiano Lombardi <>
Soumis le : lundi 5 janvier 2015 - 15:08:46
Dernière modification le : samedi 15 décembre 2018 - 01:48:38
Document(s) archivé(s) le : mercredi 3 juin 2015 - 16:42:04

Fichier

entropy.pdf
Fichiers produits par l'(les) auteur(s)

Identifiants

Collections

Citation

Sanjay Pant, Damiano Lombardi. An information-theoretic approach to assess practical identifiability of parametric dynamical systems. Mathematical Biosciences, Elsevier, 2015, pp.66-79. 〈10.1016/j.mbs.2015.08.005〉. 〈hal-01099901〉

Partager

Métriques

Consultations de la notice

372

Téléchargements de fichiers

346