Skip to Main content Skip to Navigation
Preprints, Working Papers, ...

Numerical methods for sensitivity analysis of Feynman-Kac models

Pierre-Arnaud Coquelin 1, 2 Romain Deguest 2 Rémi Munos 1
1 SEQUEL - Sequential Learning
LIFL - Laboratoire d'Informatique Fondamentale de Lille, Inria Lille - Nord Europe, LAGIS - Laboratoire d'Automatique, Génie Informatique et Signal
Abstract : The aim of this work is to provide efficient numerical methods to estimate the gradient of a Feynman-Kac flow with respect to a parameter of the model. The underlying idea is to view a Feynman-Kac flow as an expectation of a product of potential functions along a canonical Markov chain, and to use usual techniques of gradient estimation in Markov chains. Combining this idea with the use of interacting particle methods enables us to obtain two new algorithms that provide tight estimations of the sensitivity of a Feynman-Kac flow. Each algorithm has a linear computational complexity in the number of particles and is demonstrated to be asymptotically consistent. We also carefully analyze the differences between these new algorithms and existing ones. We provide numerical experiments to assess the practical efficiency of the proposed methods and explain how to use them to solve a parameter estimation problem in Hidden Markov Models. To conclude we can say that these algorithms outperform the existing ones in terms of trade-off between computational complexity and estimation quality.
Document type :
Preprints, Working Papers, ...
Complete list of metadatas

Cited literature [21 references]  Display  Hide  Download
Contributor : Pierre-Arnaud Coquelin <>
Submitted on : Thursday, January 18, 2007 - 4:58:07 PM
Last modification on : Thursday, March 5, 2020 - 6:21:31 PM
Long-term archiving on: : Tuesday, April 6, 2010 - 10:15:48 PM


Files produced by the author(s)


  • HAL Id : hal-00125254, version 1



Pierre-Arnaud Coquelin, Romain Deguest, Rémi Munos. Numerical methods for sensitivity analysis of Feynman-Kac models. 2007. ⟨hal-00125254⟩



Record views


Files downloads