Approche non paramétrique du filtrage de système non linéaire à temps discret et à paramètres inconnus

Particle filters are presently among the most powerful tools for filtering discrete time non linear systems. However the presence of unknown parameters in the system equations makes their use more complex and can even impair their convergence properties. This Note shows how an on-line consistent estimation of these parameters can be obtained simultaneously to that of the state variables to be filtered. This approach relies upon a kernel-based non parametric estimation of conditional probability densities from successive Monte Carlo generations of system particles. To cite this article: V. Rossi, J.-P. Vila, C. R. Acad. Sci. Paris, Ser. I 340 (2005).