Particle methods for optimal filter derivative: application to parameter estimation

Particle filtering techniques are a popular set of simulation-based methods to perform optimal state estimation in nonlinear nonGaussian dynamic models. However, in applications related to control and identification, it is often necessary to be able to compute the derivative of the optimal filter with respect to parameters of the dynamic model. Several methods have already been proposed in the literature. In experiments, the approximation errors increase with the dataset length. We propose here original particle methods to approximate numerically the filter derivative. In simulations, these methods do not suffer from the problem mentioned. Applications to batch and recursive parameter estimation are presented.