Bayesian filtering for hidden Markov models via Monte Carlo methods

We propose a new Monte Carlo method for Bayesian filtering of general nonlinear and non-Gaussian hidden Markov models. This method is an extension of the well known importance sampling method. It is especially well-suited to sequential simulation as it allows one to split or kill trajectories according to a given score function. The model and estimation objectives are described. The new Monte Carlo method is presented. A few results on this method are established and its application to Bayesian filtering is described. Simulation results for several nonlinear and non-Gaussian time series are presented.