Convolution particle filters for parameter estimation in general state-space models

The state-space modeling of partially observed dynamic systems generally requires estimates of unknown parameters. From a practical point of view, it is relevant in such filtering contexts to simultaneously estimate the unknown states and parameters. Efficient simulation-based methods using convolution particle filters are proposed. The regularization properties of these filters is well suited, given the context of parameter estimation. Firstly the usual non Bayesian statistical estimates are considered: the conditional least squares estimate (CLSE) and the maximum likelihood estimate (MLE). Secondly, in a Bayesian context, a Monte Carlo type method is presented. Finally these methods are compared in several simulated case studies. Key-words: Hidden Markov models, parameter estimation, particle filter, convolution kernels, conditional least squares estimate, maximum likelihood estimate