Skip to Main Content
The particle filter is a powerful filtering technique that is able to handle a broad scope of nonlinear problems. However, it has also limitations: a standard particle filter is unable to handle, for instance, systems that include static variables (parameters) to be estimated together with the dynamic states. This limitation is due to the well-known “self-resolving” phenomenon, which is caused by the gradual loss of information that occurs during the resampling steps. In the context of online Bayesian parameter estimation, some approaches to handle this problem have proposed, such as adding artificial dynamics to the parameter model. However, these approaches typically both introduce new parameters (e.g. the intensity of artificial process noise) and inherent biases to the estimation problem. In this paper, we will give a give a look at two Sequential Monte Carlo techniques that do not rely on biasing the system model: the Autonomous Multiple Model particle filter and the Rao-Blackwellized Marginal particle filter. These approaches are not new, but have not been applied yet to the problem of online Bayesian parameter estimation for non-structured models. We will derive suitable adaptations of these methods for this problem and evaluate them using simulations.