Monitoring and selection of dynamic models by Monte Carlo sampling

Time-varying signals and systems are usually modeled by state-space models. Many of these models are nonlinear and the noise that enters their description may be non-Gaussian. For such models standard Kalman filtering approaches are not optimal, and more general methods are then of interest. A group of methods that have recently attracted much attention are based on Monte Carlo sampling and propagation of relevant densities by propagating their samples. In this paper, the objective is to study the monitoring and selection of models whose processing is carried out by these methods. As a metric for model comparison, the logarithms of the models' predictive densities of the observed data are used. To demonstrate the validity of the approach, an example is presented and performance results provided.