Metropolis-hastings improved particle smoother and marginalized models

This paper combines the Metropolis-Hastings Improved Particle Smoother (MHIPS) with marginalized models. It demonstrates the effectiveness of the combination by looking at two examples; a degenerate model of a double integrator and a fifth order mixed linear/nonlinear Gaussian (MLNLG) model. For the MLNLG model two different methods are compared with the non-marginalized case; the first marginalizes the linear states only during the filtering, the second marginalizes during both the foward filtering and backward smoothing pass. The results demonstrate that marginalization not only improves the overall performance, but also increases the rate of improvement for each iteration of the MHIPS algorithm. It thus reduces the required number of iterations to beat the performance of a Forward-Filter Backward Simulator approach for the same model.