This paper provides novel particle learning (PL) methods for sequential filtering, parameter learning and smoothing in a general class of state space models. The approach extends existing particle methods by incorporating unkown fixed parameters, utilizing sufficient statistics, for parameters and/or states, and allowing for nonlinearities in the model. We also show how to solve the state smoothing problem, integrating out parameter uncertainty. We show that our algorithms outperform MCMC, as well as existing particle filtering algorithms.
Keywords: Particle learning, filtering, smoothing, mixture Kalman filer, state space models.
The manuscript is in available in PDF format.