Monte Carlo methods for inference in dynamical systems

Fredrik Lindsten
Division of Systems and Control
Department of Information Technology
Uppsala University


Abstract:

Sequential Monte Carlo (SMC, a.k.a. Particle Filters) and Markov Chain Monte Carlo (MCMC) are two of the main workhorses for Bayesian inference in general probabilistic models. In this talk I will first introduce these methods in the context of dynamical systems. I will then present a principled way of combining SMC and MCMC into so-called Particle MCMC (PMCMC) methods, enabling joint state and parameter inference in general non-linear/non-Gaussian state space models. In particular, I will present a specific PMCMC method---Particle Gibbs with Ancestor Sampling---that enables computationally efficient state and parameter inferences in general state space models.