Title: Pseudo-Marginal Hamiltonian Monte Carlo with Efficient Importance Sampling
Authors: Kjartan Kloster Osmundsen - University of Stavanger (Norway) [presenting]
Tore Selland Kleppe - University of Stavanger (Norway)
Roman Liesenfeld - University of Cologne (Germany)
Abstract: The joint posteriors of latent variables and parameters in Bayesian hierarchical models often have strong nonlinear dependencies, and thus making them challenging targets for standard Markov chain Monte Carlo methods. Pseudo- marginal methods are able to effectively explore such target distributions, by integrating out the latent variables and directly targeting the marginal posteriors of the parameters. The combination of Efficient Importance Sampling for integrating out latent variables and recently proposed pseudo-marginal Hamiltonian Monte Carlo for sampling from parameter marginal is explored. The methodology is shown to be highly efficient in the context of state space models.