Master Class: Lecture 3: Bayesian Parameter Inference in State-Space Models using Particle MCMC

SpeakerArnaud Doucet
AffiliationUniversity of Oxford
DateFriday, 05 Oct 2012
Time13:00 - 14:00
LocationRoberts G06 Sir Ambrose Fleming LT
Event seriesMaster Class: Arnaud Doucet (1-5 Oct 2012)

Standard Markov chain Monte Carlo (MCMC) methods to perform Bayesian inference for both states and parameter in state-space models can be very inefficient and/or not even applicable for complex models.

I will discuss how it is possible to come up with a new class of efficient MCMC algorithms using particle filtering proposals. This yields the class of particle MCMC methods.

One crucial practical problem for particle MCMC is the selection of the number of particles for the proposal. Essentially, a trade-off is needed. If too many particles are used, then the particle MCMC scheme has similar properties to the case where the likelihood of the parameter is exactly known but will be expensive. If too few particles are used, this is at the expense of slower mixing in the resulting Markov chain.

I will address how one should select the number of particles so as to maximize the efficiency of the particle MCMC scheme.

Slides: ucl_3.pdf

iCalendar csml_id_108.ics