Tid: 23 februari 2014 kl 15.15-16.00.Seminarierummet 3721, Institutionen för matematik, KTH, Lindstedtsvägen 25, plan 7. Karta!
Föredragshållare: Professor Błazej Miasojedow, Institute of Applied Mathematics and Mechanics, University of Warsaw
Titel: MCMC algorithm for CTBN's via particle Gibbs
Abstract In the talk will be presented an MCMC sampler for posterior distribution of CTBN trajectory given partial and noisy observation. The MCMC algorithm is based on idea of so called virtual jumps proposed in Rao&Teh 'Fast MCMC Sampling for Markov Jump Processes and Extensions', JMLR 2013. In present algorithm forward filtering backward sampling (FFBS) used in original version of Rao&Teh is replaced by particle Gibbs. For the large state space our modification is lees computationally expensive. In addition it allows us to consider an infinite state space case. Efficiency will be demonstrated by a numerical experiments.
The talk is based on joint work with W. Niemiro.
|Sidansvarig: Filip Lindskog