Jories Bierkens: Piecewise deterministic Monte Carlo
Time: Mon 2019-09-30 15.15 - 16.15
Participating: Jories Bierkens, VU Amsterdam
Markov Chain Monte Carlo (MCMC) is an essential computational tool in quantitative fields such as Bayesian statistics, physics and machine learning. The goal of MCMC is to perform computations with respect to a probability distribution of interest, which is often only implicitly specified in terms of an unnormalized density and which furthermore often cannot be used in numeric integration due to a curse of dimensionality.
In recent years piecewise deterministic Markov processes (PDMPs) have emerged as a promising alternative to classical MCMC algorithms. One of the attractive features of this approach is the possibility of `exact subsampling' of the data. In this talk PDMP based algorithms will be introduced and recent progress in our understanding of the underlying processes will be presented.