What's on in Computing Science?

Date: Thursday, 4 February, 2010
Time: 11:00
Location: Sir Alwyn Williams Building, 422 Seminar Room
[ Inference Seminar ] Hybrid Monte-Carlo in high dimensions
Alex Beskos (University College London)
We have investigated algorithmic properties of Hybrid Monte-Carlo (HMC) in high dimensions. In the simplified scenario of iid target distributions we have found that the asymptotically optimal acceptance probability, for dimensionality growing to infinity, is 0.651 (to three decimal places), irrespectively of the particular (iid) target. We have also worked on a semi-implicit version of the leapfrog integrator, which is relevant for target distributions defined as a change of measure from Gaussian laws arising in applications. We illustrate that implementation of the semi-implicit version of the integrator in such a context allows for the construction of a well-defined HMC algorithm in infinite-dimensional Hilbert spaces.

Contact: Dr Rónán Daly (rdaly@dcs.gla.ac.uk)

Add to my calendar