An Isaac Newton Institute Workshop

Recent Advances in Monte Carlo Based Inference

Sequential Monte Carlo Samplers for Parameter Estimation in Latent Variable Models

Authors: Adam Michael Johansen (University of Cambridge), Arnaud Doucet (University of British Columbia), Manuel Davy (LAGIS)

Abstract

We present a non-standard sequential Monte Carlo (SMC) algorithm for obtaining marginal maximum likelihood (or posterior) parameter estimates in latent variable models. In common with simulated annealing, our approach involves sampling from a sequenec of distributions which concentrate their support upon the set of maximisers of our objective function. This approach has two substantial advantages over traditional methods: it is resistant to trapping in local modes, and it does not require the marginal distribution to be known, even pointwise. We present results on a number of challenging problems, using real data.