SCHOOL OF COMPUTER SCIENCE AND SOFTWARE ENGINEERING
MONASH UNIVERSITY


TECHNICAL REPORT 2002/107


Message from Monte Carlo

L J Fitzgibbon, D L Dowe and L Allison

ABSTRACT

We present a first step towards a general methodology for statistical inference in the Minimum Message Length (MML) framework using Markov Chain Monte Carlo (MCMC) methods.
The Minimum Message Length concept of a two-part message is used to partition a sample from an importance sampling density of the model parameters into coding regions.  The regions contain models that are similar and hence the method  can be loosely described as model clustering.  For each region a point estimate is chosen as the representative model and we use Dowe's recent MMLD approximation as the message length criterion.
The method is made practical by using the MML Boundary Rule and importance sampling.  The MML Boundary Rule is a result that applies to MML approximations in a common form and which states that the boundary of the optimal coding region will be isometric or approximately isometric.  This means that we need only consider regions with isometric boundaries.  Importance sampling is used to efficiently evaluate expectations  and integrals, using Monte Carlo integration, over promising isometrically bounded regions of the parameter space.  We note that the regions do not have to be restricted to contigious (simply connected) regions of the parameter space and therefore efficient codebooks can be approximated.
Our approach can, in theory, be used to approximate message lengths for any problem where the prior probability density and likelihood function are known and a suitable sampler exists.  We treat the posterior as a black-box and make no assumptions about its mathematical form or the structure of its parameter space.  The method requires a sample from a suitable importance sampling density and a means of approximating the Kullback-Leibler distance between any two models.  For general use, we suggest that the posterior is suitable for the importance sampling.  In this case the Message from Monte Carlo methodology becomes aligned with Bayesian posterior sampling and can be considered as a means of introspection into the posterior sample.
We give an example of the method applied to univariate polynomial regression.  Polynomial regression was chosen because it is a useful and intuitive problem for which the MML87 method can and already has been applied.  We use the posterior as an importance sampling density and the polynomial parameters are sampled using Reversible Jump Markov Chain Monte Carlo.