This was a course project for Statistical Computing, instructed by Arnaud Doucet. It covered general Bayesian methodology, then delved into several sampling methods, especially Markov chain Monte Carlo (MCMC).

My course project investigated Bayesian fitting of Gaussian mixture models (with a variable number of components). I applied David Hastie's recent (2005) AutoMix software to the problem. This software attempts to make reversible jump (trans-dimensional) MCMC accessible to non-expert statisticians by automating the crafting of cross-model jump proposals.

Long story short, it is ill suited for performing posterior sampling for this model, and failed miserably. If you care to read up on the details, here is my: report (pdf), presentation (ppt) and code (c, Matlab). I am not particularly proud of my project, but provide it here for posterity. The included version of AutoMix may be of interest though, if you use Microsoft Visual C++ (I had to make some minor syntactic changes to get it to compile).

Back to my main page.