MCMC over SLP defined Model Structures

MCMCMS is a piece of software that constracts Markov chains over model structures defined by Stochastic Logic Programsn.

The main intuition is that an SLP can define the space of all possible models and also assign a probability which reflects our prior belief that a particular model explains the data. In addition to composing such an SLP, the user must provide a function for calculating the likelihood of the data given a model.

MCMCMS was designed to facilitate MCMC experiments over SLPs. It is modular and allows the user to add new models and associated likelihood functions in a simple way. It is written entirely in Prolog and can be run under two systems: SICStus and Yap. It was developed under Linux and is unlikely that it will run on radically different operating systems without changes.

To-date we have experimented with building SLPs that construct BNs, RPDAGs (a super-class of BNs), pedigrees and Classification and Regression Trees. The SLPs and all necessary programs for running experiments over these models are included in the sources.

• Current version : mcmcms-0_3_4-2005db3.tgz (or: mcmcms.tgz).
• User guide : MCMCMS_uguide.ps.gz, html: MCMCMS_uguide.html
• html for local install: MCMCMS_uguide_html.tgz
• Output files from the tutorial in the user guide : tutorial_out.tgz

Publications

• Nicos Angelopoulos and James Cussens. Tempering for Bayesian C&RTs. Proceedings of the 2nd International Conference on Machine Learning (ICML05), Bonn, 2005.
• Nicos Angelopoulos and James Cussens. Exploiting Informative Priors for Bayesian Classification and Regression Trees. Proceedings of the Nineteenth International Joint Conference on Artificial Intelligence , Edinburgh, Scotland, UK, July - August 2005.
• Nicos Angelopoulos and James Cussens. Extended stochastic logic programs for informative priors over C&RTs. In Rui Camacho, Ross King, and Ashwin Srinivasan, editors, Proceedings of the work-in-progress track of the Fourteenth International Conference on Inductive Logic Programming (ILP04), pages 7-11, Porto, September 2004.
• Nicos Angelopoulos and James Cussens. On the implementation of MCMC proposals over stochastic logic programs. In Colloquium on Implementation of Constraint and LOgic Programming Systems. Satellite workshop to ICLP'04, Saint-Malo, France, 2004.
• Nicos Angelopoulos and James Cussens. Markov chain Monte Carlo using tree-based priors on model structure. In Jack Breese and Daphne Koller, editors, Proceedings of the Seventeenth Annual Conference on Uncertainty in Artificial Intelligence (UAI-2001), Seattle, August 2001. Morgan Kaufmann.

Acknowledgements

The first phase of the software's development was supported by the EPSRC grant titled: Induction of Stochastic Logic Programs. The second phase was supported by the EPSRC grant: Stochastic Logic Programs for MCMC, 01/09/03--31/08/05, GR/S30993/01 under their MATHfit programme.