|
| EPSRC Reference: |
GR/S30993/01 |
| Title: |
Stochastic Logic Programs for MCMC |
| Principal Investigator: |
Dr J Cussens |
| Other Investigators: |
|
| Researcher Co-investigator: |
|
| Project Partner: |
|
| Department: |
Computer Science |
| Organisation: |
University of York |
| Scheme: |
Standard Research |
| Starts: |
01 September 2003 |
Ends: |
31 August 2005 |
Value (£): |
126,186
|
| EPSRC Research Topic Classifications: |
| Artificial Intelligence Technologies |
Statistics and Applied Probability |
|
| EPSRC Industrial Sector Classifications: |
|
| Related Grants: |
|
| Panel History: |
|
|
Summary |
|
Interest in complex probabilistic models, and specifically those using a logical or relational framework, has been growing in the Artificial Intelligence community. This project applies one such approach: stochastic logic programs (SLPs). SLPs are parameterised logic programs which compactly and declaratively represent complex probability distributions. Lying at the intersection of statistics and logic programming, this project draws on theoretical and applied work from both disciplines. The project aims to implement an SLP environment for Bayesian inference using Markov chain Monte Carlo (MCMC). MCMC methods have greatly expanded the scope of Bayesian approaches, so that now it is possible, for example, to apply Bayesian approaches to (i) 'learning' the structure of graphical models and (ii) phylogenetic analysis. SLPs allow one to declare arbitrarily complex priors over, graphical models or phylogenetic trees in such a way that all the available prior knowledge ('hard' and 'soft') can be expressed and thus exploited.
|
| Final Report Summary |
In this 2-year project, we have produced software, documentation and scientific papers for our MCMCMS system, all of which is available at the URL below.
A central result is that first-order logic can be used to conveniently express constraints on the structure of statistical models, and moreover, that MCMC-based Bayesian inference benefits from the ability to express such constraints. An empirical demonstration of this can be found in our IJCAI-05 paper.
We have also shown (in our ICML-05 paper) that "tempering" provides an effective mechanism to speed up convergence (at least when applied to Bayesin C&RT).
To meet the objective of the project it was necessary to extend the underlying formalism: Stochastic Logic Programs (SLPs). SLPs can now be parameterised by functions, not only numbers. We have also developed a more flexible SLP_based MCMC proposal mechanism by exploiting a "proof tree" representation of the state of the MCMC sampler.
In unpublished work we have applied our MCMCMS system to learn pedigrees (family trees) from DNA data, and have compared our results to that of the familias system. This work has involved collaboration with pedigree experts from Leicester and Oslo, and has led to a new proposal (currently under review) involving these collaborators.
Our most substantial piece of research is contained in an 80 page paper which we are currently finishing with a view to journal submission. This focuses on Bayesian inference of Bayesian networks and measures the effect of informative priors.
|
| Further Information: |
http://www-users.cs.york.ac.uk/%7Enicos/sware/slps/mcmcms/ |
| Organisation Website: |
http://www.york.ac.uk |
|
|