UCL Logo

James Cussens

Department of Computer Science, University of York, UK


Wednesday 15  February 2006



Seminar Room B10 (Basement)

Alexandra House, 17 Queen Square, London, WC1N 3AR



Bayesian model averaging with informative priors on model structure

A key feature of the Bayesian approach is the inclusion of non-data information (prior information) into statistical inference. However, in practical applications it can be difficult to incorporate what knowledge we have into an appropriate prior distribution. In this talk, I will present our attempts to address this problem using our MCMCMS (Markov chain Monte Carlo over Model Structures) system ( MCMCMS uses stochastic logic programs to define priors over model structure. A version of the Metropolis-Hastings algorithm is then used to approximately sample from the posterior. I will present results of applying MCMCMS to C&RT models and to Bayesian net models, including the improvement in convergence which 'tempering' brings. (This is joint work with Nicos Angelopoulos.)