Nicos Angelopoulos James Cussens
Department of
Computer Science
University of York, York, UK
We describe a collection of Prolog programs which implements a
generic method for applying MCMC over model structures defined by Stochastic
Logic Programs (SLPs). In broad terms, an SLP is used to
constructively define all possible statistical models that may
explain some input data and also to define a prior distribution
over these models. MCMC is then run over these possible
models by making moves on the tree defining all computation paths
of the SLP. The methodology is generic in two ways. Firstly, in
being applicable to any statistical model space that can be
expressed as an SLP and secondly, in allowing for a variety of
alternative proposal moves. For the former, the sole prerequisite
is that we have the means of computing the likelihood of each
generated model: that is, for computing the probability of data
given the model.
