ccprism: Probabilistic programming as a library using delimited control

This package provides several services for working with probabilistic models and is based on the functionality of PRISM. Models are written as Prolog programs enriched with extra computational effects: probabilistic choice and tabling. Programs can be run in sampling mode or explanation mode. Explanation mode results in a hypergraph representing the computation, which can then be processed to get:

• inside probabilities (generalised sum-product algorithm)
• the single best explanation (generalised Viterbi algorithm)
• any number of explanations in order of probability (lazy k-best algorithm)
• outside probabilities for computing parameter sufficient statistics Based on these several EM parameter learning methods are provided: maximum likelihood, maximum a posterior, variational Bayes, and Viterbi learning. Deterministic annealling can be used with all of these methods.

  A couple of MCMC explanation sampling methods are also provided.

Usage Examples

You can load the test module included in the examples directory like this:

swipl -g 'consult(pack(ccprism/examples/test))'

More information on how to use the system to follow... NB. the test module requires the memo pack to be installed.

There are also other examples, including lazy.pl which shows how another layer of state can be used to get lazy samplers and thereby implement random world semantics, and crp.pl, which is an experiment in implementing Dirichlet processes (fairly inefficiently) using CRPs on top of ccprism.

Note on tabling implementations

Earlier versions used a tabling implementation where non-backtrackable state was stored using services from the ccnbenv library module. This was replaced with a trie-based data structure (see https://github.com/samer--/cctable) which is a lot faster. However, in doing so, we lost the ability to store attributed variables in the variant and solution tries. The old version is still available in the branch old_tabling_with_attributes.