An embedded probabilistic programming language.
Switch branches/tags
Nothing to show
Clone or download
Latest commit 37045f8 Mar 6, 2017
Permalink
Failed to load latest commit information.
assets Update README. Mar 5, 2017
lib Add measure interpreter. Mar 5, 2017
test Initial commit. Mar 5, 2017
.gitignore Initial commit. Mar 5, 2017
LICENSE Initial commit. Mar 5, 2017
README.md Fix typo. Mar 6, 2017
Setup.hs Initial commit. Mar 5, 2017
deanie.cabal Add measure interpreter. Mar 5, 2017
stack.yaml Initial commit. Mar 5, 2017

README.md

deanie

MIT License

deanie is an embedded probabilistic programming language. It can be used to denote, sample from, and perform inference on probabilistic programs.

Usage

Programs are written in a straightforward monadic style:

mixture :: Double -> Double -> Program Double
mixture a b = do
  p      <- beta a b
  accept <- bernoulli p
  if   accept
  then gaussian (negate 2) 0.5
  else gaussian 2 0.5

You can sample from them by first converting them into an RVar from random-fu:

> sample (rvar (mixture 1 3))

Sample many times from models using standard monadic combinators like 'replicateM':

> replicateM 1000 (sample (rvar (mixture 1 3)))

Or convert them to measures using a built-in interpreter:

> let nu = measure (mixture 1 3)
> let f = cdf nu

You can perform inference on models using rejection or importance sampling, or use a simple, stateful Metropolis backend. Here's a simple beta-bernoulli model, plus some observations to condition on:

betaBernoulli :: Double -> Double -> Program Bool
betaBernoulli a b = do
  p <- beta a b
  bernoulli p

observations :: [Bool]
observations = [True, True, False, True, False, False, True, True, True]

Here's one way to encode a posterior via rejection sampling:

rposterior :: Double -> Double -> Program Double
rposterior a b =
    grejection
      (\xs ys -> count xs == count ys)
      observations (beta a b) bernoulli
  where
    count = length . filter id

Here's another, via importance sampling:

iposterior :: Double -> Double -> Program (Double, Double)
iposterior a b =
  importance observations (beta a b) logDensityBernoulli

There are also some Monte Carlo convenience functions provided, such as a weighted average for weighted samples returned via importance sampling:

> samples <- replicateM 1000 (sample (rvar (iposterior 1 1)))
> print (mcw samples)
0.6369246537796793

Background

You can read about some of the theory and ideas behind this kind of language in some blog posts I've written.