Dagger

dagger is a embedded domain-specific language for probabilistic programming in OCaml. Models are written in monadic style. Sampling from the posterior is performed using one of the following algorithms:

A basic example

Let us consider the classical "sprinkler" example (drawn from Functional Programming for Modular Bayesian Inference).

Our lawn can become wet either because it rained or because the sprinkler is activated. We construct a simple model that encodes our belief that the lawn becomes wet knowing that it rained or not and the sprinkler state.

open Lmh_inference
open Infix

let model : bool t =
  let open Infix in
  let* rain = sample (Stats_dist.bernoulli ~bias:0.2) in
  let* sprinkler = sample (Stats_dist.bernoulli ~bias:0.1) in
  let prob_lawn_wet =
    match (rain, sprinkler) with
    | (true, true) -> 0.99
    | (true, false) -> 0.7
    | (false, true) -> 0.9
    | (false, false) -> 0.01
  in
  let+ () = score prob_lawn_wet in
  rain

Each execution trace is weighted by the density of the samples taken in that trace and by calls to the score construct. The inference algorithm samples execution traces proportionally to their total weight. Scoring therefore amounts to a form of soft conditioning (and assigning a score of 0 to a trace amounts to rejecting that trace). Let us sample from the posterior implicitly defined by our model.

let rng_state = RNG.make [| 0x1337; 0x533D |]

let nsamples = 100_000

let samples =
  Lmh_inference.stream_samples model rng_state
  |> Seq.take nsamples

let freq =
  (samples
  |> Seq.map (fun x -> if x then 1 else 0)
  |> List.of_seq |> List.fold_left ( + ) 0 |> float_of_int)
  /. float_of_int nsamples

The result is that the probability that it rains knowing that the lawn is wet is approximately equal to 0.64.

More examples

Companion libraries

In the interest of modularity, sampleable distributions (such as bernoulli) are defined in auxilliary libraries. Currently, we have:

API