Skip to content
/ dice Public

Exact inference for discrete probabilistic programs. (Research code, more documentation and ergonomics to come)

License

Notifications You must be signed in to change notification settings

SHoltzen/dice

Repository files navigation

CI - Build and Test

dice is a probabilistic programming language focused on fast exact inference for discrete probabilistic programs. For more information for how dice works see the research article here. To cite dice, please use:

@article{holtzen2020dice,
  title={Scaling Exact Inference for Discrete Probabilistic Programs},
  author={Holtzen, Steven and {Van den Broeck}, Guy and Millstein, Todd},
  journal={Proc. ACM Program. Lang. (OOPSLA)},
  publisher={ACM},
  doi={https://doi.org/10.1145/342820},
  year={2020}
}

Installation

Docker

A docker image is available, and can be installed with:

docker pull sholtzen/dice

Building From Source

The following steps set up the environment for building dice. First install opam (version 2.0 or higher) following the instructions here. Then, install rust following the commands here. Then, run the following in your terminal:

opam init                  # must be performed before installing opam packages
opam switch create 4.09.0  # switch to use OCaml version 4.09
eval `opam config env`     # optional: add this line to your .bashrc
source $HOME/.cargo/env    # set up rust environment
git submodule update --init --recursive    # populate the rsdd subdirectory
opam install . --deps-only # install dependencies

Building

First follow the steps for installation. Then, the following build commands are available:

  • dune build: builds the project from source in the current directory.
  • dune exec dice: runs the dice executable.
  • dune test: runs the test suite
  • dune exec dicebench: runs the benchmark suite.

Quick Start

We will start with a very simple example. Imagine you have two (unfair) coins labeled a and b. Coin a has a 30% probability of landing on heads, and coin b has a 80% chance of landing on heads. You flip both coins and observe that one of them lands heads-side up. What is the probability that coin a landed heads-side up?

We can encode this scenario in dice as the following program:

let a = flip 0.3 in 
let b = flip 0.8 in
let tmp = observe a || b in 
a

The syntax of dice is similar to OCaml. Breaking down the elements of this program:

  • The expression let x = e1 in e2 creates a local variable x with value specified by e1 and makes it available inside of e2.
  • The expression flip 0.3 is true with probability 0.3 and false with probability 0.8. This is how we model our coin flips: a value of true represents a coin landing heads-side up in this case.
  • The expression observe a || b conditions either a or b to be true. This expression returns true. dice supports logical conjunction (||), conjunction (&&), equality (<=>), negation (!), and exclusive-or (^).
  • The program returns a.

You can find this program in resources/example.dice, and then you can run it by using the dice executable:

> dice resources/example.dice
Value	Probability
true	0.348837
false	0.651163

This output shows that a has a 34.8837% chance of landing on heads.

Optimizations

The Dice compiler has the following built-in optimizations and alternative run-time modes that are activated with the following flags:

  • -determinism: replaces deterministic probabilistic choices with non-random choices (i.e., flip 1.0 becomes true). It is recommended that this flag be enabled for most cases.
  • -eager-eval: changes the compilation order to avoid substitution during compilation. Can perform faster than the default compilation order on certain cases.
  • -flip-lifting: removes redundant flip expressions from certain classes of programs -- can increase performance.

Datatypes

In addition to Booleans, dice supports integers, tuples, and lists.

Tuples

Tuples are pairs of values. The following simple example shows tuples being used:

let a = (flip 0.3, (flip 0.8, false)) in
fst (snd a)

Breaking this program down:

  • (flip 0.3, (flip 0.8, false)) creates a tuple.
  • snd e and fst e access the first and second element of e respectively.

Running this program:

> dice resources/tuple-ex.dice
Value	Probability
true	0.800000
false	0.200000

Unsigned Integers

dice supports distributions over unsigned integers. An example program:

let x = discrete(0.4, 0.1, 0.5) in 
let y = int(2, 1) in 
x + y

Breaking this program down:

  • discrete(0.4, 0.1, 0.5) creates a random integer that is 0 with probability 0.4, 1 with probability 0.1, and 2 with probability 0.3.
  • int(2, 1) creates a 2-bit integer constant with value 1. All integer constants in dice must specify their size.
  • x + y adds x and y together. All integer operations in dice are performed modulo the size (i.e., x + y is implicitly modulo 4 in this case). dice supports the following integer operations: +, *, /, -, ==, !=, <, <=, >, >=.

Running this program:

> dice resources/int-ex.dice
Value	Probability
0	0.
1	0.4
2	0.1
3	0.5

Various distributions over integers have their own syntax. For instance,

  • uniform(3, 2, 6) creates a random 3-bit integer, containing a uniform distribution over the integers 2, 3, 4, 5.
  • binomial(3, 4, 0.5) creates a random 3-bit integer, containing a binomial distribution with parameters n=4, p=0.5

Lists

dice supports distributions over lists, possibly of different lengths.

let xs = [flip 0.2, flip 0.4] in
if flip 0.5 then (head xs) :: xs else tail xs

Breaking this program down:

  • [flip 0.2, flip 0.4] creates a list of Booleans with two elements.
  • head xs returns the first element of xs and tail xs returns a list of everything after the first element.
  • x :: xs returns a list with x added to the front of xs.

Running this program:

> dice -max-list-length 3 resources/list-ex.dice
Value   Probability
[]      0.
[true]  0.2
[false] 0.3
[true, true]    0.
[true, false]   0.
[false, true]   0.
[false, false]  0.
[true, true, true]      0.04
[true, true, false]     0.06
[true, false, true]     0.
[true, false, false]    0.
[false, true, true]     0.
[false, true, false]    0.
[false, false, true]    0.16
[false, false, false]   0.24

Functions

dice supports functions for reusing code. A key feature of dice is that functions are compiled once and then reused during inference.

A simple example program:

fun conjoinall(a: bool, b: (bool, bool)) {
  a && (fst b) && (snd b)
}
conjoinall(flip 0.5, (flip 0.1, true))

Breaking this program down:

  • A function is declared using the syntax fun name(arg1: type1, arg2: type2, ...) { body }.
  • A program starts by listing all of its functions. Then, the program has a main body after the functions that is run when the program is executed. In this program, the main body is conjoinall(flip 0.5, (flip 0.1, true)).
  • Right now recursion is not supported.
  • Functions do not have return statements; they simply return whatever the last expression that evaluated returns.

Result of running this program:

Value	Probability
true	0.050000
false	0.950000

Caesar Cipher

Here is a more complicated example that shows how to use many dice features together to model a complicated problem. We will decrypt text that was encrypted using a Caesar cipher. We can decrypt text that was encrypted using a Caesar cipher by frequency analysis: using our knowledge of the rate at which English characters are typically in order to infer what the underlying key must be.

Consider the following simplified scenario. Suppose we have a 4-letter language called FooLang consisting of the letters A, B, C, and D. Suppose that for this language, the letter A is used 50% of the time when spelling a word, B is used 25% of the time, and C and D are both used 12.5% of the time.

Now, we want to infer the most likely key given after seeing some encrypted text, using knowledge of the underlying frequency of letter usage. Initially we assume that all keys are equally likely. Then, we observe some encrypted text: say the string CCCC. Intuitively, the most likely key should be 2: since A is the most common letter, the string CCCC is most likely the encrypted string AAAA. Let's use dice to model this.

The following program models this scenario in dice:

fun sendChar(key: int(2), observation: int(2)) {
  let gen = discrete(0.5, 0.25, 0.125, 0.125) in    // sample a FooLang character
  let enc = key + gen in                            // encrypt the character
  observe observation == enc
}
// sample a uniform random key: A=0, B=1, C=2, D=3
let key = discrete(0.25, 0.25, 0.25, 0.25) in
// observe the ciphertext CCCC
let tmp = sendChar(key, int(2, 2)) in
let tmp = sendChar(key, int(2, 2)) in
let tmp = sendChar(key, int(2, 2)) in
let tmp = sendChar(key, int(2, 2)) in
key

Now we break this down. First we look at the sendChar function:

  • It takes two arguments: key, which is the underlying secret encryption key, and observation, which is the observed ciphertext.
  • The characters A,B,C,D are encoded as integers.
  • A random character gen is sampled according to the underlying distribution of characters in FooLang.
  • Then, gen is encrypted by adding the key (remember, addition occurs modulo 4 here).
  • Then, ciphertext character is observed to be equal to the encrypted character.

Next, in the main program body, we sample a uniform random key and encrypt the string CCCC. Running this program:

> dice resources/caesar-ex.dice
Value	Probability
0	0.003650
1	0.058394
2	0.934307
3	0.003650

This matches our intuition that 2 is the most likely key.

More Examples

More example dice programs can be found in the source directories:

  • The test/Test.ml file contains many test case programs.
  • The benchmarks/ directory contains example programs that are run during benchmarks.

Syntax

The parser for dice is written in menhir and can be found in lib/Parser.mly. The complete syntax for dice in is:

ident := ['a'-'z' 'A'-'Z' '_'] ['a'-'z' 'A'-'Z' '0'-'9' '_']*
binop := +, -, *, /, <, <=, >, >=, ==, !=, &&, ||, <=>, ^, ::

expr := 
   (expr)
   | ident
   | true
   | false
   | int (size, value)
   | discrete(list_of_probabilities) 
   | uniform(size, start, stop)
   | binomial(size, n, p)
   | expr <binop> expr
   | (expr, expr)
   | fst expr
   | snd expr
   | ! expr
   | flip probability
   | observe expr
   | if expr then expr else expr
   | let ident = expr in expr
   | [ expr (, expr)* ]
   | [] : type
   | head expr
   | tail expr
   | length expr

type := bool | (type, type) | int(size) | list(type)
arg := ident: type
function := fun name(arg1, ...) { expr }

program := expr 
        | function program

About

Exact inference for discrete probabilistic programs. (Research code, more documentation and ergonomics to come)

Resources

License

Stars

Watchers

Forks

Packages

No packages published