py-aiger-coins 3.3.3

Library for creating circuits that encode discrete distributions.

py-aiger-coins

warning 3.0.0 and greater are a major rewrite of this code base. I am trying to port most of the useful features.

Library for creating circuits that encode discrete distributions and Markov Decision Processes. The name comes from the random bit model of drawing from discrete distributions using coin flips.

Table of Contents

• py-aiger-coins
• Install
• Usage

Install

To install this library run:

$ pip install py-aiger-coins

Note that to actually compute probabilities, one needs to install with the bdd option.

$ pip install py-aiger-coins[bdd]

For developers, note that this project uses the poetry python package/dependency management tool. Please familarize yourself with it and then run:

$ poetry install

Usage

py-aiger-coins extends the standard py-aiger-bv and py-aiger-discrete abstractions to allow for certain bits to be set via biased coins.

The library centers around the PCirc object. The easiest way to use py-aiger-coins is to throught the pcirc function.

import aiger_bv as BV
import aiger_coins as C


expr1 = BV.uatom(2, 'x') + BV.uatom(2, 'y')

# Create distribution over bits.
my_pcirc = C.pcirc(expr1) \
            .randomize({'x': {0: 1/6, 1: 2/6, 2: 3/6}})

assert my_pcirc.outputs == {expr1.output}

# This probablistic circuit uses 2 biased coins to represent this input.
assert my_pcirc.num_coins == 2
assert my_pcirc.coin_biases == (1/3, 1/2)

# 'x' input is replaced with coinflips.
assert my_pcirc.inputs == {'y'}

# Underlying circuit now has a single input representing coin inputs.
underlying_circ = my_pcirc.circ

assert underlying_circ.inputs == {'y', my_pcirc.coins_id}

Note that aiger_coins.PCirc implements the same API as aiger_bv and aiger_coins.

Sequential Circuit API

For example, sequential and parallel composition allow combining probablistic circuits.

incr = (BV.uatom(2, 'z') + 1)
adder = (BV.uatom(2, 'x') + BV.uatom(2, 'y')).with_output('z')

# Create distribution over bits.
pcirc = C.pcirc(adder)                                         \
         .randomize({'x': {0: 1/6, 1: 2/6, 2: 3/6}})
pcirc >>= incr

or

inc_x = C.pcirc(BV.uatom(2, 'x') + 1)          \
         .randomize({'x': {0: 1/3, 2: 2/3}})                      # Pr(x=2) = 2/3

inc_y = C.pcirc(BV.uatom(3, 'y') + 1)          \
         .randomize({'y': {0: 1/3, 5: 1/3, 3: 1/3}})

inc_xy = inc_x | inc_y  #

Similarly, unroll, loopback are also implemented.

note unroll combines all coin flips into a single input in temporal order.

Finite Functions

py-aiger-coins also works well with the py-aiger-discrete API for working with arbitrary functions over finite sets. For example:

from bidict import bidict  # `pip install bidict`

# Create encoder/decoder for dice.
lookup = bidict({0: '⚀', 1: '⚁', 2: '⚂', 3: '⚃'})     # invertable dictionary.
encoder = aiger_discrete.Encoding(decode=lookup.get, encode=lookup.inv.get)

# Represent dice with a 2 bit vector.
expr1 = BV.uatom(2, '🎲')

# Add encoded dice to x. Because why not.
expr2 = BV.uatom(2, 'x') + expr1
func = aiger_discrete.from_aigbv(
    expr2.aigbv, input_encodings={'🎲': encoder}
)

# Create distribution over bits.
circ = C.pcirc(func) \
        .randomize({'🎲': {'⚀': 1/6, '⚁': 2/6, '⚂': 3/6}})

assert circ.inputs == {'x'}
assert circ.outputs == {expr2.output}
assert circ.coin_biases == (Fraction(1, 3), Fraction(1, 2))