-- |
-- Module : Control.Auto.Process.Random
-- Description : Entropy generationg 'Auto's.
-- Copyright : (c) Justin Le 2015
-- License : MIT
-- Maintainer : justin@jle.im
-- Stability : unstable
-- Portability : portable
--
-- This module provides 'Auto's (purely) generating entropy in the form of
-- random or noisy processes, as well as 'Auto's to purify/seal
-- 'Auto's with underlying entropy.
--
-- Note that every 'Auto' and combinator here is completely deterministic
-- --- given the same initial seed, one would expect the same stream of
-- outputs on every run. Furthermore, if a serializable 'Auto' is
-- serialized and resumed, it will continue along the deterministic path
-- dictated by the /original/ seed given.
--
-- All of these 'Auto's and combinators come in three flavors: one
-- serializing one that works with any serializable 'RandomGen' instance,
-- one serializing one that works specifically with 'StdGen' from
-- "System.Random", and one that takes any 'RandomGen' (including 'StdGen')
-- and runs it without the ability to serialize and resume
-- deterministically.
--
-- The reason why there's a specialized 'StdGen' version for all of these
-- is that 'StdGen' actually doesn't have a 'Serialize' instance, so a
-- rudimentary serialization process is provded with the 'StdGen' versions.
--
-- The first class of generators take arbitrary @g -> (b, g)@ functions:
-- "Generate a random @b@, using the given function, and replace the seed
-- with the resulting seed". Most "random" functions follow this pattern,
-- including 'random' and 'randomR', and if you are using something from
-- ,
-- then you can use the 'runRand' function to turn a @'Rand' g b@ into a @g
-- -> (b, g)@, as well:
--
-- @
-- 'runRand' :: 'RandomGen' g => 'Rand' g b -> (g -> (b, g))
-- @
--
-- These are useful for generating noise...a new random value at every
-- step. They are entropy sources.
--
-- Alternatively, if you want to give up parallelizability and determinism
-- and have your entire 'Auto' be sequential, you can make your entire
-- 'Auto' run under 'Rand' or 'RandT' as its internal monad, from
-- .
--
-- @
-- 'Auto' ('Rand' g) a b
-- 'Auto' ('RandT' g m) a b
-- @
--
-- In this case, if you wanted to pull a random number, you could do:
--
-- @
-- 'effect' 'random' :: ('Random' r, 'RandomGen' g) => 'Auto' ('Rand' g) a r
-- 'effect' 'random' :: ('Random' r, 'RandomGen' g) => 'Auto' ('RandT' g m) a r
-- @
--
-- Which pulls a random @r@ from "thin air" (from the internal 'Rand'
-- monad).
--
-- However, you lose a great deal of determinism from this method, as your
-- 'Auto's are no longer deterministic with a given seed...and resumability
-- becomes dependent on starting everything with the same seed every time
-- you re-load your 'Auto'. Also, 'Auto''s are parallelizable, while
-- @'Auto' ('Rand' g)@s are not.
--
-- As a compromise, you can then "seal" away the underlying monad with
-- 'sealRandom', which takes an @'Auto' ('RandT' g m) a b@, a starting @g@,
-- and turns it into a normal @'Auto' m a b@, with no underlying randomness
-- monad.
--
-- In this way, you can run any 'Auto' under 'Rand' or 'RandT' as if it was
-- a normal 'Auto' "without" underlying randomness. This lets you compose
-- your sequential/non-parallel parts in 'Rand'...and the later, use it as
-- a part of a parallelizable/potentially non-sequential 'Auto''. It's
-- also convenient because you don't have to manually split and pass around
-- seeds to every 'Auto' that requires entropy.
--
-- The other generators given are for useful random processes you might run
-- into. The first is a 'Blip' stream that emits at random times with the
-- given frequency/probability. The second works /Interval/ semantics from
-- "Control.Auto.Interval", and is a stream that is "on" or "off", chunks
-- at a time, for random lengths. The average length of each on or off
-- period is controlled by the parameter you pass in.
--
module Control.Auto.Process.Random (
-- * Streams of random values from random generators
rands
, stdRands
, rands_
, randsM
, stdRandsM
, randsM_
-- * Lifting/wrapping random functions
, arrRand
, arrRandM
, arrRandStd
, arrRandStdM
, arrRand_
, arrRandM_
-- * Random processes
-- ** Bernoulli (on/off) processes
, bernoulli
, stdBernoulli
, bernoulli_
-- ** Random-length intervals
, randIntervals
, stdRandIntervals
, randIntervals_
-- * Underlying entropy monads
-- ** Sealers
, sealRandom
, sealRandomStd
, sealRandom_
-- ** Processes
, bernoulliMR
, randIntervalsMR
) where
import Control.Applicative
import Control.Auto.Blip
import Control.Auto.Blip.Internal
import Control.Auto.Core
import Control.Auto.Effects
import Control.Auto.Interval
import Control.Category
import Control.Monad (guard)
import Control.Monad.Random
import Data.Bits
import Data.Serialize
import Data.Tuple
import Prelude hiding (id, (.), concat, concatMap, sum)
-- | Given a seed-consuming generating function of form @g -> (b, g)@
-- (where @g@ is the seed, and @b@ is the result) and an initial seed,
-- return an 'Auto' that continually generates random values using the
-- given generating funcion.
--
-- You'll notice that most of the useful functions from "System.Random" fit
-- this form:
--
-- @
-- 'random' :: 'RandomGen' g => g -> (b, g)
-- 'randomR' :: 'RandomGen' g => (b, b) -> (g -> (b, g))
-- @
--
-- If you are using something from ,
-- then you can use the 'runRand' function to turn a @'Rand' g b@ into a @g
-- -> (b, g)@:
--
-- @
-- 'runRand' :: 'RandomGen' g => 'Rand' g b -> (g -> (b, g))
-- @
--
--
-- Here is an example using 'stdRands' (for 'StdGen'), but 'rands' works
-- exactly the same way, I promise!
--
-- >>> let g = mkStdGen 8675309
-- >>> let a = stdRands (randomR (1,100)) g :: Auto' a Int
-- >>> let (res, _) = stepAutoN' 10 a ()
-- >>> res
-- [67, 15, 97, 13, 55, 12, 34, 86, 57, 42]
--
--
-- Yeah, if you are using 'StdGen' from "System.Random", you'll notice that
-- 'StdGen' has no 'Serialize' instance, so you can't use it with this; you
-- have to either use 'stdRands' or 'rands_' (if you don't want
-- serialization/resumability).
--
-- In the context of these generators, resumability basically means
-- deterministic behavior over re-loads...if "reloading", it'll ignore the
-- seed you pass in, and use the original seed given when originally saved.
--
rands :: (Serialize g, RandomGen g)
=> (g -> (b, g)) -- ^ random generating function
-> g -- ^ random generator seed
-> Auto m a b
rands r = mkState (\_ g -> g `seq` r g)
{-# INLINE rands #-}
-- | Like 'rands', but specialized for 'StdGen' from "System.Random", so
-- that you can serialize and resume. This is needed because 'StdGen'
-- doesn't have a 'Serialize' instance.
--
-- See the documentation of 'rands' for more information on this 'Auto'.
--
stdRands :: (StdGen -> (b, StdGen)) -- ^ random generating function
-> StdGen -- ^ random generator seed
-> Auto m a b
stdRands r = mkState' (read <$> get) (put . show) (\_ g -> r g)
{-# INLINE stdRands #-}
-- | The non-serializing/non-resuming version of 'rands'.
rands_ :: RandomGen g
=> (g -> (b, g)) -- ^ random generating function
-> g -- ^ random generator seed
-> Auto m a b
rands_ r = mkState_ (\_ g -> r g)
{-# INLINE rands_ #-}
-- | Like 'rands', except taking a "monadic" random seed function @g ->
-- m (b, g)@, instead of @g -> (b, g)@. Your random generating function
-- has access to the underlying monad.
--
-- If you are using something from
-- , then you
-- can use the 'runRandT' function to turn a @'RandT' g m b@ into a @g ->
-- m (b, g)@:
--
-- @
-- 'runRandT' :: ('Monad' m, 'RandomGen' g)
-- => 'RandT' g m b -> (g -> m (b, g))
-- @
--
randsM :: (Serialize g, RandomGen g, Monad m)
=> (g -> m (b, g)) -- ^ (monadic) random generating function
-> g -- ^ random generator seed
-> Auto m a b
randsM r = mkStateM (\_ g -> r g)
{-# INLINE randsM #-}
-- | Like 'randsM', but specialized for 'StdGen' from "System.Random", so
-- that you can serialize and resume. This is needed because 'StdGen'
-- doesn't have a 'Serialize' instance.
--
-- See the documentation of 'randsM' for more information on this 'Auto'.
--
stdRandsM :: Monad m
=> (StdGen -> m (b, StdGen)) -- ^ (monadic) random generating function
-> StdGen -- ^ random generator seed
-> Auto m a b
stdRandsM r = mkStateM' (read <$> get) (put . show) (\_ g -> r g)
{-# INLINE stdRandsM #-}
-- | The non-serializing/non-resuming version of 'randsM'.
randsM_ :: (RandomGen g, Monad m)
=> (g -> m (b, g)) -- ^ (monadic) random generating function
-> g -- ^ random generator seed
-> Auto m a b
randsM_ r = mkStateM_ (\_ g -> r g)
{-# INLINE randsM_ #-}
-- | Takes a "random function", or "random arrow" --- a function taking an
-- input value and a starting seed/entropy generator and returning a result
-- and an ending seed/entropy generator --- and turns it into an 'Auto'
-- that feeds its input into such a function and outputs the result, with
-- a new seed every time.
--
-- >>> let f x = randomR (0 :: Int, x)
-- >>> streamAuto' (arrRandStd f (mkStdGen 782065)) [1..10]
-- -- [1,2,3,4,5,6,7,8,9,10] <- upper bounds
-- [1,2,0,1,5,3,7,6,8,10] -- random number from 0 to upper bound
--
-- If you are using something from
-- , then you
-- can use the @('runRand' .)@ function to turn a @a -> 'Rand' g b@ into
-- a @a -> g -> (b, g)@:
--
-- @
-- ('runRand' .) :: 'RandomGen' g => (a -> 'Rand' g b) -> (a -> g -> (b, g))
-- @
--
-- (This is basically 'mkState', specialized.)
arrRand :: (Serialize g, RandomGen g)
=> (a -> g -> (b, g)) -- ^ random arrow
-> g -- ^ random generator seed
-> Auto m a b
arrRand = mkState
-- | Like 'arrRand', except the result is the result of a monadic action.
-- Your random arrow function has access to the underlying monad.
--
-- If you are using something from
-- , then you
-- can use the @('runRandT' .)@ function to turn a @a -> 'RandT' m g b@
-- into a @a -> g -> m (b, g)@:
--
-- @
-- ('runRandT' .) :: 'RandomGen' g => (a -> 'RandT' g b) -> (a -> g -> m (b, g))
-- @
arrRandM :: (Monad m, Serialize g, RandomGen g)
=> (a -> g -> m (b, g)) -- ^ (monadic) random arrow
-> g -- ^ random generator seed
-> Auto m a b
arrRandM = mkStateM
-- | Like 'arrRand', but specialized for 'StdGen' from "System.Random", so
-- that you can serialize and resume. This is needed because 'StdGen'
-- doesn't have a 'Serialize' instance.
--
-- See the documentation of 'arrRand' for more information on this 'Auto'.
--
arrRandStd :: (a -> StdGen -> (b, StdGen)) -- ^ random arrow
-> StdGen -- ^ random generator seed
-> Auto m a b
arrRandStd = mkState' (read <$> get) (put . show)
-- | Like 'arrRandM', but specialized for 'StdGen' from "System.Random", so
-- that you can serialize and resume. This is needed because 'StdGen'
-- doesn't have a 'Serialize' instance.
--
-- See the documentation of 'arrRandM' for more information on this 'Auto'.
--
arrRandStdM :: (a -> StdGen -> m (b, StdGen)) -- ^ (mondic) random arrow
-> StdGen -- ^ random generator seed
-> Auto m a b
arrRandStdM = mkStateM' (read <$> get) (put . show)
-- | The non-serializing/non-resuming version of 'arrRand'.
arrRand_ :: RandomGen g
=> (a -> g -> (b, g)) -- ^ random arrow
-> g -- ^ random generator seed
-> Auto m a b
arrRand_ = mkState_
-- | The non-serializing/non-resuming version of 'arrRandM'.
arrRandM_ :: RandomGen g
=> (a -> g -> m (b, g)) -- ^ (monadic) random arrow
-> g -- ^ random generator seed
-> Auto m a b
arrRandM_ = mkStateM_
-- | Simulates a :
-- a process of sequential independent trials each with a success of
-- probability @p@.
--
-- Implemented here is an 'Auto' producing a blip stream that emits
-- whenever the bernoulli process succeeds with the value of the received
-- input of the 'Auto', with its probability of succuss per each trial as
-- the 'Double' parameter.
--
-- It is expected that, for probability @p@, the stream will emit a value
-- on average once every @1/p@ ticks.
--
bernoulli :: (Serialize g, RandomGen g)
=> Double -- ^ probability of any step emitting
-> g -- ^ random generator seed
-> Auto m a (Blip a)
bernoulli p = mkState (_bernoulliF p)
-- | Like 'bernoulli', but specialized for 'StdGen' from "System.Random",
-- so that you can serialize and resume. This is needed because 'StdGen'
-- doesn't have a 'Serialize' instance.
--
-- See the documentation of 'bernoulli' for more information on this
-- 'Auto'.
--
stdBernoulli :: Double -- ^ probability of any step emitting
-> StdGen -- ^ random generator seed
-- (between 0 and 1)
-> Auto m a (Blip a)
stdBernoulli p = mkState' (read <$> get) (put . show) (_bernoulliF p)
-- | The non-serializing/non-resuming version of 'bernoulli'.
bernoulli_ :: RandomGen g
=> Double -- ^ probability of any step emitting
-> g -- ^ random generator seed
-- (between 0 and 1)
-> Auto m a (Blip a)
bernoulli_ p = mkState_ (_bernoulliF p)
_bernoulliF :: RandomGen g
=> Double
-> a
-> g
-> (Blip a, g)
_bernoulliF p x g = (outp, g')
where
(roll, g') = randomR (0 :: Double, 1) g
outp | roll <= p = Blip x
| otherwise = NoBlip
-- | 'bernoulli', but uses an underlying entropy source ('MonadRandom')
-- to get its randomness from, instead of an initially passed seed.
--
-- You can recover exactly @'bernoulli' p@ by using @'sealRandom'
-- ('bernoulliMR' p)@.
--
-- See 'sealRandom' for more information.
bernoulliMR :: MonadRandom m
=> Double -- ^ probability of any step emiting
-- (between 0 and 1)
-> Auto m a (Blip a)
bernoulliMR p = arrM $ \x -> do
roll <- getRandomR (0, 1)
return $ if roll <= p
then Blip x
else NoBlip
-- TODO: rewrite for AMP
-- | An 'Interval' that is "on" and "off" for contiguous but random
-- intervals of time...when "on", allows values to pass as "on" ('Just'),
-- but when "off", suppresses all incoming values (outputing 'Nothing').
--
-- You provide a 'Double', an @l@ parameter, representing the
-- average/expected length of each on/off interval.
--
-- The distribution of interval lengths follows
-- a .
-- This distribution is, as we call it in maths, "memoryless", which means
-- that the "time left" that the 'Auto' will be "on" or "off" at any given
-- time is going to be, on average, the given @l@ parameter.
--
-- Internally, the "toggling" events follow a bernoulli process with a @p@
-- parameter of @1 / l@.
--
randIntervals :: (Serialize g, RandomGen g)
=> Double -- ^ expected length of on/off intervals
-> g -- ^ random generator seed
-> Interval m a a
randIntervals l = mkState (_randIntervalsF (1/l)) . swap . random
-- | Like 'randIntervals', but specialized for 'StdGen' from
-- "System.Random", so that you can serialize and resume. This is needed
-- because 'StdGen' doesn't have a 'Serialize' instance.
--
-- See the documentation of 'randIntervals' for more information on this
-- 'Auto'.
--
stdRandIntervals :: Double -- ^ expected length of on/off intervals
-> StdGen -- ^ random generator seed
-> Interval m a a
stdRandIntervals l = mkState' (read <$> get)
(put . show)
(_randIntervalsF (1/l))
. swap . random
-- | The non-serializing/non-resuming version of 'randIntervals'.
randIntervals_ :: RandomGen g
=> Double -- ^ expected length of on/off intervals
-> g -- ^ random generator seed
-> Interval m a a
randIntervals_ l = mkState_ (_randIntervalsF (1/l)) . swap . random
_randIntervalsF :: RandomGen g
=> Double
-> a
-> (g, Bool)
-> (Maybe a, (g, Bool))
_randIntervalsF thresh x (g, onoff) = (outp, (g', onoff'))
where
(roll, g') = randomR (0 :: Double, 1) g
onoff' = onoff `xor` (roll <= thresh)
outp = x <$ guard onoff
-- should this be onoff' ?
-- | 'randIntervals', but uses an underlying entropy source ('MonadRandom')
-- to get its randomness from, instead of an initially passed seed.
--
-- You can recover exactly @'randIntervals' l@ by using @'sealRandom'
-- ('randIntervalsMR' l)@.
--
-- See 'sealRandom' for more information.
--
randIntervalsMR :: MonadRandom m
=> Double -- ^ expected length of on/off intervals
-> Interval m a a
randIntervalsMR l = flip mkStateM Nothing $ \x monoff -> do
onoff <- case monoff of
Just oo -> return oo
Nothing -> getRandom
roll <- getRandomR (0, 1)
let onoff' = onoff `xor` (roll <= thresh)
return (x <$ guard onoff, Just onoff')
where
thresh = 1/l
-- | Takes an 'Auto' over an 'Rand' or 'RandT' underlying monad as an
-- entropy source, and "seals it away" to just be a normal 'Auto' or
-- 'Auto'':
--
-- @
-- 'sealRandom' :: 'Auto' ('Rand' g) a b -> g -> 'Auto'' a b
-- @
--
-- You can now compose your entropic 'Auto' with other 'Auto's (using '.',
-- and other combinators) as if it were a normal 'Auto'.
--
-- Useful because you can create entire programs that have access to an
-- underlying entropy souce by composing with 'Rand'...and then, at the end
-- of it all, use/compose it with normal 'Auto's as if it were a "pure"
-- 'Auto'.
sealRandom :: (RandomGen g, Serialize g, Monad m)
=> Auto (RandT g m) a b -- ^ 'Auto' to seal
-> g -- ^ initial seed
-> Auto m a b
sealRandom a g0 = mkAutoM (sealRandom <$> resumeAuto a <*> get)
(saveAuto a *> put g0)
$ \x -> do
((y, a'), g1) <- runRandT (stepAuto a x) g0
return (y, sealRandom a' g1)
-- | The non-serializing/non-resuming version of 'sealRandom_'. The random
-- seed is not re-loaded/resumed, so every time you resume, the stream of
-- available randomness begins afresh.
sealRandom_ :: (RandomGen g, Serialize g, Monad m)
=> Auto (RandT g m) a b -- ^ 'Auto' to seal
-> g -- ^ initial seed
-> Auto m a b
sealRandom_ a g0 = mkAutoM (sealRandom_ <$> resumeAuto a <*> pure g0)
(saveAuto a)
$ \x -> do
((y, a'), g1) <- runRandT (stepAuto a x) g0
return (y, sealRandom_ a' g1)
-- | Like 'sealRandom', but specialized for 'StdGen' from "System.Random",
-- so that you can serialize and resume. This is needed because 'StdGen'
-- doesn't have a 'Serialize' instance.
--
-- See the documentation of 'sealRandom' for more information on this
-- combinator.
--
sealRandomStd :: Monad m
=> Auto (RandT StdGen m) a b -- ^ 'Auto' to seal
-> StdGen -- ^ initial seed
-> Auto m a b
sealRandomStd a g0 = mkAutoM (sealRandomStd <$> resumeAuto a <*> (read <$> get))
(saveAuto a *> put (show g0))
$ \x -> do
((y, a'), g1) <- runRandT (stepAuto a x) g0
return (y, sealRandomStd a' g1)