{-# LANGUAGE GeneralizedNewtypeDeriving, CPP, BangPatterns, TupleSections, FlexibleInstances, TypeSynonymInstances #-} -- | -- Module : Math.Probable.Distribution.Finite -- License : BSD3 -- Maintainer : alpmestan@gmail.com -- Stability : experimental -- Portability : GHC -- -- /Fun with finite distributions!/ -- -- This all pretty much comes from Eric Kidd's series -- of blog posts at <http://www.randomhacks.net/probability-monads/>. -- -- I have adapted it a bit by making it fit into my own -- random generation/sampling scheme. -- -- The idea and purpose of this module should be clear after going -- through an example. First, let's import the library and 'vector'. -- -- > import Math.Probable -- > import qualified Data.Vector as V -- -- We are going to talk about Books, and particularly about whether a given -- book is interesting or not. -- -- > data Book = Interesting -- > | Boring -- > deriving (Eq, Show) -- -- Let's say we have very particular tastes, and that we think -- that only 20% of all books are interesting (that's not so small actually. oh well). -- -- > bookPrior :: Finite d => d Book -- > bookPrior = weighted [ (Interesting, 0.2) -- > , (Boring, 0.8) -- > ] -- -- 'weighted' belongs to the 'Finite' class, which represents -- types that can somehow represent a distribution over a finite set. -- That makes our distribution polymorphic in how we will use it. Awesome! -- -- So how does it look? -- -- > λ> exact bookPrior -- in ghci -- > [Event Interesting 20.0%,Event Boring 80.0%] -- -- 'exact' takes 'Fin' 'a' and gives you the -- inner list that 'Fin' uses to represent the distribution. -- -- Now, what if we pick two books? First, how do we even do that? -- Well, any instance of 'Finite' must be a 'Monad', so you have your -- good old /do notation/. The ones provided by this package also -- provide 'Functor' and 'Applicative' instances, but let's use -- do. -- -- > twoBooks :: Finite d => d (Book, Book) -- > twoBooks = do -- > book1 <- bookPrior -- > book2 <- bookPrior -- > return (book1, book2) -- -- Nothing impressive. We pick a book with the prior -- we defined above, then another, pair them together -- and hand the pair back. What this will actually do -- is behave just like in the list monad, but in addition -- to this it will combine the probabilities of the various -- events we could be dealing with in the appropriate way. -- -- So, how about we verify what I just said: -- -- > λ> exact twoBooks -- > [ Event (Interesting,Interesting) 4.0% -- > , Event (Interesting,Boring) 16.0% -- > , Event (Boring,Interesting) 16.0% -- > , Event (Boring,Boring) 64.0% -- > ] -- -- Nice! Let's take a look at a more complicated scenario now. -- -- What if we wanted to take a look at the same distribution, -- with just a difference: we want at least one of the books to -- be an Interesting one. -- -- > oneInteresting :: Fin (Book, Book) -- > oneInteresting = bayes $ do -- notice the call to bayes -- > (b1, b2) <- twoBooks -- > condition (b1 == Interesting || b2 == Interesting) -- > return (b1, b2) -- -- We get two books from the previous distribution, and use 'condition' -- to restrict the current distribution to the values of b1 and b2 -- that verify our condition. This lifts us in the 'FinBayes' type, -- where our probabilistic computations can "fail" in some sense. -- If you want to discard values and restrict the ones on which you'll -- run further computations, use 'condition'. -- -- However, how do we view the distribution now, without having all -- those 'Maybe's in the middle? That's what 'bayes' is for. It runs -- the computations for the distribution and discards all the ones -- where any 'condition' wasn't satisfied. In particular, it means -- it hands you back a normal 'Fin' distribution. -- -- If we run this one: -- -- > λ> exact oneInteresting -- > [ Event (Interesting,Interesting) 11.1% -- > , Event (Interesting,Boring) 44.4% -- > , Event (Boring,Interesting) 44.4% -- > ] -- -- Note that these finite distribution types support random sampling too: -- -- * If one of your distributions has a type like "Finite d => d X", -- you can actually consider it as a 'RandT' value, from which you can sample. -- -- * If you have a 'Fin' distribution, you can use 'liftF' (lift 'Fin') -- to randomly sample an element from it, by more or less following -- the distribution's probabilities. -- -- > -- example of the former -- > sampleBooks :: RandT IO (V.Vector Book) -- > sampleBooks = vectorOf 10 bookPrior -- -- > λ> mwc sampleBooks -- > fromList [Interesting,Boring,Boring,Boring,Boring -- > ,Boring,Boring,Interesting,Boring,Boring] -- -- > λ> mwc $ listOf 4 (liftF oneInteresting) -- example of the latter -- > [ (Boring,Interesting) -- > , (Boring,Interesting) -- > , (Boring,Interesting) -- > , (Interesting,Boring) -- > ] module Math.Probable.Distribution.Finite ( -- * Probability type P(..), prob -- * 'Event' type , Event(..), never -- * 'EventT' monad transformer , EventT(..) -- * Finite distributions: 'Finite' and 'Fin' , Finite(..), Fin, exact, uniformly, liftF -- * Bayes' rule: 'FinBayes' , FinBayes, bayes, condition, onlyJust ) where #if !MIN_VERSION_base(4, 8, 0) import Control.Applicative #endif import Control.Monad import Control.Monad.Primitive import Control.Monad.Trans import Control.Monad.Trans.Maybe import Data.List import Math.Probable.Random -- | Probability type: wrapper around Double -- for a nicer Show instance and for more easily -- enforcing normalization of weights newtype P = P Double deriving (Eq, Ord, Fractional, Num, Real, RealFrac) -- | Get the underlying probability -- -- > λ> prob (P 0.1) -- > 0.1 prob :: P -> Double prob (P x) = x instance Show P where show (P p) = show intPart ++ "." ++ show fracPart ++ "%" where digits = round (1000 * p) intPart = digits `div` 10 :: Int fracPart = digits `mod` 10 :: Int -- | An event, and its probability data Event a = Event a {-# UNPACK #-} !P deriving (Eq, Show) -- | This event never happens (probability of 0) -- -- > never = Event undefined 0 never :: Event a never = Event undefined 0 instance Functor Event where fmap f (Event evt p) = Event (f evt) p instance Applicative Event where pure evt = Event evt 1 Event f p1 <*> Event e p2 | p1 == 0 || p2 == 0 = never | otherwise = Event (f e) (p1*p2) instance Monad Event where return evt = Event evt 1 (Event evt p) >>= f | p == 0 = never | otherwise = Event e' (p*p') where Event e' p' = f evt -- | 'EventT' monad transformer -- -- It pairs a value with a probability within the 'm' monad newtype EventT m a = EventT { runEventT :: m (Event a) } instance Monad m => Functor (EventT m) where fmap = liftM instance (Functor m, Monad m) => Applicative (EventT m) where pure = return mf <*> x = mf >>= \f -> fmap f x instance Monad m => Monad (EventT m) where return = lift . return m >>= f = EventT go where go = do ph <- runEventT m case ph of Event e p1 | p1 == 0 -> return never | otherwise -> do Event e' p2 <- runEventT (f e) return $ Event e' (p1 * p2) instance MonadTrans EventT where lift x = EventT (liftM return x) -- | Create a 'Finite' distribution over the values in -- the list, each with an equal probability -- -- > λ> exact $ uniformly [True, False] -- > [Event True 50.0%,Event False 50.0%] uniformly :: Finite d => [a] -> d a uniformly = weighted . map (,1) {-# INLINE uniformly #-} -- | 'Fin' is just 'EventT []' -- -- You can think of 'Fin a' meaning '[Event a]' -- i.e a list of the possible outcomes of type 'a' -- with their respective probability type Fin = EventT [] -- | See the outcomes of a finite distribution and their probabilities -- -- > λ> exact $ uniformly [True, False] -- > [Event True 50.0%,Event False 50.0%] -- -- > λ> data Fruit = Apple | Orange deriving (Eq, Show) -- > λ> exact $ uniformly [Apple, Orange] -- > [Event Apple 50.0%,Event Orange 50.0%] -- -- > λ> exact $ weighted [(Apple, 0.8), (Orange, 0.2)] -- > [Event Apple 80.0%,Event Orange 20.0%] exact :: Fin a -> [Event a] exact = runEventT {-# INLINE exact #-} -- | 'FinBayes' is 'Fin' with a 'MaybeT' layer -- -- What is that for? The 'MaybeT' lets us express -- the fact that what we've drawn from the distribution -- isn't of interest anymore, using 'condition', -- and observing the remaining cases, using 'bayes', -- to get back to a normal finite distribution. Example: -- -- > data Wine = Good | Bad deriving (Eq, Show) -- > -- > wines :: Finite d => d Wine -- > wines = weighted [(Good, 0.2), (Bad, 0.8)] -- > -- > twoWines :: Finite d => d (Wine, Wine) -- > twoWines = (,) <*> wines <$> wines -- > -- > decentMeal :: FinBayes (Wine, Wine) -- > decentMeal = do -- > (wine1, wine2) <- twoWines -- > -- we only consider the outcomes of 'twoWines' -- > -- where at least one of the two wines is good -- > -- because we're having a nice meal and are looking -- > -- for a decent pair of wine -- > condition (wine1 == Good || wine2 == Good) -- > return (wine1, wine2) -- > -- > -- to view the distribution, applying -- > -- Bayes' rule on our way: -- > exact (bayes decentMeal) type FinBayes = MaybeT Fin -- | This is the core of 'FinBayes'. If the 'Bool' is false, -- the current computation is shortcuited (sent to a 'Nothing' -- in 'MaybeT') and won't be included when running the distribution -- with 'bayes'. See the documentation of 'FinBayes' for an example. condition :: Bool -> FinBayes () condition = MaybeT . return . toMaybe where toMaybe True = Just () toMaybe False = Nothing -- | This functions discards all the elements of the distribution -- for which the call to 'condition' yielded 'Nothing'. -- While 'condition' does the mapping to 'Maybe' values, -- this function discards all of those values for which the condition -- was not met. bayes :: FinBayes a -> Fin a bayes = onlyJust . runMaybeT -- | Keeps only the 'Just's and remove the 'Maybe' layer -- in the distribution. onlyJust :: Fin (Maybe a) -> Fin a onlyJust dist | total > 0 = EventT (map adjust filtered) | otherwise = EventT [] where filtered = catMaybes' (runEventT dist) total = sum (map proba filtered) adjust (Event x p) = Event x (p / total) proba (Event _ p) = p -- value (Event x _) = x -- | This function, used by 'onlyJust', discards all the events -- holding a 'Nothing'. catMaybes' :: [Event (Maybe a)] -> [Event a] catMaybes' [] = [] catMaybes' (Event Nothing _ : xs) = catMaybes' xs catMaybes' (Event (Just x) p : xs) = Event x p : catMaybes' xs -- | T distribution of probabilities -- over a finite set. class (Functor d, Monad d) => Finite d where -- | The only requirement is to somehow -- be able to represent the distribution -- corresponding to the list given as argument, e.g: -- -- > weighted [(True, 0.8), (False, 0.2)] -- -- It should also be able to handle the normalization for you. -- -- > weighted [(True, 8), (False, 2)] weighted :: [(a, Double)] -> d a instance Finite Fin where weighted l = EventT $ map weight l where weight (x, w) = Event x $ P (w / total) total = foldl' (\w (_, w') -> w + w') 0 l instance PrimMonad m => Finite (RandT m) where weighted = liftF . weighted -- | Make finite distributions ('Fin') citizens of -- 'RandT' by simply sampling an element at random -- while still approximately preserving the distribution -- -- > λ> mwc . liftF $ uniformly [True, False] -- > False -- > λ> mwc . liftF $ uniformly [True, False] -- > True -- > λ> mwc . liftF $ weighted [("Haskell", 99), ("PHP", 1)] -- > "Haskell" liftF :: PrimMonad m => Fin a -> RandT m a liftF dist = do d <- double pick (P d) (exact dist) -- | Look for an 'Event' in that list that has -- a probability superior to the one given as an -- argument. pick :: Monad m => P -> [Event a] -> m a pick _ [] = error "pick: no value to pick" pick n (Event x p : evts) | n <= p = return x | otherwise = pick (n-p) evts instance Finite FinBayes where weighted = lift . weighted