Unique sample drawing & searches with List and StateT — “Send more money”

SourceMarkdownLaTeXPosted in Haskell, TutorialsComments


Nothing too crazy today, just a cute (basic/intermediate) haskell trick as a response to Mark Dominus’s excellent Universe of Discourse post on Easy exhaustive search with the list monad intended for people new or unfamiliar with haskell demonstrating the common “list monad as a constraint solver” approach that is standard fare for learning Haskell. I myself have literally done an entire series of blog posts on this usage.

Mark’s use case however incorporates a bit of an extra pattern not typically discussed. The list monad is good for taking “independent samples” of things (looking at different samples from a list):

ghci> do x <- "abc"
         y <- "abc"
         z <- "abc"
         return [x,y,z]
["aaa","aab","aac","aba","abb" ... ]

However, what if you wanted to instead “draw” from a pool, and represent different drawings? Traditionally, the answer was something like:

ghci> do x <- "abc"
         y <- filter (/= x) "abc"
         z <- filter (/= y) . filter (/= x) $ "abc"
         return [x,y,z]
"abc","acb","bac","bca","cab","cba"]

This is a little bit awkward…and it definitely gets a lot worse (O(n^2)) when you have more items. Also, it relies on an Eq constraint — what if our thing doesn’t have an Eq instance? And this also falls apart when our list contains duplicate items. If we had used "aabc" instead of "abc", the result would be the same — despite having more 'a's to pick from!

Important note: After writing this article, I found out that Douglas Auclair in 11th issue of the Monad Reader solved this exact same problem with pretty much the exact same approach. (Oops!) If you want to do further reading, check it out! :D

StateT

There’s a type in the transformers library that provides a very useful monad instance:

data StateT s m a = StateT (s -> m (a, s))

A StateT s m a is a function that takes an initial state s and returns a result a with a modified state…in the context of m.

Specialize for m ~ [] and we get

data StateT s [] a = StateT (s -> [(a, s)])

Which is basically describing a function from a initial state to a list of ways you can modify the state, and different results from each one. It returns a list of “all ways you can mutate this state”.

For example,

foo :: StateT Int [] Bool
foo = StateT $ \x -> [(even x, x+1), (odd x, x-1), (x > 0, negate x)]

So foo takes a number, x, and says, “here are three ways we might proceed from having this number. We can return whether or not it’s even, in which case the new state is x+1…we can return whether or not it’s odd, in which case the new state is x-1….or we can return whether or not it’s positive, in which case the new state is negate x

What the monad instance does is that it allows you to “chain” forks, and go along different forks, and gather together “all possible forks” you could have taken. At the end, it outputs all possible forks. So if you did foo >> foo, there’d be nine results — one result for when you took the first route (the x+1) twice, one for when you took the first and then the second (x-1), one for when you took the first and the third….and the second and the first…etc., etc.

MonadPlus

One other tool we have at our disposal is guard:

guard :: Bool -> StateT Int [] ()

which is a StateT action that says “kill this current branch if given False, or go on if given True

The Problem

The problem, as stated, was to find distinct digits for each letter to solve:

    S E N D
+   M O R E
-----------
  M O N E Y

So SEND is a four-digit number, MORE is a four-digit number, and MONEY is a five-digit number that is the sum of the two. The first digit of MONEY has to be the first digit of MORE, the last digit of MORE has to be the second digit of SEND, etc.

The previous approach was done using the entire “pick from all possibilities…except for the ones already chosen”, using (/=) and filtering over all of the things seen vs all of the things to pick from.

However, we can abstract over “picking dependently from a sample” by defining a function called select, which really should be in the base libraries but isn’t for some reason:

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/misc/send-more-money.hs#L7-9
select :: [a] -> [(a, [a])]
select []     = []
select (x:xs) = (x,xs) : [(y,x:ys) | (y,ys) <- select xs]

(Implementation thanks to Cale, who has fought valiantly yet fruitlessly to get this into base for many years.)

select will take a list [a] and return a list of different “selected” as, with the rest of the list, too:

ghci> select "abcd"
[('a',"bcd"),('b',"acd"),('c',"abd"),('d',"abc")]

But, hey…does the type signature of select look like anything familiar?

It looks exactly like something that StateT is supposed to describe! Give an initial state ([a]), and returns a list of all possible ways to “mutate” that state (by removing one element from the state), and a result from each mutation (the removed element).

StateT select :: StateT [a] [] a

And armed with this…we have all we need

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/misc/send-more-money.hs#L3-35
import Control.Monad             (guard, mfilter)
import Control.Monad.Trans.State
import Data.List                 (foldl')

asNumber :: [Int] -> Int
asNumber = foldl' (\t o -> t*10 + o) 0

main :: IO ()
main = print . flip evalStateT [0..9] $ do
    s <- StateT select
    e <- StateT select
    n <- StateT select
    d <- StateT select
    m <- StateT select
    o <- StateT select
    r <- StateT select
    y <- StateT select
    guard $ s /= 0 && m /= 0
    let send  = asNumber [s,e,n,d]
        more  = asNumber [m,o,r,e]
        money = asNumber [m,o,n,e,y]
    guard $ send + more == money
    return (send, more, money)

Remember, StateT here operates with an underlying state of [Int], a list of numbers not yet picked. StateT select picks one of these numbers, and modifies the state to now only include the items that were not picked. So every time you sequence StateT select, select draws from a smaller and smaller pool of numbers, and makes the state list smaller and smaller. What sequencing StateT does is allow us to explore all of the possible ways we could pick and modify state, all at once. Using guard, we then “close off” and kill off the paths that don’t end up how we’d like.

asNumber takes a list like [1,2,3] and turns it into the number 123; credit to the source blog.

And, to test it out…

$ ghc -O2 send-more-money.hs
$ ./send-more-money
# [(9567,1085,10652)]

It returns the one and only solution, SEND = 9567, MORE = 1085, and MONEY = 10652.1

Aside

We can make things a little bit more efficient with minimal cost in expressiveness. But not that it matters…the original version runs fast already.

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/misc/send-more-money.hs#L38-59
select' :: [a] -> [(a,[a])]
select' = go []
  where
   go xs [] = []
   go xs (y:ys) = (y,xs++ys) : go (y:xs) ys

main' :: IO ()
main' = print . flip evalStateT [0..9] $ do
    s <- mfilter (/= 0) $ StateT select'
    m <- mfilter (/= 0) $ StateT select'
    e <- StateT select'
    n <- StateT select'
    d <- StateT select'
    o <- StateT select'
    r <- StateT select'
    y <- StateT select'
    let send  = asNumber [s,e,n,d]
        more  = asNumber [m,o,r,e]
        money = asNumber [m,o,n,e,y]
    guard $ send + more == money
    return (send, more, money)

This is a more performant version of select courtesy of Simon Marlow that doesn’t preserve the order of the “rest of the elements”.

Also, we use mfilter to “eliminate bad s and ms” right off the bat, before having to pick any more things. mfilter can be thought of as “killing the fork immediately” if the action doesn’t satisfy the predicate. If the s picked doesn’t match (/= 0), then the entire branch/fork is immediately ruled invalid.

By the way, isn’t it neat that it does all of this in “constant space”? It just keeps track of the output list, but the actual search processes is in constant space. You don’t need to keep track of all 10! combinations in memory at once. Hooray laziness!

Other Applications

Using select and StateT, we can do a lot of things involving picking from a sample, or permutations. Anything that you used to awkwardly do by using filter not-equal-to’s can work now. You can do things like drawing from a deck:

pokerGame :: [Ordering]
pokerGame = flip evalStateT [0..51] $ do
    p2Hand <- replicateM 5 (StateT select)
    p1Hand <- replicateM 5 (StateT select)
    return $ pokerCompare p1Hand p2Hand

Which would draw five distinct cards from a deck of [0..51], and return who won for each draw (assuming you had a suitable pokerCompare :: [Card] -> [Card] -> Ordering). Note that if you use runStateT, you’d get the results (the winner), as well as the leftover cards in the deck for each path!

You can even combine the two sorts of drawings — sampling independently (like rolling dice) using lift, and drawing from an underlying deck. For example, you might encode some game logic from a board game like monopoly:

combo = flip evalStateT initialDeck $ do
    roll <- lift [1..6]
    draw <- StateT select
    ...

Whenever you want a dice roll, use lift [1..6]…and whenever you want to draw from the deck, use StateT select.

What you get in the end, remember, is a list of “all possible paths”. You’ll get a list of every possible result from all of your different rolling and drawing choices.

Happy Haskelling!


  1. For some reason this runs pretty slowly if you use runghc/runHaskell, but it runs in the blink of an eye when you actually compile it (and especially with optimizations on). The difference is pretty striking…and I don’t really know what’s going on here, to be honest. If anyone does know a good explanation, I’d love to hear it :)

Comments powered by Disqus