awkward monad transformer stack - haskell

Question

Solving a problem from Google Code Jam (2009.1A.A: "Multi-base happiness") I came up with an awkward (code-wise) solution, and I'm interested in how it could be improved.
The problem description, shortly, is: Find the smallest number bigger than 1 for which iteratively calculating the sum of squares of digits reaches 1, for all bases from a given list.
Or description in pseudo-Haskell (code that would solve it if elem could always work for infinite lists):
solution =
head . (`filter` [2..]) .
all ((1 `elem`) . (`iterate` i) . sumSquareOfDigitsInBase)
And my awkward solution:
By awkward I mean it has this kind of code: happy <- lift . lift . lift $ isHappy Set.empty base cur
I memoize results of the isHappy function. Using the State monad for the memoized results Map.
Trying to find the first solution, I did not use head and filter (like the pseudo-haskell above does), because the computation isn't pure (changes state). So I iterated by using StateT with a counter, and a MaybeT to terminate the computation when condition holds.
Already inside a MaybeT (StateT a (State b)), if the condition doesn't hold for one base, there is no need to check the other ones, so I have another MaybeT in the stack for that.
Code:
import Control.Monad.Maybe
import Control.Monad.State
import Data.Maybe
import qualified Data.Map as Map
import qualified Data.Set as Set
type IsHappyMemo = State (Map.Map (Integer, Integer) Bool)
isHappy :: Set.Set Integer -> Integer -> Integer -> IsHappyMemo Bool
isHappy _ _ 1 = return True
isHappy path base num = do
memo <- get
case Map.lookup (base, num) memo of
Just r -> return r
Nothing -> do
r <- calc
when (num < 1000) . modify $ Map.insert (base, num) r
return r
where
calc
| num `Set.member` path = return False
| otherwise = isHappy (Set.insert num path) base nxt
nxt =
sum . map ((^ (2::Int)) . (`mod` base)) .
takeWhile (not . (== 0)) . iterate (`div` base) $ num
solve1 :: [Integer] -> IsHappyMemo Integer
solve1 bases =
fmap snd .
(`runStateT` 2) .
runMaybeT .
forever $ do
(`when` mzero) . isJust =<<
runMaybeT (mapM_ f bases)
lift $ modify (+ 1)
where
f base = do
cur <- lift . lift $ get
happy <- lift . lift . lift $ isHappy Set.empty base cur
unless happy mzero
solve :: [String] -> String
solve =
concat .
(`evalState` Map.empty) .
mapM f .
zip [1 :: Integer ..]
where
f (idx, prob) = do
s <- solve1 . map read . words $ prob
return $ "Case #" ++ show idx ++ ": " ++ show s ++ "\n"
main :: IO ()
main =
getContents >>=
putStr . solve . tail . lines
Other contestants using Haskell did have nicer solutions, but solved the problem differently. My question is about small iterative improvements to my code.

Answer 1

Your solution is certainly awkward in its use (and abuse) of monads:
It is usual to build monads piecemeal by stacking several transformers
It is less usual, but still happens sometimes, to stack several states
It is very unusual to stack several Maybe transformers
It is even more unusual to use MaybeT to interrupt a loop
Your code is a bit too pointless :
(`when` mzero) . isJust =<<
runMaybeT (mapM_ f bases)
instead of the easier to read
let isHappy = isJust $ runMaybeT (mapM_ f bases)
when isHappy mzero
Focusing now on function solve1, let us simplify it.
An easy way to do so is to remove the inner MaybeT monad. Instead of a forever loop which breaks when a happy number is found, you can go the other way around and recurse only if the
number is not happy.
Moreover, you don't really need the State monad either, do you ? One can always replace the state with an explicit argument.
Applying these ideas solve1 now looks much better:
solve1 :: [Integer] -> IsHappyMemo Integer
solve1 bases = go 2 where
go i = do happyBases <- mapM (\b -> isHappy Set.empty b i) bases
if and happyBases
then return i
else go (i+1)
I would be more han happy with that code.
The rest of your solution is fine.
One thing that bothers me is that you throw away the memo cache for every subproblem. Is there a reason for that?
solve :: [String] -> String
solve =
concat .
(`evalState` Map.empty) .
mapM f .
zip [1 :: Integer ..]
where
f (idx, prob) = do
s <- solve1 . map read . words $ prob
return $ "Case #" ++ show idx ++ ": " ++ show s ++ "\n"
Wouldn't your solution be more efficient if you reused it instead ?
solve :: [String] -> String
solve cases = (`evalState` Map.empty) $ do
solutions <- mapM f (zip [1 :: Integer ..] cases)
return (unlines solutions)
where
f (idx, prob) = do
s <- solve1 . map read . words $ prob
return $ "Case #" ++ show idx ++ ": " ++ show s

Answer 2

The Monad* classes exist to remove the need for repeated lifting. If you change your signatures like this:
type IsHappyMemo = Map.Map (Integer, Integer) Bool
isHappy :: MonadState IsHappyMemo m => Set.Set Integer -> Integer -> Integer -> m Bool
This way you can remove most of the 'lift's. However, the longest sequence of lifts cannot be removed, since it is a State monad inside a StateT, so using the MonadState type class will give you the outer StateT, where you need tot get to the inner State. You could wrap your State monad in a newtype and make a MonadHappy class, similar to the existing monad classes.

Answer 3

ListT (from the List package) does a much nicer job than MaybeT in stopping the calculation when necessary.
solve1 :: [Integer] -> IsHappyMemo Integer
solve1 bases = do
Cons result _ <- runList . filterL cond $ fromList [2..]
return result
where
cond num = andL . mapL (isHappy Set.empty num) $ fromList bases
Some elaboration on how this works:
Had we used a regular list the code would had looked like this:
solve1 bases = do
result:_ <- filterM cond [2..]
return result
where
cond num = fmap and . mapM (isHappy Set.empty num) bases
This calculation happens in a State monad, but if we'd like to get the resulting state, we'd have a problem, because filterM runs the monadic predicate it gets for every element of [2..], an infinite list.
With the monadic list, filterL cond (fromList [2..]) represents a list that we can access one item at a time as a monadic action, so our monadic predicate cond isn't actually executed (and affecting the state) unless we consume the corresponding list items.
Similarly, implementing cond using andL makes us not calculate and update the state if we already got a False result from one of the isHappy Set.empty num calculations.