Preliminary remarks
I am learning Haskell.
A question that I answered some days ago gave me the inspiration for this exercise in Haskell, which gave the opportunity for experimenting with the few things that I've learned up to now, and also left me with questions :)
Problem statement
Given a rectangle A of width w
and height h
find the best rectangle B that fits n
times within A, where best means having the smallest perimeter.
My attempt
I've started with the basic idea of generating the set of sub-rectangles of A having an area equal to div (w * h) n
, and then picking the one having the smallest perimeter.
Here are the three implementations of this idea that I came up with; they're in chronological order: I got the inspiration for the third after having done the second, which I got after having done the first (OK, there's a version 0, in which I didn't use data Rectangle
but just a tuple (x, y)
):
Implementation 1
data Rectangle = Rectangle { width :: Integer,
height :: Integer
} deriving (Show)
subRectangles :: Rectangle -> Integer -> [ Rectangle ]
subRectangles r n = [ Rectangle x y | x <- [1..w ], y <- [1..h], x * y == (w * h) `div` n ]
where w = width r
h = height r
bestSubRectangle :: [ Rectangle ] -> Rectangle
bestSubRectangle [ r ] = r
bestSubRectangle (r:rs)
| perimeter r < perimeter bestOfRest = r
| otherwise = bestOfRest
where bestOfRest = bestSubRectangle rs
perimeter :: Rectangle -> Integer
perimeter r = (width r) + (height r)
Implementation 2
data Rectangle = Rectangle { width :: Integer,
height :: Integer
} deriving (Show)
subRectangles :: Rectangle -> Integer -> [ Rectangle ]
subRectangles r n = [ Rectangle x y | x <- [1..w ], y <- [1..h], x * y == (w * h) `div` n ]
where w = width r
h = height r
bestSubRectangle :: [ Rectangle ] -> Rectangle
bestSubRectangle xs = foldr smaller (last xs) xs
smaller :: Rectangle -> Rectangle -&g开发者_如何学编程t; Rectangle
smaller r1 r2
| perimeter r1 < perimeter r2 = r1
| otherwise = r2
perimeter :: Rectangle -> Integer
perimeter r = (width r) + (height r)
Implementation 3
import Data.List
data Rectangle = Rectangle { width :: Integer,
height :: Integer
} deriving (Show, Eq)
instance Ord Rectangle where
(Rectangle w1 h1) `compare` (Rectangle w2 h2) = (w1 + h1) `compare` (w2 + h2)
subRectangles :: Rectangle -> Integer -> [ Rectangle ]
subRectangles r n = [ Rectangle x y | x <- [1..w ], y <- [1..h], x * y == (w * h) `div` n ]
where w = width r
h = height r
bestSubRectangle :: [ Rectangle ] -> Rectangle
bestSubRectangle = head . sort
Questions
Which approach is more idiomatic?
Which approach is better in terms of performance?
bestSubRectangle
in Implementation 3 depends onsort
, which is at best O(n lg n), while in Implementation 1, 2bestSubRectangle
requires only scanning the array returned bysubRectangles
, thus making it O(n). However I'm not sure if/how Haskell laziness works onbestSubRectangle = head . sort
: willsort
produce only the first element of the sorted array, because ofhead
requiring only the first element (head (x:_) = x)
?In implementation 3, when making
Rectangle
an instance ofOrd
should I also define the other methods of theOrd
class? Is this the right way to makeRectangle
an instance ofOrd
?
Any further suggestion/recommendation to improve is highly welcome.
To answer your questions about Haskell (and not about the algorithm you've choosen):
- I'd say your second implementation is most idiomatic.
- You are right that the problem needs only a linear scan, so that
sort
is likely to be more expensive than needed. You are also asking the right question when asking if, due to laziness,head . sort
will compute only the first element of the result of thesort
. It will, but depending on howsort
is implemented, that may very well rely on sorting the whole list before it can return the first element. You should assume that it does. - You can tell from the documentation for
Ord
thatcompare
is sufficient. The key phrase is Minimal complete definition: eithercompare
or<=
. Many type classes have similar use patterns. You should feel free to write a minimal implementation where they do.
Some other observations about your code (again, not the algorithm):
- Function calls bind tighter than operators, as in most languages - just in Haskell there are no parenthesis around the arguments. Hence you would write
perimeter r = width r + height r
- If you are folding over a list that must have at least one element, you can use
foldr1
as inbestSubRectangle xs = foldr1 smaller xs
- Your instance of
Ord
forRectangle
doesn't agree with the derived instance ofEq
. That is, there are values which willcompare
asEQ
but would returnFalse
for==
. In this case, I wouldn't try to bendRectangle
's general purpose instance ofOrd
for purposes of perimeter comparison, and instead go with thesmaller
predicate you wrote in the second implementation.
Seems like you are calculating too much, inductively going through posibilities of n (number of rectangles we wish to fill up our given rectangle) we should get:
- n = 1 => always returns the given rectangle
- n = 2 => always returns the 2 rectangles given by bisecting the given rectangle on the longest side, i.e. gives the 2 most square rectangles
- n = 3 => same as 2 using 3, divide equally along longest side.
- n = 4 More complicated, essentially the question comes up is it better to put the same rectangle 4 times in a row OR is is better to bisect both length and width into a 2x2 group of rectangle. For larger numbers of n it is also a question of these configuration of factors > 1
--Essentially this is a question of factoring n, dividing the length and width by the each of the factors THEN choosing the configuration where the divided length & width (i.e. the length and width of the resulting filler rectangles) are MOST SIMILAR, (most square-like). Also note it is not necessary to try all factors against all sides. The most square rectangle will be taking the larger factor against the longer side.
- n = number of filler rectangles
- l = longest of width or height of container rectangle
- s = shortest of width or height container rectangle
- h1 = height of candidate rectangle 1
- w1 = width of candidate rectangle 1
- d = difference of candidate dimensions
- t = smallest difference in candidate height and width (initialize t to l)
- b = best fit
Therefore the steps become:
1: Factor n
2: for each factor pair of n:
(l/largest factor) - (s/smaller factor) = d
if d < t:
t = d
store current best candidate rectangle
3: Return best fit remaining after all factors have been tried
精彩评论