4

I'M looking for an algorithm (preferably polynomial) for the following problem:

Input:
- an M x N grid, where each square can be either empty or occupied
- a list of items
- a list of constraints

Output:
the grid, such that all items are placed and the constraints aren't violated.

Each item is a 1d block of length k. Different items may have different length, and there may be several items with the same length. Items can't ovelap each other.

Example 1: Given 1 x 5 grid and 2 items of lengths 1, 2 respectively. The following solutions are acceptable:

11100

10110

11010

and so forth.

Example 2: Given 3 x 5 grid and 2 items of lengths 1, 2 respectively. The following solutions are acceptable:

00110
10000
00000

11000
10000
00000

etc.

Possible constraints:
a. Isolation - items may not touch each other. e.g. the following solutions is illegal:

Example 1:

11100

Example 2:

11000
10000
00000

However this is a legal solution:

01100
10000
00000

i.e. touching diagonals are legal.

b. Facing - 2 items must be placed one in front of the other. e.g.

11000
00000
11000

What I did so far is to find a solution for a simpler problem, where the grid is 1d and the only possible constraint is Isolation. I solved it by representing it as a search problem, and using A* with the heuristic: sum of lengths of items not placed yet. The problem is that it suffers from state explosion which will only deteriorate in higher dimensions.

dimid
  • 141
  • 3
  • This smells NP-hard. A reduction from partition could work, but I don't have time atm to work out the details. (I'd try with a $2 \times\frac m2$-grid for elements with a weight sum of $m$. Some adjustment might be necessary to forbid placing elements with size 2 in the wrong orientation.) – FrankW May 04 '14 at 21:42

1 Answers1

1

The 1-dimensional case (a $1 \times n$ grid) is straightforward. I'm not sure why you are having difficulty finding solutions when the grid is 1-dimensional. If you don't have any constraints, then you just add up the lengths of the items; if their total length is $\le n$, you can solve it (just put the items next to each other), otherwise you can't. To deal with the isolation constraint, you can note that the problem with tiles lengths $\ell_1,\dots,\ell_k$ and an isolation constraint is solvable if and only if the problem with tile lengths $\ell_1+1,\dots,\ell_{k-1}+1,\ell_k$ and no isolation constraint is solvable.

The 2-dimensional case (a $m \times n$ grid) looks much harder. One approach, in practice, might be to use a SAT solver. Introduce boolean variables $x_{i,j,k,d}$ and $y_{i,j}$, with the intent that $x_{i,j,k,d}=1$ if the $k$th tile starts at location $(i,j)$ on the grid and proceeds in direction $d$, and $y_{i,j}=1$ if location $(i,j)$ is covered by some tile. Now you can formulate your problem as a set of boolean constraints on the $x$'s and $y$'s, then feed it to an off-the-shelf SAT solver. You could experiment with this for the parameter sizes that tend to crop up for you in practice and see how well it works. SAT solvers often work surprisingly well, though there is no guarantee: they are not a silver bullet, and the worst-case running time will remain exponential. There are many other ways to represent this as a SAT problem, too, so you could try other representations as well to see which tends to work best with existing SAT solvers.

I'm not sure you're going to be able to do much better than this: as FrankW says, for 2-dimensional grids this smells like it might be NP-complete in the general case.

D.W.
  • 159,275
  • 20
  • 227
  • 470
  • can you provide a quick illustration for the 2d formulation to the SAT solver please? – evandrix Dec 29 '20 at 05:47
  • @evandrix, I'll let you work through an example yourself and see if you have any specific questions as you go through that exercise. There are lots of materials and tutorials on SAT solvers available on the Internet, if that's what you are looking for. – D.W. Dec 29 '20 at 19:52
  • my question was motivated by me looking at the answer here, which proposes the same representation as you (so perhaps this is THE standard transformation for such 2d problems to something a SAT solver can easily consume), and not understanding how to extend the step "Then, write clauses that prohibit two crossing segments. If segment 1 and segment 2 are both length 2, then we add the clause (!L_seg1_0_0_right || !L_seg2_1_0_right)" for >2 line segments, is it going to be N choose 2 i.e. pairwise, for N line segments? – evandrix Dec 29 '20 at 21:12
  • if so, won't that be many separate constraints, would it choke the SAT solver, and produce non-viable runtime? – evandrix Dec 29 '20 at 21:14
  • which is why i requested for a simple example, so that i can compare what you are proposing, vs what https://stackoverflow.com/questions/12099244#12099815 is proposing – evandrix Dec 29 '20 at 21:15
  • the problem i have is arguably simpler - like https://stackoverflow.com/questions/12099244#12099815, but line segments can only be horizontal, plus the constraint stated in this question "a. Isolation - items may not touch each other", and i also need to output a valid assignment, not just a YES/NO result – evandrix Dec 29 '20 at 21:18
  • there aren't as many examples as you might think (?) or i'm just dull, but the only other real thing of some help i found was like https://stackoverflow.com/questions/47918792 – evandrix Dec 29 '20 at 21:19
  • did you have some particular resource/link online in mind when you wrote "lots of materials and tutorials"? – evandrix Dec 29 '20 at 21:20
  • re:my earlier comment above, for some reason i can't edit it anymore, https://stackoverflow.com/questions/12099244#12099815 writes: "There is a pretty natural mapping." – evandrix Dec 29 '20 at 21:21
  • @evandrix, if you want to ask about that answer, I suggest trying to work out the details as much as you can, then use the 'Ask Question' button to ask a new question specifically about it. This comment thread is probably not the right place to understand what is going on in that answer. You might start with https://cs.stackexchange.com/q/12087/755 and https://cs.stackexchange.com/q/30790/755 and try Googling. – D.W. Dec 29 '20 at 23:51
  • i did google, a lot, finding e.g. http://www.hakank.org/google_or_tools – evandrix Dec 30 '20 at 00:26