Pipes, Clojure and Choco (part 3: Smart Search)

UPDATED: I found a bug in my code, which made the smart search less efficient, so I have re-written this post (more details below; at last I can delete the part that said “I do not understand”).

This is the third in a series of posts that explore constraint programming. It’s always more fun to learn with a good example, so we’re writing a solution for this problem.

In the previous article I described a “direct search”, but even there we were using constraints:

It’s true I didn’t explain this in detail (if you’re interested, valid-move? in isti.tubes.puzzle is the main test), but these constraints are so obvious it would have been “stupid” to ignore them.

How can we make this smarter? Norvig wrote an excellent article on constraints within Sudoku. Can we do something similar here?

Perhaps we can exploit something from knot theory? That would be interesting, but would require a complete rethink of how we describe the problem. A simpler alternative is to more carefully choose where we move.

Currently we add new tiles from the top left, raster scanning horizontally (next-empty in isti.tubes.direct). But it would make sense to place any tiles we are certain of first, so that we don’t waste time using them elsewhere. More generally, it seems like it would be best to place tiles on the most constrained cells first.

We can do this by tracking which tiles are available for each empty cell. Doing this efficiently is quite challenging. Originally I used an implementation with three separate maps, which avoided any scans of the data, but had very complex logic for updates. Here I will show a simpler approach with a single map from location to a set of masks.

There is one additional wrinkle: because we place moves in a “random” way we may generate the same pattern more than once (in one branch we could move T1 to C1 then T2 to C2; in another T2 to C2 then T1 to C1). This is because the moves no longer have a fixed, exclusive order1. We can avoid repeated moves by storing a hash of each tried position and testing against that - this must also go in the state.

OK, so on to the code. The state at each node will consist of the current puzzle, the map from location to possible masks (options), and the mutable set of hashes (visited, mutable so that we can always get the most complete set, even when lazy sequences make the evaluation order unclear):

(defn extract [[puzzle options visited]]

(defn solve [puzzle trace]
    [puzzle (initial-options puzzle) (atom #{})]
    (comp complete? extract)
    expand (atom 0) trace))

Expansion to child nodes (expand) uses the options map described earlier - sorted by the number of masks, then reduced to a sequence of locations.

(defn sort-by-constrained [options]
  (sort-by (comp count second) options))

(defn expand [[puzzle options visited]]
  (expand-masks puzzle options visited
    (first (sort-by-constrained options))))

where expand-masks do the mechanical work of generating all the possible moves (wrapped in lazy sequences so that we only pay for the children we use).

Embedded in that function, expand-masks, (see full source) is the logic for updating the state. We must do three things:

(defn update-options [puzzle options ij mask]
  (let [tiles (:tiles puzzle)
        cells (:cells puzzle)
        options (remove-location options ij)
        options (remove-last-mask options tiles mask)
        options (update-neighbours options cells ij puzzle)]

At this point I should explain my the reason for the “update” at the top of this article: originally the expansion into possible moves included all possible cells. If the most-constrained cell failed to work, others were tried in turn. But that was wrong! If the most-constrained cell has no consistent solution then our current solution is wrong and we can backtrack immediately. This makes the code much faster.

One more interesting snippet is an additional check: if we have a cell with no options (after the second of the three actions described just above) then we will be unable to find a solution, so we should prune the search tree at this point:

(defn check-options [puzzle options]
  (= (expt (:n puzzle) 2) (+ (count (:cells puzzle)) (count options))))

And, as I said above, the full source and the entire repo are available for you to study if curious.

How does it perform? Interestingly! Here are some typical solutions:

1       2       3       4       5       6       7       8       9       10   
▒═▒▒▒   ▒═▒▒▒   ╔═▒▒▒   ╔═▒▒▒   ╔═▒▒▒   ╔═▒▒▒   ╔═▒▒╗   ╔═▒▒╗   ╔═▒▒╗   ╔═▒▒╗
▒═▒▒▒   ▒═▒▒▒   ▒═▒▒▒   ▒═▒▒▒   ╠═▒▒▒   ╠═▒▒▒   ╠═▒▒▒   ╠═▒▒▒   ╠═▒▒▒   ╠═▒▒▒
▒╔▒╬▒   ▒╔▒╬▒   ▒╔▒╬▒   ║╔▒╬▒   ║╔▒╬▒   ║╔▒╬▒   ║╔▒╬▒   ║╔▒╬╝   ║╔═╬╝   ║╔╬╬╝
▒▒▒▒╗   ▒▒▒▒╗   ▒▒▒▒╗   ▒▒▒▒╗   ▒▒▒▒╗   ║▒▒▒╗   ║▒▒▒╗   ║▒▒▒╗   ║▒▒▒╗   ║▒▒▒╗
╚▒▒▒▒   ╚▒▒▒╝   ╚▒▒▒╝   ╚▒▒▒╝   ╚▒▒▒╝   ╚▒▒▒╝   ╚▒▒▒╝   ╚▒▒▒╝   ╚▒▒▒╝   ╚▒▒▒╝

11      12       13     14      15      16      17      18      19      20   
╔═▒▒╗   ╔═▒▒╗   ╔═▒▒╗   ╔═▒▒╗   ╔═▒▒╗   ╔═▒▒╗   ╔═▒╔╗   ╔═╗╔╗   ╔═╗╔╗   ╔═╗╔╗
╠═▒▒▒   ╠═▒▒▒   ╠═▒▒▒   ╠═▒▒▒   ╠═▒▒▒   ╠═▒▒▒   ╠═▒▒▒   ╠═▒▒▒   ╠═╣▒▒   ╠═╣▒║
║╔╬╬╝   ║╔╬╬╝   ║╔╬╬╝   ║╔╬╬╝   ║╔╬╬╝   ║╔╬╬╝   ║╔╬╬╝   ║╔╬╬╝   ║╔╬╬╝   ║╔╬╬╝
║▒▒╚╗   ║▒▒╚╗   ║▒▒╚╗   ║║▒╚╗   ║║║╚╗   ║║║╚╗   ║║║╚╗   ║║║╚╗   ║║║╚╗   ║║║╚╗
╚▒▒▒╝   ╚▒▒═╝   ╚▒╚═╝   ╚▒╚═╝   ╚▒╚═╝   ╚╝╚═╝   ╚╝╚═╝   ╚╝╚═╝   ╚╝╚═╝   ╚╝╚═╝


1         2        ...   16        17       ...   34       35    
▒▒▒▒▒▒	  ▒▒▒▒▒▒	 ╔╗▒╗╔╗	   ╔╗▒╗╔╗	  ╔╗▒╗╔╗   ╔╗╔╗╔╗
▒╠▒▒▒▒	  ▒╠▒▒▒▒	 ║╠╣▒▒║	   ║╠╣▒║║	  ║╠╣║║║   ║╠╣║║║
▒▒▒▒▒▒	  ▒▒▒▒▒▒	 ▒▒▒▒▒▒	   ▒▒▒▒▒▒	  ║║║║║║   ║║║║║║
▒▒▒▒▒▒	  ▒▒▒▒▒▒	 ▒▒▒▒▒▒	   ▒▒▒▒▒▒	  ║║║║║║   ║║║║║║
▒▒▒▒╣▒	  ▒▒▒▒╣║	 ▒▒▒▒╣║	   ▒▒▒▒╣║	  ║║║╠╣║   ║║║╠╣║
▒▒▒▒▒▒	  ▒▒▒▒▒▒	 ╚╝╚╝╚╝	   ╚╝╚╝╚╝	  ╚╝╚╝╚╝   ╚╝╚╝╚╝

The direct search in the previous article required around 40 moves to solve the first problem. The new, smarter approach uses half that, and only mis-steps slightly on step 9 (if you read the earlier version of this post then you may remember that this was the point where the search becaome “lost”; that no longer happens when we discard alternative cells).

For the second example, the new strategy is even better. The number of iterations is reduced by 90% (to the minimum) because the clue tiles provide strong constraints - every move is fixed.

Unfortunately, the extra calculations require more time per step. Here are the timing results:

examples               concentric                     per iteration
         5x5          6x6       8x8           10x10            average
smart    2ms/21       3ms/35    900ms/9,000   40s/300,000      100us
direct   0.2ms/40     2ms/300   200ms/30,000  10s/1,000,000    5-10us

This is a little disappointing. Even in the “perfect” second example, we are no faster than the simpler search. And the overhead continues to outweigh the advantage from the reduced number of steps on the concentric problems.

Again: you can see all the code in the repo.

I mentioned earlier that I had a more complex version of this code. That took 7ms for the 5x5 example and 16ms for 6x6, so . Next article I’ll show how to use Choco - will it do the impossible and combine smarts with speed?


this (please comment if you do).

  1. Sorry, but I don’t have a good, clear theoretical explanation for

Related Posts

blog comments powered by