For Colin and Serg,
The code that I used, was a modified version of my code for the 17-clue search.
A complete explanation would be long and complicated, but the basic idea, involves puzzles like this, in the first stage:
- Code: Select all
+-------+-------+-------+
| . . . | . . . | x x x |
| . . . | . . . | x x x |
| . . . | . . . | x x x |
+-------+-------+-------+
| | . . . |
| <= 7 clues | . . . |
| | . . . |
+-------+-------+-------+
| g g g | g g g | g g g |
| g g g | g g g | g g g |
| g g g | g g g | g g g |
+-------+-------+-------+
The 9c (or 5c) box, is the box with x's, and really it can be any of the 6 boxes in bands 1 & 2.
Band 3 is a band fill with one of the 44 "gangster (mini-column) signatures", and bands 1 & 2 in the puzzles' solution grids, are minlex with respect to transformations that preserve the gangster signature in band 3. There are 983,959,110 (ED) cases like that ... "gangster signature" + "compatible bands 1&2 fill".
For the first part of the search, band 3 can be any of the bands that have the required gangster signatures, and puzzles like the one above, are identified.
For the next part, there's a loop over the compatible band 3 fills, and an attempt to clear the band 3 box that's under the box with the x's, and then reduce the clue count in the other boxes, to be small enough to make a 9plus14 (or less) puzzle, or for the 5+15 search, to make a 5plus15.
I'm leaving a bunch of details out ... like having tables for each canonical band type, listing thier "5c one box puzzles", "7c <= two box puzzles", and so on.
and translating those to match the particular band 1 & band 2 fills.
Both problems (9+14 and 5+15) have a diagonal reflection symmetry, that I didn't exploit ... so I ended up finding each puzzle twice, basically.
The 9+14 search took 36 hours, and the 5+15 search took 45 hours, on a machine with 8 (physical) cores, running at 4.5 GHz.
Is this enough of an anwer ?