Top 10000 Sudokus

Ruud · Site Admin

There is a new collection that you can use to test your high-end solvers. The Top 10000 only contains Sudokus which SudoCue can only solve with tabling.

The failing techniques are:
- Singles
- Locked candidates
- Naked/Hidden subsets
- Fish, Finned Fish, Sashimi Fish
- Skyscraper, 2-string kite, Empty Rectangle
- Simple colors, Multi-colors
- Remote pairs
- XY-Wing, XYZ-Wing
- APE
- Sue-de-coq
- ALS-xz rule
- XY-chains
- 3D coloring w. weak interactions (AIC w/o embedded patterns)
- UR types 1-4
- BUG+1
- Nishio (single-digit patterns)

As usual, my puzzle collections are free for personal use.
_________________
Meet me at sudocue.net

whizter · Posted: Mon Aug 06, 2007 10:36 pm Post subject:

My solver can solve 69 of these, and it only uses techniques that you mentioned.

gsf · Posted: Tue Aug 07, 2007 4:49 am Post subject:

with my solver:
5321 are solved with basic methods and XY cycles
another 4452 are solved with Y-knots (failed Y cycles aka contradiction chains using singles)
and 32 more are solved with rookery overlays (templates)
the remaining 195 require backtracking

a r n e · Posted: Thu Mar 27, 2008 8:48 pm Post subject:

Hi.

The SuDoKu-Cracker only tries to solve as fast as posible. Your Top 10000 are solved in 31 sec (30.907).

arne.

gsf · Posted: Thu Mar 27, 2008 9:18 pm Post subject:

a r n e · Posted: Fri Mar 28, 2008 8:42 am Post subject:

Thanks, you gave me an idea.

My Solver works fastest with full featured solver limited to singles + backtracking.

On an AMD Sempron; 548 MHz; Windows XP
It takes 10 sec (9.890) to solve this Top 10000
On a mixed (Easy to Expert) database, it takes 2.6 sec to solve 10000 puzzles.

arne.

gsf · Posted: Fri Mar 28, 2008 2:04 pm Post subject:

Sisophon2001 · Posted: Sun Mar 30, 2008 8:25 am Post subject:

I am new to this level of Sudoku, although I have made a puzzle solver, and generator for the more basic methods. I am working at the moment on expanding my understanding of solving techniques.

Using the development code I am working on, my results more or less match Ruud's, and perhaps the others who responded. I did not implement tabling, and use forcing nets at the last stop before using backtracking.

1041 puzzles need backtracking and the rest are solvable only with forcing nets. I must assume (if my code does not have any bugs) that Ruud's implementation of tabling goes more than one step deep, so that it can investigate cases where there is more than one chain present. Forcing nets mirrors what I would do myself is stuck with a difficult puzzle, so I think this is useful to implement as it provides some information useful to the user.

Looking at whizter results I would think that his/hers implementation of 3D Medusa is picking up even more than my forcing net. 3DMedusa looks like it should covers almost everything, from what I read. What do others think?

gsf is using techniques not documented in the sudopedia - I checked his web site, and it would take much more time that I gave it to absorb the information provided - so I am not sure if these techniques are easy to learn for human solvers. If they are human solving techniques, then they look worth learning, judging by his results.

These are the techniques I currently implement,
- Singles
- Locked candidates
- Naked/Hidden subsets
- Fish, Finned Fish, Sashimi Fish
- XY-Wing, XYZ-Wing
- Sue-de-coq
- ALS-XZ rule ( in progress - still buggy so not used in tests)
- Forcing Chains
- XY-chains
- BUG+1 & BUG+2
- Nishio
- Forcing Net

Bifurcation is use internally for analysis, but not used as a solving technique.

I am printing the results to console, and redirecting to file, so this is not the best way to test for speed. My code with all test methods enabled is slow enough that I can have a shower while waiting for the results of 10,000 tests. I don't want to time it. I will worry about testing for speed when I understand the solving techniques.

I enjoyed the tests, and reading the replies.

Garvan