The New Sudoku Players' Forum

[coloin]

Well.....
with my understanding of the process - its not exactly clear if splitting the search into 3 will save any time,
however the search for the 222222221 puzzles must surely be quick.

I made a very determined effort to find 17-puzzles which have 4 clues in a box.
I generated several [ over 3 million ] 22 clue puzzles with a full box [ that would be 9 clues plus 13 others]
By removing clues I found 90 % of the known 17-puzzle with 4 clues. I stopped the program - because it would never be exhaustive....

It is also not proven that band completions with 5 or more clues in a box are not worthwhile to test, but I suspect i have found almost all of the 9 clues plus 12 others
I found only 152 puzzles - the six 5plus12s were in there.
and of course no puzzle with 9 clues plus 11 others was found.

[coloin]

As ever the breadth of the sudoku puzzle space scuppers many ideas.

I had hoped that it would have been possible to generate almost all those '9plus13' puzzles, and that my generation methods were sound.
It seems that its very easy to rattle around the non-remote puzzles .....
With decreasing puzzles found - each time i had hoped that I had reached the tail of the puzzles at 3.5M

but thanks to blues's help

Using "random grid, random box choice" sampling ... 2,512,000 samples ...
I got an estimate of (18.1 +/- 1.4) million 9+13's.

which explains a lot

and for completeness the number of 9plus14s were estimated as well ....

From 1000 grids, 936 of them had at least one 9+14.
From 100, 93 of them did, for a total of 3033 puzzles.
Extrapolating from the samples, gave:
(5.122 +/- 0.083) billion grids with a 9+14.
(166 +/- 40) billion 9+14's total.

anyway ... too many 9plus14s ,,,, and no way to get an exhaustive 9plus13s count ...

[champagne]

On my side, I started the transfer of the code in my last frame for the phase 2 and the revision of this code.
I am revising deeply the UAs generator and the GUAs generator and also the brute force 2 bands to simplify the process and improve the performance.
If all works as expected (generally, this is not true), I have a reasonable margin to start the phase2 in April when I am back from the southern hemisphere.
The new code is expected to be at the end "gcc compatible" and I could have some free cycles under LINUX to go ahead.

[blue]

All 17 clues sudokus having one or more band/stack with >=7 clues are known.

I can confirm that, now.

Like champagne, regarding the validity of the code: I can only say that it found all of the known puzzles (of that form), each in all of the forms that it was expected to be found in ... 44982 forms for 36659 ED puzzles ... and it didn't find anything else.

[champagne]

All 17 clues sudokus having one or more band/stack with >=7 clues are known.

I can confirm that, now.

Like champagne, regarding the validity of the code: I can only say that it found all of the known puzzles (of that form), each in all of the forms that it was expected to be found in ... 44982 forms for 36659 ED puzzles ... and it didn't find anything else.

Hi blue,

Good to have another view on this code.


For the next step, distribution 566 656, the code is now ready and released int his thread here just because it is part of my new common frame to all programs.

When I closed the first program (17 clues with one band/stack having more than 6 clues) I had in hand some results of the version ready ro tpocess the 566 656 distribution but I was not quite satisfied with the timings I got.

We were in October 2018, I had planned three months off from mid January to mid Aptil 2019, so I decided to try to improve the code during these 5 months.

In the new version, most changes affect the first phase. Nearly nothing has changed in the band 3 processing.

The UAs collectors have been significantly changed and more UAs collected at the start, especially for socket UAs.
The primary filter works now in 128 bits mode and up to 4 x 128 bits UAs can be processed
A new step 2x3y has been added for high index bands 1

I can only test on my laptop, so I have to make extrapolations, but my feeling is that the new version is significantly better then the former one.

The search should start in one month from now, meantime, I'll continue testing with my limited power.

[champagne]

2 and half month later, after a first long test showing tables overflows, the code seems ok.
The comments on the code are requiring a serious revision.
All should be ready within a week for the last release of the code/comments and preliminary results.
As free cycles is a scarce resource, the process will continue with my limited power.

[Mathimagics]

Some related questions:

• how many essentially different (ED) grids are now known to have 17-clue puzzles?

• is a list of these grids readily available?

Cheers
MM

[champagne]

Some related questions:

• how many essentially different (ED) grids are now known to have 17-clue puzzles?

• is a list of these grids readily available?

Cheers
MM

The count is 49157 for long. AFAIR, the last were produced by blue.
I have a file of these 17 in the canonical format of my storage of puzzles, that I can upload somewhere.
I don't remember where is the original file but surely somebody will give the link

[1to9only]

The 49157 17-clue puzzles can be downloaded here: https://sites.google.com/site/dobrichev/sudoku-puzzle-collections.

[blue]

Some related questions:

• how many essentially different (ED) grids are now known to have 17-clue puzzles?

• is a list of these grids readily available?

Cheers
MM

Answering the actual question(s):

• 46300 grids.
• I don't know. Champagne ?

Cheers,
Blue.

[dobrichev]

https://sites.google.com/site/dobrichev/sudoku-puzzle-collections/17grids46300.zip

[blue]

Fantastic !
Thanks Mladen.

[Mathimagics]

Thanks!!

[champagne]

Answering the actual question(s):

• 46300 grids.
• I don't know. Champagne ?

Cheers,
Blue.

I have many files linking 17s to specific canonical forms of the solution grids, but they are debugging/checking oriented and I did not have the file posted by mladen.

BTW, I am slowly adjusting post 1 and I added the first links to the comments on the code currently running pass2 (665 distribution)

[champagne]

Scan of solution grids to search 17s milestone.

The search has been restarted mid may. The first test on band 6 led to table overflows.

The new run was first divided in three parts

A full expansion of the band1 index 6
A full expansion of band1 index 189
Full expansion of bands of High index.

Later, bands of low index were started depending on the availability of cores.


Where we are

All expected 17s are there, no new appeared

In low index Bands index 0,2,3,6 are closed
in high index Bands 331-416 are closed.

This gives 89 bands 1 closed out of 416, but less than 2% of the solution grids processed.

Bands 1 4 5 are started, expected to be closed quickly (relatively small count of bands 2/3)
Band 189 has been started just because many 17s are expected. It is a "middle range band". Both run time and check that all 17s are there are important.
Bands 7 a big one ( as band 6 ) and band 330 are started.

Comments on the process and a "cleaned" version of the code are available, the links are in post 1.