The New Sudoku Players' Forum

[denis_berthier]

Here is the link to MithsBigList

Hi coloin
Thanks for making your work public and avoiding us all the necessity of re-doing it. A few minimals that are not real minimals in the original collection is enough to make computations inconsistent.

I tried to do a bit more expanding keeping TE3 but mith seems to have done this pretty well.
I was able to get possibly 46 more puzzles ... but need to check fully.

I think he has been very exhaustif in his collection. But my purpose is different: I'm not searching for new puzzles. I'm now more interested in the (+1+BRT) layers above the min-expands in T&E(3) and in the border with T&E(2). The example I referred to before shows there are at least 9 layers in T&E(3).
I've studied similar questions in [HCCS2], but I now have scripts in CSP-Rules able to call SHC and gsf's solver and I can now deal with larger collections.
.

[denis_berthier]

.
Hi coloin

One more point: when you say "minlex" you mean gsf's solution minlex (i.e. with the -f'%#.c' setting) ?
.

[coloin]

Ive included a list of min expand puzzles - One minlex puzzle [largest minlex] for each solution grid..

I should have written
"Ive included a list of min expand puzzles [in minlex solution grid format] with one largish lexographically puzzle for each solution grid."
So rather than processing 300K/4M puzzles, 67K puzzles - one from each solution grid is quite likely to be representative.

I guess what we need is a filter for puzzles which have a single tridagon elimination as a first step ....

adding clues to the TE3 and then to BxB5 or BxB6 - would give many BxB5 or BxB6 puzzles to minimize furthur.
BxB >6 is very rare as we have found out

[denis_berthier]

"Ive included a list of min expand puzzles [in minlex solution grid format] with one largish lexographically puzzle for each solution grid."
So rather than processing 300K/4M puzzles, 67K puzzles - one from each solution grid is quite likely to be representative.

... but representative of what?

I guess what we need is a filter for puzzles which have a single tridagon elimination as a first step ....

We already know that this is a very special case - and not necessarily the simplest one.

adding clues to the TE3 and then to BxB5 or BxB6 - would give many BxB5 or BxB6 puzzles to minimize furthur.
BxB >6 is very rare as we have found out

Adding clues to the TE3 is very different from starting from B5B or B6B. In the latter case, vicinity search has few chances of finding T&E(3) or a BxB > 7
But when crossing the T&E(3) border by adding a clue to a T&E(3)-expand, one can get very large B or BxB, as I've shown in [HCCS2].
.
I'll say a word of what I'm currently doing.
I'm splitting the collection of minimals into the full sub-collections associated with each solution grid. (That's a longest part of the computations.)
Every puzzle/solution is in gsf's solution minlex format, so that "P1 included in P2" is easy to check and 1-expansions are trivial.
Basic operations are:
- 1-expansion (add 1 candidate from the solution) (via SudoRules)
- BRT-expansion (exp. by Singles) (via SHC)
- elimination of redundancies (via gsf's solver)
- filter by T&E (or BxB) (via SHC)
I'm developing CSP-Rules scripts that will allow to automatise all the calculations for a collection of grids and associated minimals - and that will work 100% grid by grid, so as to optimise the hardest step, i.e. the elimination of redundancies.
My purpose is not to find new puzzles - though this could happen - because various levels of expansion may have more minimals.
.

[coloin]

... but representative of what?[

The min expand is representative of all the puzzles inherent with less clues .. It could be wrong but I cant see it ! Chosing one random minimal puzzle for each grid may well be similar in effect.
I can see a case that the max expand puzzle might be less complex than some of its inherent lesser puzzles
If you are able to process all the minimals..then that would obviously be more complete !

I ordered the representative puzzles with the gsf -q2 rating, as quite a good index of complexity ..... except the top rated puzzle had a single tridagon elimination reducing 11.7 to 9.5.

Code: Select all

........94.7............1.52.684....78.3.2....34...8.234..28.6..726.4...6.873.... # 99379 FNBP C30/M2.9.2916   ED=11.7                                                                                                                     
........94.7............1.52.6847...78.3.2....34...8.234..28.6..726.4...6.873.... # 30686 FNBP C31/M2.115.57   ED= 9.5

I think Mith did several rounds of expansion/minimization to get new puzzles as well a {-1+1} on the minimals to make new minimals and test TE3.
He also generated the associated twin puzzles.
Ive been generating new TE3 by a {-2+2} to produce minimals ... but this is limited as 1 puzzle gives ~ 5000 new minimals.
TE3 puzzles were also found relatively commonly when minimizing Bxb6 puzzles...

I like the idea that you can generate Bxb6 puzzles form a TE3 by adding a clue. All minimal puzzles which have the new clue will be BxB6
Finding Bxb>6 was only possible by minimizing an expanded BxB6 and very very rarely a higher order minimal puzzle was found.

Grouping the puzzles by grid solution is handled very ably and reliably with gridchecker, it also removes redundant clues, handling big files super fast.

[denis_berthier]

I can see a case that the max expand puzzle might be less complex than some of its inherent lesser puzzles
If you are able to process all the minimals..then that would obviously be more complete !

By definition of a max-expand (or T&E(3)-expand in my more precise vocabulary), it is still in T&E(3).
It can certainly be less complex, in many senses: fewer guardians for the tridagon, maybe simpler resolution after the tridagon rules have been used.

I like the idea that you can generate Bxb6 puzzles form a TE3 by adding a clue. All minimal puzzles which have the new clue will be BxB6
Finding Bxb>6 was only possible by minimizing an expanded BxB6 and very very rarely a higher order minimal puzzle was found.

That's the whole idea: starting from known harder puzzles, because the chances of finding hard puzzles by vicinity search from less hard ones are very low.
When you're on the T&E(3)-T&E(2) border, on the T&E(3) side, i.e. when you have a T&E(3)-expand puzzle, if you add a single clue (1-expansion), then you can get (non-minimal) puzzles in any of T&E(0, 1, 2). Most of them are not interesting. But you can get a few with high B or BxB classifications - and still harder ones by minimisation.
I've done this on a small scale in [HCCS2], but the scripts Im' writing will allow anyone to do it on a larger scale.

Grouping the puzzles by grid solution is handled very ably and reliably with gridchecker, it also removes redundant clues, handling big files super fast.

I've already lost too much time trying to compile on my Mac a program that is based on obscure x86 code and restricted to specific variants of Unix.
I wrote CLIPS functions to do that. However slow it is, it works on my Mac and it costs me only little time to program. I let one core run this in the background.
In my view, this is only preparatory work for an approach that will by based on totally separate processes for different solution grids.
However, if you can do that easily and reliably on your PC with gridchecker, I'm interested. That would free one processor of my Mac.
(BTW, the work you have done on mith's collection, cleaning the minimals and the solution grids, has already been very helpful. In the night, I got my first batch of minimals for each of the first 1000 solutions and I can start applying my (still partial) scripts to it.)
.