The New Sudoku Players' Forum

[pierrenay2]

I have been looking for a solver/ rating system while testing an AI classic Sudoku generator, i came across Hodoko which appears to work quite in delivering a simple score/time based rating system and can be batch processed. as a non master sudoku player, could I ask if the Hodoku rating system is somewhat of the gold standard ? or are there other systems out there ?

[denis_berthier]

I have been looking for a solver/ rating system while testing an AI classic Sudoku generator, i came across Hodoko which appears to work quite in delivering a simple score/time based rating system and can be batch processed. as a non master sudoku player, could I ask if the Hodoku rating system is somewhat of the gold standard ? or are there other systems out there ?

The Hodoku rating system is worth 0.
The only generally accepted system is the Sudoku Explainer Rating.

I have some doubts about an AI generator, unless it includes a non-AI very fast solver.

[pierrenay2]

I have been looking for a solver/ rating system while testing an AI classic Sudoku generator, i came across Hodoko which appears to work quite in delivering a simple score/time based rating system and can be batch processed. as a non master sudoku player, could I ask if the Hodoku rating system is somewhat of the gold standard ? or are there other systems out there ?

The Hodoku rating system is worth 0.
The only generally accepted system is the Sudoku Explainer Rating.

I have some doubts about an AI generator, unless it includes a non-AI very fast solver.

thank you for your directive, trying to understand the scoring system:
Sudoku Explainer default label of HARD = Difficulty rating: 1.7 > 2.5 (Direct Hidden Pair)
The term HARD doesn't necessarily indicate that the puzzle is HARD to solve although it sits 3rd tier up the rank in difficulty / i.e. : difficulty rating more based on difficulty to generate.
therefore, Are the labels used in S.Explainer such as HARD the standard to refer to ( as per excel screen grab? )

re AI Generator: I cant find a concise document that offers info on its output , hence the need to find a solver/rating system that can be said to represent the current standard and work out if this generator is simply a randomizer. as for Ai solvers, as far as i know , there are simple backtracking algorithms using grunt force, not much use really.

[denis_berthier]

Sudoku Explainer default label of HARD = Difficulty rating: 1.7 > 2.5 (Direct Hidden Pair)
The term HARD doesn't necessarily indicate that the puzzle is HARD to solve although it sits 3rd tier up the rank in difficulty / i.e. : difficulty rating more based on difficulty to generate.
therefore, Are the labels used in S.Explainer such as HARD the standard to refer to ( as per excel screen grab? )

Here are the SE ratings:

Code: Select all

How does the Sudoku Explainer rate the difficulty of a Sudoku?
The difficulty is rated by the hardest solving technique that is required to solve it. This is not necessary the best rating scheme, and many other solvers might use different rating for some solving techniques (see previous question).
More precisely, the Sudoku Explainer uses the following difficulty ratings of the solving techniques:
1.0: Last value in block, row or column
1.2: Hidden Single in block
1.5: Hidden Single in row or column
1.7: Direct Pointing
1.9: Direct Claiming
2.0: Direct Hidden Pair
2.3: Naked Single
2.5: Direct Hidden Triplet
2.6: Pointing
2.8: Claiming
3.0, 3.2, 3.4: Naked Pair, X-Wing, Hidden Pair
3.6, 3.8, 4.0: Naked Triplet, Swordfish, Hidden Triplet
4.2, 4.4: XY-Wing, XYZ-Wing
4.5 - 5.0: Unique rectangles and loops
5.0, 5.2, 5.4: Naked Quad, Jellyfish, Hidden Quad
5.6 - 6.0: Bivalue Universal Graves
6.2: Aligned Pair Exclusion
6.5 - 7.5: Bidirectioal X-Cycles and Y-Cycles
6.6 - 7.6: Forcing X-Chains
7.0 - 8.0: Forcing Chains, Bidirectional Cycles
7.5 - 8.5: Nishio
8.0 - 9.0: Cell/Region Forcing Chains
8.5 - 9.5: Dynamic Forcing Chains
9.0 - 10.0: Dynamic Forcing Chains (+)
> 9.5: Nested Forcing Chains
It seems that the difficulties 1.2, 1.5, 1.7 and >= 2 correspond to the degrees 1, 2, 3 and 4 of some newspapers. Note that many solvers are rating Naked Singles easier than Hidden Singles (which is reasonable when candidates are always visible). According to some sources, "diabolical" Sudokus are those that are not solvable without trial and error. But there is a great controversy on what "trial and error" means.

You can see that HARD doesn't appear in the SE ratings.

as for Ai solvers, as far as i know , there are simple backtracking algorithms using grunt force, not much use really.

Using backtracking isn't really AI.
The only AI solver I know of is mine: https://github.com/denis-berthier/CSP-Rules-V2.1
It will be of no use for generating puzzles. But it can be used to filter puzzles generated by a fast generator. I do it when I propose a puzzle in the "Puzzles" section.

[StrmCkr]

problem with rating is that it uses a prefixed order of operations:
which when creating a rating system is good to approximate how hard something could potentially be: however

once you get past the stuff known as basics:
subsets naked/hidden, blr, x-wings

order changes the rating dramatically as some techniques might be simpler in method but ranked harder as they are lumped into a parent technique that has a high class of over all ranking
{they don't act dynamically to change when the used information is smaller}

if you play around with hodokus settings you may find this out very easily by moving around techniques in its hierarchy ratings can drop or go up quickly


that is why the community uses one standard:

sudoku explainer the order is known, the rough rating of each lvl is known
and it is not muddled by words of - easy, medium,hard, extreme, diabolical.
these really have no meaning as no one that ever used them clearly define the order system the rating program uses.

there has been group of users on this form working on updating it to be more modernized ie containing more advanced techniques
as se creator hasn't updated it to reflect new found methods in 10 years.

[ghfick]

I agree with StrmCkr.
HoDoKu was designed to give the human solver complete freedom to design the rating based on the human solver's own current views as to the difficulty of a particular step. HoDoKu's rating is then a sum of ALL of the scores for every step in a solution path.
Sudoku Explainer [SE] uses 3 facets of a puzzle's difficulty : explainer rating (ER), the pearl rating (EP) and the diamond rating (ED). For example, SE might give an ER of 11.0 but the actual solution path might contain just one 11.0 step or many such steps at 11.0 or close to 11.0. So should the puzzle with many 11.0 steps be rated harder than a puzzle with just one 11.0 step?
Both HoDoKu and SE are not current. champagne has detailed a 'potential hardest' list based on SE but he has also separated off these really hard puzzles [based on SE] into those that can now be solved with steps like Junior Exocet, MSLS and other more recently developed techniques.

[denis_berthier]

Sudoku Explainer [SE] uses 3 facets of a puzzle's difficulty : explainer rating (ER), the pearl rating (EP) and the diamond rating (ED).

Only the ER rating is widely used. The other two are for the Patterns Game.

For example, SE might give an ER of 11.0 but the actual solution path might contain just one 11.0 step or many such steps at 11.0 or close to 11.0. So should the puzzle with many 11.0 steps be rated harder than a puzzle with just one 11.0 step?

This is totally unrelated to the EP and ED.
SER is a rating of the hardest step; this is a clear limitation; but any other rating based on the full path is much worse.

Both HoDoKu and SE are not current. champagne has detailed a 'potential hardest' list based on SE but he has also separated off these really hard puzzles [based on SE] into those that can now be solved with steps like Junior Exocet, MSLS and other more recently developed techniques.

champagne's list of hardest puzzles (compiled from other lists) is totally irrelevant to SE being "current" or not. The "hardest" in "hardest list" means the rating is computed with SER.

The only fact is, SER is the only rating widely accepted by everybody. We all know its shortcomings, but it remains the only common reference.

[StrmCkr]

http://forum.enjoysudoku.com/post60757.html#p60757 this post is probably the more actuate way to gauge a grids difficulty

however the process of doing such a task is very daunting: my own attempts at such a method are arduous at best
my program can do find all moves at each phase/step of solving - which tends to be massive long run times in a solution path with hundreds/thousands of active methods with same/similar or extra eliminations on-top of thousands of dead ends techniques with no eliminations.

then if i go into identifying each move as a subset branch or a specific path i end up with millions of trees with various steps lengths

hierarchy preset move search executed one after another has faster run times: but always misses potential short cuts from order changes.

in short i don't think my theory method for a 3 fold method of evaluation is viable with current tech as it is also appears to be p=np problem.

[pierrenay2]

Sudoku Explainer default label of HARD = Difficulty rating: 1.7 > 2.5 (Direct Hidden Pair)
The term HARD doesn't necessarily indicate that the puzzle is HARD to solve although it sits 3rd tier up the rank in difficulty / i.e. : difficulty rating more based on difficulty to generate.
therefore, Are the labels used in S.Explainer such as HARD the standard to refer to ( as per excel screen grab? )

Here are the SE ratings:

Code: Select all

How does the Sudoku Explainer rate the difficulty of a Sudoku?
The difficulty is rated by the hardest solving technique that is required to solve it. This is not necessary the best rating scheme, and many other solvers might use different rating for some solving techniques (see previous question).
More precisely, the Sudoku Explainer uses the following difficulty ratings of the solving techniques:
1.0: Last value in block, row or column
1.2: Hidden Single in block
1.5: Hidden Single in row or column
1.7: Direct Pointing
1.9: Direct Claiming
2.0: Direct Hidden Pair
2.3: Naked Single
2.5: Direct Hidden Triplet
2.6: Pointing
2.8: Claiming
3.0, 3.2, 3.4: Naked Pair, X-Wing, Hidden Pair
3.6, 3.8, 4.0: Naked Triplet, Swordfish, Hidden Triplet
4.2, 4.4: XY-Wing, XYZ-Wing
4.5 - 5.0: Unique rectangles and loops
5.0, 5.2, 5.4: Naked Quad, Jellyfish, Hidden Quad
5.6 - 6.0: Bivalue Universal Graves
6.2: Aligned Pair Exclusion
6.5 - 7.5: Bidirectioal X-Cycles and Y-Cycles
6.6 - 7.6: Forcing X-Chains
7.0 - 8.0: Forcing Chains, Bidirectional Cycles
7.5 - 8.5: Nishio
8.0 - 9.0: Cell/Region Forcing Chains
8.5 - 9.5: Dynamic Forcing Chains
9.0 - 10.0: Dynamic Forcing Chains (+)
> 9.5: Nested Forcing Chains
It seems that the difficulties 1.2, 1.5, 1.7 and >= 2 correspond to the degrees 1, 2, 3 and 4 of some newspapers. Note that many solvers are rating Naked Singles easier than Hidden Singles (which is reasonable when candidates are always visible). According to some sources, "diabolical" Sudokus are those that are not solvable without trial and error. But there is a great controversy on what "trial and error" means.

Thank you for the detailed list & breakdown, with that, the solver's scores are to be derived using Explainer rating, this is most helpful.

[pierrenay2]

The only AI solver I know of is mine: https://github.com/denis-berthier/CSP-Rules-V2.1
It will be of no use for generating puzzles. But it can be used to filter puzzles generated by a fast generator. I do it when I propose a puzzle in the "Puzzles" section.

Thank you for the github link :you've given me a lot of homework :ugeek , i am also very curious about the other puzzles you've chosen to solve with your CSP system : I will take some time off to read on
in terms of sudoku generator, aside from the experimental AI generator mentioned earlier : I have been using QQWING, assume there is no issue there

[1to9only]

Here is a finer breakdown of SE Ratings I copied off a very old thread here, found it whilst browsing/searching for somethings else!

Code: Select all

 1.0 Single
 1.2 Hidden Single in box
 1.5 Hidden Single in line
 1.7 Direct Pointing
 1.9 Direct Claiming
 2.0 Direct Hidden Pair
 2.3 Naked Single
 2.5 Direct Hidden Triplet
 2.6 Pointing
 2.8 Claiming
 3.0 Naked Pair
 3.2 X-Wing
 3.4 Hidden Pair
 3.6 Naked Triplet
 3.8 Swordfish
 4.0 Hidden Triplet
 4.2 XY-Wing
 4.3 [Direct Hidden Quad]
 4.4 XYZ-Wing
 4.5 UR Types 1 or 2 or 4 or 3 w/ hidden pair
 4.6 UR Type 3 w/ naked pair or hidden triplet
     UL Types 1 or 2 or 4 or 3 w/ hidden pair (6 cells)
 4.69 UL Type 3 w/ a naked pair or hidden triplet (6 cells)
 4.7 UR Type 3 w/ naked triplet or hidden quad
     UL Types 1 or 2 or 4 or 3 w/ hidden pair (8 cells)
 4.8 UR Type 3 w/ naked quad
     UL Type 3 w/ naked triplet [or hidden quad] (6 cells)
     UL Type 3 w/ naked pair or hidden triplet (8 cells)
 4.89 UL Type 3 w/ naked quad (6 cells)
 4.9 [UL Type 3 w/ naked triplet or hidden quad (8 cells)]
 5.0 Naked Quad or UL 1 or 2 or 4 (>=10 cells)
 5.1 UL Type 3 w/ naked pair (>=10 cells)
 5.2 Jellyfish
 5.3 Unknown
 5.4 Hidden Quad
 5.5 Unknown
 5.6 BUG Type 1
 5.7 BUG Type 2 or 4
 5.8 BUG Type 3 w/ naked pair
 5.9 BUG Type 3 w/ naked triplet
 6.0 BUG Type 3 w/ naked quad
 6.1 BUG Type 3 w/ naked quint
 6.2 Aligned Pair Exclusion
 6.3 Unknown
 6.4 Unknown
 6.5 Bidirectional X-Cycle or Bidirectional Y-Cycle (1-4 nodes)
 6.6 Turbot Fish
     Forcing X-chain or Bidirectional Y-Cycle (5-6 nodes)
 6.69 Forcing X-Chain (7-8 nodes)
 6.7 Bidirectional Y-cycle (7-8 nodes)
 6.8 Forcing X-Chain or Bidirectional Y-cycle (9-12 nodes)
 6.9 Forcing X-Chain or Bidirectional Y-cycle (13-16 nodes)
 7.0 Bidirectional Y-cycle (17-24 nodes)
     Forcing Chain or Bidirectional Cycle (1-4 nodes)
 7.1 Forcing Chain or Bidirectional Cycle (5-6 nodes)
 7.2 Forcing Chain or Bidirectional Cycle (7-8 nodes)
 7.3 Forcing Chain or Bidirectional Cycle (9-12 nodes)
 7.4 Forcing Chain (13-16 nodes)
 7.5 Forcing Chain (17-24 nodes)
     Aligned Triplet Exclusion
 7.6 Forcing Chain (25-36 nodes)
     Nishio Forcing Chain (5-6 nodes)
 7.7 Nishio Forcing Chain (7-8 nodes)
 7.8 Nishio Forcing Chain (9-12 nodes)
 7.9 Nishio Forcing Chain (13-16 nodes)
 8.0 Nishio Forcing Chain (17-24 nodes)
 8.1 Nishio Forcing Chain (25-36 nodes)
 8.2 Multiple (7-8 nodes) Region Forcing Chains
 8.3 Multiple (9-12 nodes) Cell/Region Forcing Chains
 8.4 Multiple (13-16 nodes) Cell/Region Forcing Chains
 8.5 Multiple (17-24 nodes) Cell/Region Forcing Chains
 8.6 Multiple (25-36 nodes)
     Dynamic (5-6 nodes) Cell/Region Forcing Chains
 8.7 Dynamic (7-8 nodes) Cell/Region Forcing Chains
 8.8 Dynamic (9-12 nodes) CRCD Forcing Chains
 8.9 Dynamic (13-16 nodes) CRCD Forcing Chains
 9.0 Dynamic (17-24 nodes) CRCD Forcing Chains
 9.1 Dynamic (25-36 nodes) CRCD Forcing Chains
 9.2 Dynamic (37-48 nodes) CRCD Forcing Chains
 9.3 Dynamic (49-72 nodes)
     Dynamic + (9-12 nodes) CRCD Forcing Chains
 9.4 Dynamic (73-96 nodes)
     Dynamic + (13-16 nodes) CRCD Forcing Chains
 9.5 Dynamic + (17-24 nodes) CRCD Forcing Chains
 9.6 Dynamic + (25-36 nodes) CRCD Forcing Chains
 9.7 Dynamic + (37-48 nodes) CRCD Forcing Chains
 9.8 Dynamic + (49-72 nodes) CRCD Forcing Chains
 9.9 Dynamic + (73-96 nodes) CRCD Forcing Chains
10.0 Dynamic + (97-144 nodes)
     Dynamic + Forcing Chains (17-24 nodes) CRCD Forcing Chains
10.1 Dynamic + (145-192 nodes)
     Dynamic + Forcing Chains (25-36 nodes) CRCD Forcing Chains
10.2 Dynamic + Forcing Chains (37-48 nodes) CRCD Forcing Chains
10.3 Dynamic + Forcing Chains (49-72 nodes) CRCD Forcing Chains
10.4 Dynamic + Forcing Chains (73-96 nodes) CRCD Forcing Chains
10.5 Dynamic + Forcing Chains (97-144 nodes) CRCD Forcing Chains
10.6 Dynamic + Forcing Chains (145-192 nodes) CRCD Forcing Chains
10.7 Dynamic + Forcing Chains (193-288 nodes) CRCD Forcing Chains
10.8 Dynamic + Forcing Chains (289-384 nodes) CRCD Forcing Chains
10.9 Dynamic + Multiple Forcing Chains (73-96 nodes) CRCD Forcing Chains
11.0 Dynamic + Multiple Forcing Chains (97-144 nodes) CRCD Forcing Chains
11.1 Dynamic + Multiple Forcing Chains (145-192 nodes) CRCD Forcing Chains
11.2 Dynamic + Multiple Forcing Chains (193-288 nodes) CRCD Forcing Chains
11.3 Dynamic + Multiple Forcing Chains (289-384 nodes) CRCD Forcing Chains
11.4 Dynamic + Multiple Forcing Chains (385-576 nodes) CRCD Forcing Chains
11.4 [Dynamic + Dynamic Forcing Chains (73-96 nodes) Region/Contradiction Forcing Chains]
11.5 [Dynamic + Dynamic Forcing Chains (97-144 nodes) Region Forcing Chains]
11.6 [Dynamic + Dynamic Forcing Chains (145-192 nodes) Cell Forcing Chains]
11.7 [Dynamic + Dynamic Forcing Chains (193-288 nodes) Double Forcing Chains]

CRCD=Cell/Region/Contradiction/Double
UL=Unique Loop
UR=Unique Rectangle
BUG=Bivalue Universal Grave



[denis_berthier]

Here is a finer breakdown of SE Ratings I copied off a very old thread here, found it whilst browsing/searching for somethings else!

That's great. I had been looking for this detailed list, but I was unable to find it.

[pierrenay2]

Here is a finer breakdown of SE Ratings I copied off a very old thread here, found it whilst browsing/searching for somethings else!

that's fantastic. May i suggest, as a newbe having struggled to get more information on classic sudoku, that this list be christened into the annals of sudoku forum :0

[denis_berthier]

in terms of sudoku generator, aside from the experimental AI generator mentioned earlier : I have been using QQWING, assume there is no issue there

I had never heard of it before. As long as it produces puzzles with a unique solution (preferably minimal), I don't think there can be any problem.
You could compare generation speed with other generators, for instance the classical top-down suexg-td.c - which I have included here for reference: https://github.com/denis-berthier/Controlled-bias_Sudoku_generator_and_collection/tree/master/PROGRAMS

You're probably already aware that different types of generators output different distributions of puzzles (wrt complexity). Otherwise, see chapter 6 of [PBCS].

[pierrenay2]

in terms of sudoku generator, aside from the experimental AI generator mentioned earlier : I have been using QQWING, assume there is no issue there

I had never heard of it before. As long as it produces puzzles with a unique solution (preferably minimal), I don't think there can be any problem.
You could compare generation speed with other generators, for instance the classical top-down suexg-td.c - which I have included here for reference: https://github.com/denis-berthier/Controlled-bias_Sudoku_generator_and_collection/tree/master/PROGRAMS

You're probably already aware that different types of generators output different distributions of puzzles (wrt complexity). Otherwise, see chapter 6 of [PBCS].

thanks for your suexg-cb link, ill put it in motion as well and yes , all the puzzles generated are intended to have only 1 unique solution (tested accordingly) , given the time, ill map out a summary of results as i go along