The New Sudoku Players' Forum

[denis_berthier]

.
This is to announce the publication of a software by François Cordoliani (Defise on this forum) on GitHub, the Sudoku Hierarchical Classifier (SHC): https://github.com/denis-berthier/Sudoku_Hierarchical_Classifier

The SHC computes 4 classifications of 9x9 puzzles:

T&E-depth
B-rating for the puzzles in T&E(1)
BxB classification for the puzzles in T&E(2)
BxBB classification for the puzzles in T&E(3)

The development of the SHC was based on the same references as CSP-Rules or SudoRules but the implementations were totally independent.

All the classification results of the SHC and of SudoRules completely coincide on the large collections of puzzles used to compare them.

The SHC is much faster than SudoRules for the functionalities it implements.
It is hoped that the T&E-depth and BpB parts will be very useful in the search for the "hardest" puzzles in T&E(2).

Details are available on GitHub.
See also :http://forum.enjoysudoku.com/hierarchical-classifications-in-constraint-satisfaction-t42076.html for theoretical background, classification results, computation times and their comparison with SER times.

[Edit:] Renamed BpB and BpBB as BxB and BxBB
.

[denis_berthier]

.
Today's update on GitHub includes a few speed improvements (up to 30% for the BpB calculations).
In the examples directory, "message" files" have been updated to reflect the new computation times.
.

[denis_berthier]

.
Today's update on GitHub includes large speed improvements made by François. Here are the factors computation times have to be divided by, wrt to the previous version, for the standard examples:

TE-depth : 18,6
B : 4,2
BpB : 5,3
BpBB : 10,5

In the examples directory, "message" files" have also been updated to reflect the new computation times.
.

[Hajime]

I read the pdf 2023-HCCS.pdf.
In it is a table from SER=10.6 to 11.9 and from SHC=B2B to B7B.
Apart from a higher granularity in SER (14 steps) versus SHC (6 steps) , how is the SHC rating below SER 10.6 ?

[denis_berthier]

I read the pdf 2023-HCCS.pdf.
In it is a table from SER=10.6 to 11.9 and from SHC=B2B to B7B.
Apart from a higher granularity in SER (14 steps) versus SHC (6 steps) , how is the SHC rating below SER 10.6 ?

Note that there's no SHC rating per se. SHC is a program that computes 4 classifications.

Table 2.2 is for puzzles in T&E(2).
All the classification results in T&E(2) must be taken with some grain of salt. We have no unbiased collection of T&E(2) puzzles - and not even one with a known bias.
I chose eleven's collection for this table because it was built without using patterns (contrary to the ph2010 database). But even so, we don't know how its distribution is related to the real distribution within T&E(2).

As for puzzles with SER < 10.6 [and > 9.1, because those with SER < 9.1 are usually in T&E(1)]:
- we have no collection that could serve as a reference;
- the T&E(1) vs T&E(2) line cuts somewhere in this range of SERs (but depending on each puzzle);
- therefore, different sub-classifications will apply (B vs BpB)
- based on the results of table 2.2, one can conjecture that, for the T&E(2) part, each value B1B, B2B, B3B spreads over several SER values.

In short, I have no real answer.
.

[denis_berthier]

.
A few minor changes today.
Following the discovery of puzzles in BpB for p>7 (http://forum.enjoysudoku.com/the-bpb-classification-of-t-e-2-puzzles-t41922-5.html), the default values of buffer-size have been increased, because such cases may require more partial-braids.
.

[denis_berthier]

.
fix rare out-of-memory bug
.

[denis_berthier]

.
1) In order to make it easier to identify which version number you are using and to report possible problems, a command has been added:

Code: Select all

java -jar SHC.jar version

gives the version number: *** SHC version 5.2 ***


2) In case a negative value is obtained for the classification of some puzzle in a file (which might be unnoticed by the user), a message is now issued in the Terminal / Command Window:

Code: Select all

===> Problem encountered while processing your input file.
===> Check your output file for negative values and see the manual for their interpretation.

The manual gives the following interpretations:

Code: Select all

-1 : this puzzle is not in the right format, or it has no solution, or it has several solutions; check your puzzle;
-2 : <max-time> allowed is too short for this puzzle; try increasing it;
-3 : <buffer-size> is too small for this puzzle; try increasing it;
-4 :  other problems encountered, such as:
         . <classif> is not  relevant for this puzzle (e.g. applying the BpB rating to a puzzle in T&E(3)); check the T&E-depth of this puzzle;<br>
         . <max-length> is too small for this puzzle (can appear mainly in case <classif> = B); try increasing it;
                 ( note that if you increase <max-length>, you may also have to increase <max-time> and/or <buffer-size>).


[denis_berthier]

.
Only one small change today, but it will make it easier to include SHC in scripts (in particular in the search for the "hardest" puzzles).
(This was suggested by Coloin and Hendrik. Thanks to them for their remarks.)
Version number is 5.3

A new option appears in the command line: (-auto-end <auto-end>). <auto-end> is true by default. It means that, at the end of computations, the process will terminate automatically. If SHC is embedded in a script, the script will not be blocked.
Set <auto-end> to false if you want to restore the old behaviour (that's useful only on Windows, if you want to have time to see the messages in the console at the end of computations).

[Edit]: One more change I forgot to mention: the -erase option is now true by default. (Keeping old values in files led me to too many errors when trying to extract puzzles with fixed rating.) You can change it to false it you prefer keeping the previous results.
.

[denis_berthier]

.
SHC 5.4 fixed a bug that occurred in case there were 81 clues.
Not a situation very likely to happen - but it happened to one user, so we made the road safer.
.

[denis_berthier]

.
A new utility has been added: expansion by Singles. Syntax:
java -jar Expand.jar (-input <input-file>) (-output <output-file>) (-erase <erase>) (-auto-end <auto-end>) (-puzzle <puzzle>)
with all the options working the same way as in SHC.jar.
.

[denis_berthier]

.
Following the discovery of puzzles in B14B by Hendrik and Coloin, the default max-length value for the BpB rating is now 14. (You can always change it with the options).
.

[denis_berthier]

.
In order to avoid any confusion between the BpB classification and the BpB rating of a puzzle, the BpB classification has been renamed BxB classification. This is more consistent with the common usage that x is the unknown.
Reminder:
- in "BxB classification", the "x" is the value to be found for the puzzle.
- in "BpB rating", p is fixed and what's to be found is the minimal length of Bp-braids necessary to solve the puzzle.

As a result, the SHC line command has changed: in <classif>, BxB and BxBB must now be used instead of BpB and BpBB.
.

[denis_berthier]

.
The new release 6.1 of the SHC brings speed increases for the B and BxB calculations.
Little change will appear for small values of B or BxB, but the larger the value of x, the larger the difference will be in the mean.
The gain factor may be as large as 7 for high BxB and as large as 24 for the highest known B values.
This is expected to be very useful in the current hunting of snarks with high BxB values.

BEWARE that the meaning of parameter <max-time> has changed; it now applies only to the full list of of puzzles in a file. See the README.md for details.

.

[denis_berthier]

.
Following the speed improvements brought by François to SHC 6.1, I've expanded the standard set of BxB examples. See thehttp://forum.enjoysudoku.com/hierarchical-classifications-in-constraint-satisfaction-t42076.html thread for the corresponding updated table.
.