dobrichev wrote:The last night I did an experiment how many hard puzzles could be found after depth relabeling followed by minimization of the appeared non-minimals.
I use the term relabeling for pattern-preserving transformation, where the givens' positions are fixed, and {-x+x} transformation is performed on a seed puzzle.
.........
The whole exercise took 21 hours - that is the time stamp difference between latest and first files in the directory. This time include several hours of idle CPU and much time in single-thread load, although the relabelling and later the rating have been done in parallel.
I know most of the listed puzzles are already discovered but hope there are few new 11+ there. I have no plans to rate the possible new ones with SE.
Cheers,
MD
The result, if we look at new entries, looks disappointing. Only one new puzzle rating 11.1 (skfr).
In fact, the test has been done in the area 21_23 clues. That area has been deeply searched during the last five years and my feeling is that more than 90% of possibles are in the data base.
Now some personal comments on the process you described:
The runtime grows very fast with the depth.
Within the pattern (relabelling for you), exceeding a depth +-3 must be done with a specific target.
"in" Changing the pattern +-1 is ok, +-2 already very long +-3 nearly infeasible.
"out"But the acceptable processes (+-3
in +-1
out) do not extend easily the area where puzzles are searched.
Finding new seeds is a key point and the process you describe is one way to find new seeds in virgin areas.
In the pattern game, where the pattern is locked, I tested both a guided scan and a deep search.
The guided scan has better chances to give seeds covering all the field, but it is very expensive. Anyway, it is not working here.
Looking for fresh seeds, you can expand any puzzle, but you have then to have a highly permissive filter for the next cycle(s).
I am running a +-4
in search in the 25 clues area. The yield is around 3% of fresh seeds with my cut_off at the skfr rating 10.0.
I process about 1000 puzzles per core/day, very very low. I get less fresh seeds than the number of puzzles processed, but I hope this will search in new areas in the next cycle.
In the pattern game, I often extend the search to +-4 for highest ratings, usually not more.
In the current game, I did not find Patrice's 10.2 submission (I never use submissions for the generation, it would pollute my test). I compared the puzzle with my collection of ratings over 10. The closest puzzle seems to differ by 9 clues. A +-9 would not be acceptable in the scope of the game. The results would come after game closure.
Last comment : I never keep non minimal puzzles.
in the pattern game, it has no value,
in the search for potential hardest, as I am studying "n" clues when "n-1" is nearly covered, it has a poor added value.
dobrichev wrote:champagne wrote:The file I published is my data base of seeds used in the search of "potential high ratings"
It does not contains the "credit" if you mean the reference to the first known publication
That means most of the puzzles are not published previously, collection is initially composed by you, and therefore is credited to you. Due to the all possible complications I proposed not to credit individual puzzles.
I think there is no credit issue. Many of these puzzles are probably in the working files of eleven.
There are sources and naming issues.
Usually, when anybody uses a file, the source is given, but it's enough to have a name of the file as a first step (as the taxonomy file of gsf).
Naming a puzzle is another point. If a puzzle (and only in that case to my view) plays a key role in the discussions, then it's good to give him a name.
So far, in the file of "potential hardest" we have a sequence number. IDs are just showing the source and the date of first publication for the last puzzles entered. Only very old puzzles have a name.
A file of puzzles having received a name could be kept separately. Important puzzles as "fata morgana", "Platinum Blonde" are not in my data base of "potential hardest.
BTW, platinum blonde, classified as the most difficult in the study pointed by JPF has a serate rating of 10.4 only. In that study, it is classified as tougher than Golden Nugget.
