Paquita wrote:One thing I noticed is that among the hardest puzzles some solutions occur quite often. For the about 50.000 high rated puzzles I selected there were 33625 different solutions, the highest occurence was 53 times the same solution. It can be a result of vicinity search or maybe there are solutions that create harder puzzles. The patterns of givens on the other hand are not in clusters that occur often.
Vicinity search has an obvious impact. Probably much higher than what you observed, if you have tested equality of solutions instead of isomorphism.
I don't know for patterns of givens, but patterns in the usual sense occur very often: sk-loops in the ph database, tridagons in mith's database...Obviously, such patterns are much more stable under vicinity search than patterns of givens.
Paquita wrote:Anyway right now I am puzzled by some rating results. If I rate minlex instead of maxlex, some 11.8 become 11.9. But some other 11.8 become 11.7.
Also, when I rate minlex minimals derived from rated expanded minlex puzzles, if the expanded have 11.9, the minimals sometimes are 11.8. This is against my expectation that a minimal puzzle has less information so would be potentially harder. (I use PGE for rating now, skfr does not work on my newer processor anymore).
It would be nice, when searching for the hardest puzzles, if the rating would not show such glitches. Not just because the search is a game, but also to have a stable, reliable rating method that can define what is hardest.
Mith wrote:SErate has the same bug as PGE
Such problems of SER have been known for quite a long time.
Unfortunately, it's not a programming bug in SE (or at least not only that). It's before all a conceptual bug.
Sudoku Explainer is not based on a set of resolution rules.
SE "rules", i.e. (dynamic)-contradiction-(forcing)-chains, are not resolution rules in the precise sense I've given to this expression: due to their counting of "nodes", i.e. of inference steps, they can't be written as pure logic formulae. From this and from the arbitrary thresholds in the numbers of nodes defining the SER levels, "rules" at some SER level don't have the confluence property: as a result, different paths can lead to different ratings. Isomorphisms may imply different paths...
The only ratings/classifications I know of that don't have this problem are those I've defined (B, gB, SB and BB ratings and T&E-depth)
SER is useful to give an idea of difficulty. Don't expect of it more than it can provide.
There's no unique rating that could define what's hardest. You always have to specify hardest wrt what.
.