JPF wrote:gfroyle wrote:Why does replacing the puzzle with an isomorph take 2 extra forcing chains? [At least the rating is the same!]
I will leave the experts to answer your question
I
recently found ratings 8.5 and 8.9 for 2 equivalent puzzles.
the SE algorithm was not designed to take isomorphs into account
dukosu's suexrat uses a tree search to rate and is prone to local minima/maxima
it works around this (imperfectly) by re-running on random isomorphs of the same puzzle and reporting the average rating
my solver's -q1 rating attacks the problem by using a singles only tree search (like suexrat)
but it orders the tree search by degree (e.g., bivalue/bilocation cells first, then trivalue/trilocation, etc.)
and counts all propositions for each of the cells of each degree
and continues until the collected propositions yield a solution
there is still an inherent ordering bias in the singles logic, but that tends to affect the last three digits
of 5 digit ratings, so taking the first two digits as n.m should be consistent across isomorphs of the same puzzle
a test on 10 random copies of the easter monster produced a consistent 99408
it took < 6min for all ten copies
on this puzzle
- Code: Select all
600000002090400050001000700050084000000020000000305040200000600030009080007000001
the rating for the puzzle and 9 random isomorphs was consistently 6[01]*** and took 8s @1.9Ghz (ok, second digit affected here)
SE would have take 10*10hr = ~4day to do the same
the -q1 rating for |G|=47308 takes 1m20s @3.2Ghz
tree search profiles of -q1 rating searches are failrly consistent across random isomorphs
so it looks like most of the rating number slop is due to the ordering bias in the singles logic
