![[Welcome to Sensei's Library!]](images/stone-hello.png "Sensei's Library")
- 2004-Nov: GNU Go 3.6 released. Somewhat stronger than 3.4, and probably of comparable speed (though perhaps slightly slower). At KGS GNU Go 3.6 based bots are usually rated at 12k.
- 2003-Jul: GNU Go 3.4 released. Around two handicap stones stronger than 3.2 and is approximately as fast as 3.2.
- 2002-Apr: GNU Go 3.2 released. About three handicap stones stronger than 3.0, uses less RAM, and about as fast.
- 2001-Aug: GNU Go 3.0 released. About four handicap stones stronger than 2.6.
- 2000-Feb: GNU Go 2.6 released. At KGS its rated at 17k.
Seems like there's some internal inconsistency here. Starting with the known KGS rating of v2.6, I'd expect v3.6 to be "somewhat stronger than" 8k! At least one of those versions wasn't quite the quantum leap in strength mentioned...:-)
Actually all above can be correct. KGS rating have been adjusted. Rating algorith has been revised several times. KGS rating has known problem of drifting. So yyou just cannot compare 2000 rating into 2004 rating. But also if strenght gains were measured agains Gnu go itself then it is quite expected that results are not too trust worthy.
Lets assume that A > B > C in strength when they play each other. I suspect that usually (A-B) + (B-C) > A-C. More there are version jumps, probably bigger error when going from version A to B to C to D to ...: They should play against every version and not only previous/next version. Also difference between current and previoius version might be much bigger than when playing against humans.
Gnu Go has 10 levels of strength. Is it a kyu-like or a dan-like ordering, i.e. which one is supposed to be stronger? I set up two games between level 10 and level 1, with the following results (no handicap, komi 6.5):
Level 10 sure did take its time thinking, though…
I understand this is not an entirely correct way to compare them, but still curious.