If rating can’t go below 100, how can you know if someone is 68k?
Is the 100 minimum a glicko thing, or an ogs add on?
“Twenty plus kyus” have so little knowledge that their moves are almost random.
It’s pointless digging under that. 68k has no sense. Under 25k you just don’t know how to play.
You can translate rating numbers into kyus without limits. I did that with the previous rating system an still there were no players under 30k on OGS. I don’t know what would happen with current rating. But the sense is always the same: ranks are just useful for coupling players, and they must have at least a little knowledge of general strategy and life-and-death situations. Absolute beginners don’t have it.
2 Likes
I strongly disagree with that. I have fond memories of my 22kyu friend who used to give me 9 stones. Then one day could only give me 8 stones. Then 7 stones. And so on. That was quite a progress!
The rankings from the French go federation can go up to 30 kyu.
Children go schools also typically have kyu badges that go up to 50 kyu.
And of course there is the very impressive progression of LeelaZero, who started playing literally random moves, then excruciatingly slowly improved until it was about 60 kyu, and then slowly but very measurably improved until it was about 45 kyu, and then improved super fast and achieved superhuman strength.
12 Likes
It seems to me that “almost random” only gets you halfway to the worst possible play.
1 Like
we need Leela Zero with inverted goal
3 Likes
Is there any reason to believe that there is even a “worst possible play”, assuming we’re discounting a player or bot that always resigns by move X.
Resigning on your first move would certainly be the worst possible. Equally I suppose always passing would be only slightly better.
But once you get into random play, is it clear there’s a worst possible random?
What if there’s distributions that play better than others so that you could arbitrarily bad players.
Once you have a player/bot with 0% chance of beating another bot, they’re already on incomparable Elo scales.
1 Like
Pass should be impossible. Then pass should happen when normal bot would pass. So, bad vs bad game would make sense. It probably would look like full board seki without eyes in Chinese rules.
For Japanese rules its possibly better to force bot in reverse Atari Go. No pass, who captures first, loses.
Are they?
Imagine you take KataGo, and then you build a family of bots R_100, R_99.5, R_99, R_98.5, …, R_0.5, R_0, with the following rule: bot R_x has x% probability of choosing KataGo’s move, and (100-x)% of choosing a move uniformly at random on the board, at every move. The random process is repeated independently at every move.
This means that bot R_100 is identical to KataGo, and bot R_0 plays fully randomly. Now of course R_100 will always win against R_0, but R_100 will not always win against R_99.5 (if R_99.5 plays 150 moves in a game, then R_99.5 has 99.5%^150 probability of never deviating from KataGo’s suggestions, and that’s about 50%), so they can be compared. Likewise R_99.5 can be compared to R_99, which can be compared to R_98.5, etc., down to R_0, so that you can end up with an ELO scale that contains both KataGo and the fully random bot.
5 Likes
A random bot which makes 120 moves in a game has probability about 1/( 361x359x…x123) to play like Katago, so has a winning probability about 10-285. This represents an Elo difference of about 100000.
6 Likes
In theory it sounds good, but it might be that R_0.5 has 100% winning rate against R_0, or anywhere along the chain something like this could happen.
Elo doesn’t model the case where one player wins 100% of the time. The probability of one player winning is always non-zero.
The point I also was making is that “random” doesn’t just mean uniform at random. You can pick any distribution you want, and it could be that some distributions win 100% of the time against others.
If we assume katago is completely deterministic, so that “playing like katago” isn’t a random distribution, then the uniform random bot does have a small non-zero chance to play “like katago” by
assuming finite games, so some area rules with superko.
1 Like
What api is best for getting the rank data for these graphs?
I think api/v1/players/ should be reasonably fast, but I don’t know how one would generate the “veterans” graph without api/v1/players/%%/full.
You could try do some kind of tree crawling, like take a recent game, find the players, find their recent opponents and their recent opponents and so on?
1 Like
I used tree walk on the rating data
https://online-go.com/termination-api/player/605979/v5-rating-history
2 Likes
What percentile is all time 9d and 9d+?