Thanks to the hard work by @dexonsmith we have some long sought updates:
1) Draws and Ing komi value updates
The first is that draws are now officially supported as a game result. This primarily affects New Zealand rules as most rules have fractional komi to specifically avoid draws, but New Zealand rules explicitly allow for them, however if you are using non standard whole number komi on any rule set it will also result in a Draw as this is intuitive.
Because of this change, the other similarly veined change has been that Ing rules now use komi of 7.5 for even games and 0.5 for handicap games. This deviates from the official rules that call for 8 handicap and 0 handicap respectively, where black wins would be ties in even games and white wins would be ties in handicap games, so the outcome is unchanged, but the values used align with our new expectations that if the score is equal, the game result is a draw.
2) Ranked games are now allowed beyond 9 ranks
It’s been an ongoing issue where players couldn’t play bots that were beyond their rank range, forcing them to rank down by losing against strong bots before getting to appropriately ranked bots. That limitation has now been lifted and is generally applicable for all players. We ran the numbers and determined that the rating system should be fine with the change, particularly for new players trying to establish their rank against established bots and players. We’ll be keeping on eye on the change to see how much it affects bot ranks moving forward, but we hope this change should make the initial and continued experience better for new players in particular.
3) Handicap and komi changes for small boards
For years we’ve used the “Old Japanese Recommendation” for our handicap and komi selections, which gives an unfair advantage to Black. @dexonsmith not only fixed a bug in the old system which inappropriately awarded too many handicap stones (further increasing the problem), but also revamped the allocation of handicap stones and komi based on a lot of research and number crunching to be what we think is a lot more fair allocation of handicap stones and komi. Below are the new tables showing the number of handicap stones and komi given for a given rank difference.
9x9
Rank Difference
Handicap Stones
AGA
Chinese
Ing
Japanese
Korean
New Zealand
0
0
7.5
7.5
7.5
6.5
6.5
7
1
1
5.5
5.5
5.5
4.5
4.5
5
2
1
3.5
3.5
3.5
2.5
2.5
3
3
1
1.5
1.5
1.5
0.5
0.5
1
4
1
-0.5
-0.5
-0.5
-1.5
-1.5
-1
5
1
-2.5
-2.5
-2.5
-3.5
-3.5
-3
6
2
7.5
7.5
7.5
6.5
6.5
7
7
2
5.5
5.5
5.5
4.5
4.5
5
8
2
3.5
3.5
3.5
2.5
2.5
3
9
2
1.5
1.5
1.5
0.5
0.5
1
13x13
Rank Difference
Handicap Stones
AGA
Chinese
Ing
Japanese
Korean
New Zealand
0
0
7.5
7.5
7.5
6.5
6.5
7
1
1
3.5
3.5
3.5
2.5
2.5
3
2
1
-0.5
-0.5
-0.5
-1.5
-1.5
-1
3
2
7.5
7.5
7.5
6.5
6.5
7
4
2
3.5
3.5
3.5
2.5
2.5
3
5
2
-0.5
-0.5
-0.5
-1.5
-1.5
-1
6
3
7.5
7.5
7.5
6.5
6.5
7
7
3
3.5
3.5
3.5
2.5
2.5
3
8
3
-0.5
-0.5
-0.5
-1.5
-1.5
-1
9
4
7.5
7.5
7.5
6.5
6.5
7
As usual, let us know how these changes feel in practice and of course let us know if you run into any issues!
It’s been an ongoing issue where players couldn’t play bots that were beyond their rank range, forcing them to rank down by losing against strong bots before getting to appropriately ranked bots. That limitation has now been lifted and is generally applicable for all players. We ran the numbers and determined that the rating system should be fine with the change, particularly for new players trying to establish their rank against established bots and players. We’ll be keeping on eye on the change to see how much it affects bot ranks moving forward, but we hope this change should make the initial and continued experience better for new players in particular."
WOOOOOOOOOOOOOOOOOO!!! Maybe I can actually find a game! lol Also, I have really wanted to improve new users experience
As a point of interest, in the first version of Ing rules 8 komi and black wins ties is not equivalent to 7.5 komi because of the way points in sekis are scored fractionally and can be shared asymmetrically, e.g if on board white has 40 and 2/3 points and black has 48 1/3 then white wins with 8 komi but loses with 7.5. Later versions of Ing rules abandoned this fractional points rule.
P. S there’s a double typo of handicap for komi in the quote.
If we were using the asymmetric-shared-points-in-seki counting, I guess we could set Ing komi to 0.01 (handicap) and 7.99 (even)… glad that we don’t need to account for it though .
Hello! I mostly play 9x9 games here, and I’m interested in what should be the correct komi for a 9x9 game (assuming no handicap or adjustment due to rank difference).
I’ve had a lot of experience playing games and afterward reading the computer analysis that this site provides – and as far as I can tell, the computer was saying that with a 5.5 komi, black started with a 0.5 point advantage. Which means, as far as I can tell, that black would probably win every time with perfect play.
However, now that you’ve switched to the 6.5 komi, it seems that white starts with a small advantage, which as far as I can tell might mean that white now wins every time with perfect play.
Can you try this by pitting strong AIs against each other? If I’m right, then it might prove that the only komi that leads to a draw with perfect play is 6.0.
And if so, shouldn’t we adopt that komi? Especially now that draws are an option via the same update, it seems even more opportune to do so!!
For the record, wasn’t it a bug in the old rating system which incorrectly accounted for handicap? Or was there also some bug in how it computed the stones?
And will the newly-calibrated rating calculations retroactively apply to old games when they eventually get recomputed?
It was how stones were computed, and yes plans are in the works for the next iteration of the rating system, as that rolls out old unfair games will be recomputed.
You’re right; from the data gathering I did, komis of 6 (territory) and 7 (area) are currently the best estimates for “perfect” komi. However, most rulesets by convention do not allow draws, and give an extra 0.5 as a tie-breaker. The change you’re seeing here brings small boards in line with 19x19.
Two options now if you want to play with komi that allows draws (“perfect” or otherwise):
Use the New Zealand ruleset, which allows draws by default.
Play unranked games () and set a custom komi.
That said, we might have more options in the future. The plumbing that @anoek and I are working on (to retroactively fix ratings due to old small-board handicap games) will in theory enable the ratings system to understand non-standard komi in general.
We can do a bit better than playing two strong AIs against one another, as doing that gives you some data about one possible path through the game but not all the other possible openings.
Here’s an opening book for Japanese-like rules https://katagobooks.org/book9x9jp/root/root.html on 6.0 komi that builds out a large tree of the plausible openings, using hundreds of billions (yes billions with a b) of visits. Branching very widely initially and narrowing down as you go deeper.
Used for the book generation is a special neural net that trained almost exclusively on 9x9 games for several months starting from a very wide range of different openings in those training games, and which improved enough on 9x9 to almost always win or draw against prior versions without that training (mostly draw, since prior versions still were pretty good) even when run with a modest computational disadvantage.
Funny thing though is that the special 9x9 net, despite such focused training, is still not that confident about some lines, or about winning/losing when komi is unfair by half a point. 9x9 is still very hard to solve! But yes 6.0 komi is the current best guess for fair komi for Japanese rules perfect play until someone pushes further with even more training and trillions of visits instead of hundreds of billions.