Draws, handicap & komi changes, and ranking limits lifted

Announcements OGS Announcements
81

Is this 5.5 komi and Leela being the pre-zero AI as opposed to LeelaZero? I can kind of imagine the zero AI like katago might be able to play 9x9 better. Have you tried having katago play black and leela on your machine play white with 5.5 komi? It could be interesting to see if there’s any finer lines.

Or one could look at https://katagobooks.org/book9x9jp/root/root.html for example, that was linked by @hexahedron above

You can play vs low playout (2-50) katago on 9x9 on aisensei or you could try an unranked game vs katago-micro, or you could do a low playout review on aisensei with a free account.

Other than that installing katago/katrain etc, or possibly paying for some service that does ai reviews, ogs, zbaduk, aisensei etc.

1 Like
82

It’s Leelazero. Thanks for the recommendation. Will try it out. Have a feeling it’s insufficient due to the 6 komi setting. 0.5 komi is huge difference in it’s evaluation.

1 Like
83

Huh, I don’t think I remember a Leela Zero network being trained for 9x9. It sounds like it’s decently good?

Anyways, at very high levels of play for 9x9, fair komi is believed to be 6 on Japanese rules and 7 on Chinese rules, so switching Japanese 9x9 komi from 5.5 to 6.5 does change it to “unfairly” favor White, but at least at high levels of play it’s not clear that it favors White now any more than it used to “unfairly” favor Black at 5.5.

For 9x9 specifically I’d probably be in favor of making the default komi to be the fair integer komi (as I hear is the case on GoQuest). Unlike 13x13 or 19x19, on 9x9 the game is so much tighter that the half point offset from theoretical fair really can be meaningful. Particularly for people who’ve studied opening lines, I would guess that high-level human players can occasionally play literally perfect games, as in literally every move the player plays is game-theoretically win/loss/draw-optimal.

11 Likes
84

As @hexahedron mentioned, best guesses for “perfect” komi, for all board sizes, are:

  • 6 for territory scoring
  • 7 for area scoring

Convention in most rulesets is that ties go to white, so 6.5/7.5 it is. Ing SST rules dictate 8 komi with ties going to black (still 7.5). New Zealand rules allow draws and use 7 komi (maybe you’d like to try that out).

It seems like some people find 9x9 harder as black, others harder as white. At 6.5, it does favour white by half a point (whereas 5.5 favoured black). The in-development ratings system change will include komi in ratings update to model this (dis)advantage.

6 Likes
85

I reckon this (dis)advantage is not very big for lower rated players using territory scoring on 19x19 boards.

In points, ranks on 19x19 are spaced by a full handicap stone, which corresponds to 2 times komi (~13 points), so the (dis)advantage of 0.5 points is about 4% of a rank.

Around 1d EGF, the value of a full handicap stone is about 100 Elo points, so for equally rated players of that level, this point (dis)advantage corresponds to about 4 Elo points shift, which corresponds to about 0.5 percentage points winrate shift away from 50%.

For equally rated DDK players (handicap stone value about 33 Elo points for EGF ranks), a 0.5 point (dis)advantage may correspond to about 0.2 percentage point winrate shift, and for pros (handicap stone value about 300 Elo points for EGF ranks) it may be about 1.7 pp.

At 9x9, when using 6 ranks per full handicap stone, 0.5 points may correspond to about 3.5 percentage point winrate shift for 1d level, about 1.2 pp for DDK level and about 10 pp for pro level.

With area scoring, the score resolution is about twice as coarse as with territory scoring, so you probably need to double the Elo shifts. In that case, using 7.5 instead of 7 komi may shift the winrate on 9x9 boards in white’s advantage by about 6.5 pp for 1d level, 2.1 pp for DDK level and about 19 pp for pro level.

So I think it is a good idea to account for a 0.5 point advantage in the rating system. It will make the rating system more accurate, especially for stronger players using area scoring on smaller boards.

3 Likes
86

Hallelujah for finally fixing komi – now OGS may become my favorite server!

4 Likes
87

According to Katago the correct komi for all odd-sized boards above 7x7 is 6 for territory scored games and 7 for area scored games. This means that default Japanese and Chinese rules favor white by making ties their victory (white has actually had a demonstrably better win rate than black in professional games in recent years, potentially due to this) So NZ is probably the fairest ruleset in this server (area scoring with 7 komi) – unfortunately I don’t think there is a ruleset with territory scoring and fair komi.

4 Likes
88

With Japanese rules, the size of komi is only weakly linked to the rules.

The official Japanese rules don’t even mention the size of komi, and dates where komi was updated (to 4.5 during the 30s and 40s, gradually shifting to 5.5 during the 50s and 60s, switching to 6.5 in 2002) don’t coincide with dates where the rules were updated (1949, 1989).

When playing under Japanese rules in Europe, it’s customary to follow komi as used in concurrent Japanese pro tournaments, but some European tournaments have been played with integer komi.

6 Likes
89

Just looking at their own native score rating already shows white winning by half point right from the get go so I don’t know how they can believe this is fair scoring.

90

It’s going to be either white winning by half a point from the get-go, or black winning by half a point from the get-go, unless we used an integer komi of 6.

You can set the komi to 6 for unranked games: I guess you’re arguing that we ought to be able to have ranked games by Japanese rules with komi of either 6.5 (traditional) or 6 (what you seem to be suggesting). Or have I misunderstood?

3 Likes
91

Yes I suggest we go for the whole number komi route just for 9x9. Otherwise the komi increase is meaningless. If you’re going for fairness but just arbitrarily deciding on which side to favour then it’s as good as zero change.

92

Why not integer komi (6.0 for territory rules; 7.0 for area rules) in rated games for all 3 standard sizes (obviously just as an option for rulesets which don’t usually have it: the change would be in teaching the rating system how to interpret the advantage of custom komi correctly and then allowing rated games to use custom komi)? I hope that becomes possible with some of the talk that’s been going around about rating system and handicap updates

1 Like
93

I suggest to allow any komi that is more than -1 and less than 10 in ranked games. Ranked system will calculate them differently than standard even games just like it do it for handicap games.
Standard komi should still be default in ranked games.

1 Like
94

I’d make the range [0, 2n] where n is the default komi for that ruleset, so the range is 1 stone. Of course, I’d really prefer to support reverse komi at least up to 9 stones, but if we’re going for some smaller range, I think increments of komi make more sense

95

Allowing too big difference from standard may lead to problems similar to 2 minutes absolute time control in 19x19 ranked game.

96

Komi 0 and standard komi works ok in practice. Therefore anything between it works ok too. But its unknown if most users would be able to adapt to numbers of komi beyond that range.

1 Like
97

Another komi feature that i would enjoy to try on OGS is the “pie” komi

I propose a komi and you chose the color.

2 Likes
98

I don’t think that’s that interesting since fair komi is all but known now for standard board sizes with strong AI

99

I think the fault there lies with 2 minutes absolute time. I have nothing against players playing it, but if it doesn’t work with too extreme custom komi, I think the solution if you like those time controls, is to not play with very extreme custom komi

100

I think the same could be said of Handicap stones. Some players will handle it better than others, but they will either learn, or play even