Komi power

[SmoothOper]

I wonder how accurate the statistics are for komi. Are there enough games in the pros to discriminate between a 6.5 9dan and 5.5 9dan pro, or is increments of +-6.5 as much statistical power as they can get.

[Shaddy]

I think I remember seeing that at 5.5 komi, the winrate for Black was ~51%, and at 6.5, it's ~49? I don't remember where I saw these numbers, though.

[SmoothOper]

I think I remember seeing that at 5.5 komi, the winrate for Black was ~51%, and at 6.5, it's ~49? I don't remember where I saw these numbers, though.

That is an over all komi, I guess I was thinking about komi as a fine grained ranking system. For example some 9 dan only win like 30% of there games against other 9 dan. What I am proposing is that maybe there are enough games to say statistically that while they are definitely stronger than any 8dan, they can only win with maybe 4.5 komi as black against a stronger 9dan. Statistically I am not sure there are enough games to make such a distinction. I see this on go servers where IGS will add or subtract increments of as much komi as 6.5 for weaker and stronger players of the same rank.

[Polama]

I wonder how accurate the statistics are for komi. Are there enough games in the pros to discriminate between a 6.5 9dan and 5.5 9dan pro, or is increments of +-6.5 as much statistical power as they can get.

Professional Rankings don't work like amateur ones. First, they measure accomplishment rather than skill: You can have very strong 2-dan who haven't had enough time to climb the ranks further yet, and 9-dan who are no longer playing at their peak ability.

Secondly, they don't use the 1 rank=1 stone handicap rule amateurs do. A professional rank is only worth a fraction of a stone. So the difference between two ranks is already just a question of a couple points of komi.

[SmoothOper]

I wonder how accurate the statistics are for komi. Are there enough games in the pros to discriminate between a 6.5 9dan and 5.5 9dan pro, or is increments of +-6.5 as much statistical power as they can get.

Professional Rankings don't work like amateur ones. First, they measure accomplishment rather than skill: You can have very strong 2-dan who haven't had enough time to climb the ranks further yet, and 9-dan who are no longer playing at their peak ability.

Secondly, they don't use the 1 rank=1 stone handicap rule amateurs do. A professional rank is only worth a fraction of a stone. So the difference between two ranks is already just a question of a couple points of komi.

I am aware of the inflation problem as well as the timing problem, my question is are there enough games to discern the difference of a few points of komi in strength.

[Magicwand]

depended on how much time given percentage differ.
for games 3 hours + white wins more game
for 1hour - black seem to win more.


it also depended on how strong you are and style you prefer.
amatures prefer black when komi is 6.5
strong pros prefer white when komi is 6.5

[SmoothOper]

depended on how much time given percentage differ.
for games 3 hours + white wins more game
for 1hour - black seem to win more.


it also depended on how strong you are and style you prefer.
amatures prefer black when komi is 6.5
strong pros prefer white when komi is 6.5

Are there any categories that have enough power? At a certain point I think the flavor in their morning bowl of miso soup may make a difference.

[Uberdude]

strong pros prefer white when komi is 6.5

I guess Gu Li is a weak pro then...

[billywoods]

Why don't you go do the statistics yourself?

A while back, I looked at the first 400 games on gokifu.com. Over 300 of them were resigned before scoring, with only about 80 making it through to scoring, and 15 of those 80 were wins (for both black and white fairly equally) by more than 7 points even after komi - that is, wins in which komi didn't matter anyway. Absurd results like B+20 and W+30 were rare, but happened. I then took the average score difference, and compared the black/white win ratio, and got pretty much exactly what you'd expect if komi was correct, but it was such a small sample size with such varied results that I can't convincingly say it wasn't an accident.

[speedchase]

and 15 of those 80 were wins (for both black and white fairly equally) by more than 7 points even after komi - that is, wins in which komi didn't matter anyway.

That is just about the most dubious claim ever.

[Bantari]

Why don't you go do the statistics yourself?

A while back, I looked at the first 400 games on gokifu.com. Over 300 of them were resigned before scoring, with only about 80 making it through to scoring, and 15 of those 80 were wins (for both black and white fairly equally) by more than 7 points even after komi - that is, wins in which komi didn't matter anyway. Absurd results like B+20 and W+30 were rare, but happened. I then took the average score difference, and compared the black/white win ratio, and got pretty much exactly what you'd expect if komi was correct, but it was such a small sample size with such varied results that I can't convincingly say it wasn't an accident.

Two issues I can see here:
1. How many of the resignation games would not have been resigned if komi were different? Hard to tell...
2. How did size of komi influence the results in games won/lost by 7 or more points? Or any game for that matter? Hard to tell...

I am not saying that the answer to these questions would or would not influence your final conclusion - I just don't know. All I am saying is that it disagree with just dismissing these two cases in your calculations as you apparently did. I think it might be slightly less simplistic than just looking at the surface numbers. Better approximation would be to look at all games, and only compare win/loss ratio, regardless of size of result.

[billywoods]

All I am saying is that it disagree with just dismissing these two cases in your calculations as you apparently did.

I didn't dismiss them. We all know what happens when you count the number of wins and losses, but I find that dull - I wanted to know some more honest raw numbers. But because of issues like resignation and psychological factors, it seems next to impossible. The data are kind of poor. That's my main conclusion here.

[Uberdude]

Resignations don't mean komi wasn't relevant. You can count that you lose by half a point with normal yose and thus start some complications which don't quite work but give the opponent an opportunity to make a mistake: finding a place to resign. Here's my traditional illustrative OGS game (you might recognise the problem, I posted it here before): white could have spent a defensive move inside his territory but then he would lose by half a point, so he left it and I found a sequence inside his territory which worked so he resigned.

[Bill Spight]

First, komi (at least in Japan) was not historically decided by statistics. Modern komi started out at 4.5 points. Then for several years some events used 4.5 komi, while some used 5.5 komi. By the mid-1970s the shift was complete to 5.5 komi, AFAIK. In 76 or 77 Terry Benson asked me to review an article for the AGA Journal before publication, which claimed that the correct komi was 7. (At that time I had already predicted that Japanese komi would be 6.5 by the year 2000. Oops!) I made the comment that, while the stats suggested that 5.5 komi was too small, the current attitude was that Black had to be enterprising to overcome komi. A change in attitude to the effect that White is the one who needs to be enterprising, even with komi, might make a difference in the stats.

Terry sent me the data upon which the article was based, the results of 1400 pro games at 5.5 komi and 1400 pro games at 4.5 komi. To my surprise, the stats for the 4.5 komi games suggested a komi of 6.5. That being the case, it is surprising that the 4.5 komi lasted as long as it did.

[SmoothOper]

First, komi (at least in Japan) was not historically decided by statistics. Modern komi started out at 4.5 points. Then for several years some events used 4.5 komi, while some used 5.5 komi. By the mid-1970s the shift was complete to 5.5 komi, AFAIK. In 76 or 77 Terry Benson asked me to review an article for the AGA Journal before publication, which claimed that the correct komi was 7. (At that time I had already predicted that Japanese komi would be 6.5 by the year 2000. Oops!) I made the comment that, while the stats suggested that 5.5 komi was too small, the current attitude was that Black had to be enterprising to overcome komi. A change in attitude to the effect that White is the one who needs to be enterprising, even with komi, might make a difference in the stats.

Terry sent me the data upon which the article was based, the results of 1400 pro games at 5.5 komi and 1400 pro games at 4.5 komi. To my surprise, the stats for the 4.5 komi games suggested a komi of 6.5. That being the case, it is surprising that the 4.5 komi lasted as long as it did.

I think you could begin to use data from the different eras to construct the case for a komi based handicap system. Say, by looking at the statistics for weaker pros that had strong winning records as black with 4.5 komi, but then declined when they had to win with 5.5 komi. IE some players were 9 dan 4.5 komi as black and some were 9 dan 5.5 komi as black. Of course players strengths varying with age would have to be taken into account, as well as the number of games on record for each player.