Statistically derived komidashi?

magic · #1

Has there ever been studied which komidashi leads to a win/loss rate for white that is the closest to 50/50 when two equally good players play against each other?

Bill Spight · #2

magic wrote:

Has there ever been studied which komidashi leads to a win/loss rate for white that is the closest to 50/50 when two equally good players play against each other?

Yes. There was an article in the American Go Journal in the 1970s based upon Japanese pro statistics that concluded that the proper komi was 7.

Ing proposed what is effectively a komi of 7.5 by area scoring. I rather expect that that was based upon statistics.

HermanHiddema · #3

http://senseis.xmp.net/?Komi#toc6

John Fairbairn · #4

Quoting from the GoGoD "History of Komi":

Quote:

In September 2005, the Nihon Ki-in carried out a survey of 1,042 completed games at the new 6.5 komi. It found that Black had won 50.51% and White 49.49%. This was the nearest to parity ever achieved. The equivalent figures in the Nihon Ki-in's last survey under 5.5 komi showed Black won 52.42%.

Figures from Japan in 2008 showed that the latest komi of 6.5 still leaves Black a slight edge. In 19,702 games from 6 November 2002 up to 31 December 2007, Black won 50.59% and White won 49.41%. The usual thinking in western forums has been that figures for the initial period after a komi adjustment may be a little unreliable as players try new strategies. To allow for that, we can look at figures for last year alone (i.e. 1 January 2007 to 21 December 2007): 4,014 games, Black 50.62%, White 49.38%. Figures for 1998 ~ 2003 at 5.5 komi were: 10,054 games. Black 52.4%, White 47.6%. In all periods, a breakdown of the figures suggests that Black has even more of an advantage in the later reaches of tournaments, i.e. the strongest players know how to use first mover better. In 2007's title matches the score was 23-19 to Black. In the period 2002-2007, Black won 51.34% of games at the final 9-dan dominated knockout level and 51.58% of games where the final stage was a league.

RobertJasiek · #5

A few years ago, at the International Go Rules Forum, similar statistics for Chinese or Korean professionals gave similar results of Black having a slight advantage, but closer to 50% than with earlier komi values. However, such values did not measure only games of equally strong players.

As we have discussed about 10 years ago, the komi increment was an improvement, but to really find out the proper komi only by empirical means, also greater komi values must be tested (exclusively or by bidding). (Non-statistical theory claims that 7 is the optimal integer komi.)

Bill Spight · #6

John Fairbairn wrote:

Quoting from the GoGoD "History of Komi":

Quote:

In September 2005, the Nihon Ki-in carried out a survey of 1,042 completed games at the new 6.5 komi. It found that Black had won 50.51% and White 49.49%. This was the nearest to parity ever achieved. The equivalent figures in the Nihon Ki-in's last survey under 5.5 komi showed Black won 52.42%.

Figures from Japan in 2008 showed that the latest komi of 6.5 still leaves Black a slight edge. In 19,702 games from 6 November 2002 up to 31 December 2007, Black won 50.59% and White won 49.41%. The usual thinking in western forums has been that figures for the initial period after a komi adjustment may be a little unreliable as players try new strategies. To allow for that, we can look at figures for last year alone (i.e. 1 January 2007 to 21 December 2007): 4,014 games, Black 50.62%, White 49.38%. Figures for 1998 ~ 2003 at 5.5 komi were: 10,054 games. Black 52.4%, White 47.6%. In all periods, a breakdown of the figures suggests that Black has even more of an advantage in the later reaches of tournaments, i.e. the strongest players know how to use first mover better. In 2007's title matches the score was 23-19 to Black. In the period 2002-2007, Black won 51.34% of games at the final 9-dan dominated knockout level and 51.58% of games where the final stage was a league.

Those statistics are in line with the theory that komi, in the sense of achieving a 50:50 result, increases with the skill of the players. It isn't just that Black knows how to exploit the first move, it is that White also plays better.

magic · #7

Interesting to ready what you have written.

RobertJasiek wrote:

to really find out the proper komi only by empirical means, also greater komi values must be tested (exclusively or by bidding).

That is why we should really store the score difference and the komi that was used for each played game, and not just who won. In that way it would be easy to calculate what the komi should have been in order to get the closest to a 50:50 ration between white's wins and black's wins, simply by calculating komi + black's score - white's score for each game and calculating the median of that. The median is then what komi really should be.

skydyr · #8

magic wrote:

Interesting to ready what you have written.

RobertJasiek wrote:

to really find out the proper komi only by empirical means, also greater komi values must be tested (exclusively or by bidding).

That is why we should really store the score difference and the komi that was used for each played game, and not just who won. In that way it would be easy to calculate what the komi should have been in order to get the closest to a 50:50 ration between white's wins and black's wins, simply by calculating komi + black's score - white's score for each game and calculating the median of that. The median is then what komi really should be.

The problem is that stronger players modify their play based on the komi, so you can't extrapolate from one komi to another without taking that into account.

EDIT:
Also, when you can count well and know what the score will be, you can play more conservatively to preserve a lead, resulting in a win by a small number of points where more were possible, to keep from making mistakes.

shapenaji · #9

If there were a large number of bangneki games available somewhere, that would be an interesting place to determine exact komi. Unfortunately, such things are never recorded.

Addendum:
Supposing that 7 is, in fact, the correct komi. Would it be worth allowing Jigo again to make the game fairer?

magic · **#10**

skydyr wrote:

The problem is that stronger players modify their play based on the komi, so you can't extrapolate from one komi to another without taking that into account.

EDIT:
Also, when you can count well and know what the score will be, you can play more conservatively to preserve a lead, resulting in a win by a small number of points where more were possible, to keep from making mistakes.

Ah, I see. So a good player who has the upper hand would rather play defensive to maximize his chance of winning, even if that meant that he would maybe win by 2 or 3 points less in average than if he would played more offensive?

Bill Spight · **#11**

magic wrote:

skydyr wrote:

The problem is that stronger players modify their play based on the komi, so you can't extrapolate from one komi to another without taking that into account.

EDIT:
Also, when you can count well and know what the score will be, you can play more conservatively to preserve a lead, resulting in a win by a small number of points where more were possible, to keep from making mistakes.

Ah, I see. So a good player who has the upper hand would rather play defensive to maximize his chance of winning, even if that meant that he would maybe win by 2 or 3 points less in average than if he would played more offensive?

Except that the player who is behind will play more enterprisingly, which may result in the loss of more points than would occur with perfect play.

magic · **#12**

Okay, so it can really swing in both directions. Hard to know in which way the komi will affect the median score difference then? (not including komi)

Mef · **#13**

shapenaji wrote:

If there were a large number of bangneki games available somewhere, that would be an interesting place to determine exact komi. Unfortunately, such things are never recorded.

Addendum:
Supposing that 7 is, in fact, the correct komi. Would it be worth allowing Jigo again to make the game fairer?

Large numbers of strong bangneki games would truly be a great source for evaluating all sorts of things in go. Of course creating, maintaining, and analyzing such a database runs counter to the interests of strong bangneki players...another one of life's conundrums.

Edit: though as I reconsider it, many of the benefits coming from such a database might be offset by noise/bad data coming from a classic sandbag hustle...nevertheless it would still be interesting to have...

skydyr · **#14**

magic wrote:

Okay, so it can really swing in both directions. Hard to know in which way the komi will affect the median score difference then? (not including komi)

Well, I think it's safe to say that increasing the komi, i.e. giving white a larger benefit, will benefit white after things have shaken out for a few years. The question is how much.

Arguably, the larger the komi, the more white is pressured to aim for a low scoring game, since then the komi will make up a higher percentage of the total score: if white and black score around 60 points combined, those 6.5 or 7 or however many points count for more than if both sides scored around 120 together.

Matti · **#15**

I think we also have enough data for 7½ komi. Does some one have them at hand and be able to tell whether black still has an advantage? If yes, then we would need, say 1000 games with 8½ or preferably 9½ to know better.

HermanHiddema · **#16**

Matti wrote:

I think we also have enough data for 7½ komi. Does some one have them at hand and be able to tell whether black still has an advantage? If yes, then we would need, say 1000 games with 8½ or preferably 9½ to know better.

At http://senseis.xmp.net/?Komi%2FStatistics there is a section I posted about 7.5 or 8.0 (Ing) komi, which has white winning 50.5% of games in a sample of 410 games from gobase.org where both players are at least 8d pro.

hyperpape · **#17**

shapenaji wrote:

Addendum:
Supposing that 7 is, in fact, the correct komi. Would it be worth allowing Jigo again to make the game fairer?

A good question, and I think you can make an argument that draws should be possible in some ideal sense. But if komi suffices to reduce the advantage to less than a 52% win percent (for either White or Black), that's excellent balance. I don't find it worth tolerating a 5% chance of draws in competition to remove that advantage. Perhaps if it were the old days in which a series could continue as long as it needed, that would be ok, but we already have scheduling conflicts.

(Btw: even allowing draws would not necessarily make the result 50-50. We could be talking about moving the needle by less than 2% advantage to a point that is still slightly unbalanced).

HermanHiddema · **#18**

shapenaji wrote:

Supposing that 7 is, in fact, the correct komi. Would it be worth allowing Jigo again to make the game fairer?

Depends on the circumstances. If would theoretically be fairer, but might have other downsides.

In round-robin leagues it would not have any real downsides.

In Swiss/MacMahon it would work, but introduce slightly more up/down pairings due to the rarity of players with half points.

In knock-out it could break the system (but see: ResolvingJigoByKoThreats)

ez4u · **#19**

HermanHiddema wrote:

Matti wrote:

I think we also have enough data for 7½ komi. Does some one have them at hand and be able to tell whether black still has an advantage? If yes, then we would need, say 1000 games with 8½ or preferably 9½ to know better.

At http://senseis.xmp.net/?Komi%2FStatistics there is a section I posted about 7.5 or 8.0 (Ing) komi, which has white winning 50.5% of games in a sample of 410 games from gobase.org where both players are at least 8d pro.

I am not sure why you think that 8p both sides is particularly relevant in the present pro environment (consider who gets excluded from the results). Anyway my copy of GoGoD (summer 2013) has 9,340 games with komi of 3.75 (i.e. Chinese rules with 7.5). White wins 52.3%. The not quite 19,000 games with komi of 6.5 have Black winning 50.2%

HermanHiddema · **#20**

ez4u wrote:

I am not sure why you think that 8p both sides is particularly relevant in the present pro environment (consider who gets excluded from the results). Anyway my copy of GoGoD (summer 2013) has 9,340 games with komi of 3.75 (i.e. Chinese rules with 7.5). White wins 52.3%. The not quite 19,000 games with komi of 6.5 have Black winning 50.2%

Is the weaker player (lower pro rank) more likely to get black in that sample? I wanted to avoid that option.

Statistically derived komidashi?

Who is online