The size of komi

[sorin]

I was very surprised to read this overview of 2018 with respect to komi - this is Nihon Ki-in reporting on their Facebook page:

"In 2018, 4,208 official pro games were played in total with Komi 6.5 points in Japan and the winning percentage of white is 50.48% while black is 49.52%.
This case (a winning percentage in White became higher than Black) only happened four times (2005, 2015, 2016 and 2018) since Japan changed 6.5 points (from 5.5 points) in 2002.
For your reference, in the case of Iyama Yuta of 2018, his result of Black is 13 wins -16 losses while his result of White is 22 wins -10 losses.
The followin question may arise: if Komi 6.5 points should be changed to 5.5 points?"

The first thing I found surprising is that in most years black still wins more games than white - I wish they published the overall numbers since this may be a weird artifact of dividing the data into years.

Second surprising thing is the last thought in their report - it never occurred to me that people would want to go back to 5.5 komi.

[macelee]

I checked the Go4Go database. For all 6.5-komi games played in 2018, black won 1596 (49.67%) and white won 1617 (50.33%). It may be the case that playing the AI way makes white slightly easier than before. But I think this sample size is too small. Even if this statistics is accurate, I don't think the gap is big enough to justify switching back to 5.5 komi.

[ez4u]

Translated and slightly rearranged from Mamumamu's Analysis of 6.5 Komi (see table 2 there).
http://sports.geocities.jp/mamumamu0413 ... .html#res3

Code: Select all

Year	Games	Black 	White 	Black Win %
2002～
2003	 3,645 	 1,836 	 1,809 	50.37%
2004	 3,680 	 1,872 	 1,808 	50.88%
2005	 3,864 	 1,906 	 1,958 	49.33%
2006	 3,375 	 1,736 	 1,639 	51.44%
2007	 3,469 	 1,767 	 1,702 	50.94%
2008	 3,500 	 1,820 	 1,680 	52.00%
2009	 4,126 	 2,075 	 2,051 	50.29%
2010	 3,820 	 1,927 	 1,893 	50.45%
2011	 3,747 	 1,886 	 1,861 	50.33%
2012	 3,746 	 1,943 	 1,803 	51.87%
2013	 4,020 	 2,013 	 2,007 	50.07%
2014	 3,922 	 1,992 	 1,930 	50.79%
2015	 4,740 	 2,368 	 2,372 	49.96%
2016	 4,579 	 2,340 	 2,239 	51.10%
2017	 4,325 	 2,178 	 2,147 	50.36%
Total  58,559    29,660	 28,899	 50.65%

Interesting that his figures disagree with the Nihon Ki'in for 2016. Mamumamu wrote that he used all domestic games in his analysis. The NK may be including international games as well.

[John Fairbairn]

Mamumamu's data relates to the Nihon Ki-in only, it seems. The Nihon Ki-in has discussed its own figures many times and has postulated various factors that affect the results.

The biggest factor is, of course, the size of the komi - over the long run. When komi was changed from 5.5. to 6.5, White was favoured by about two percentage points in the Nihon Ki-ins own survey (1st survey at 6.5, with just over 1,000 games: 50.5%, down from 52.4% in last multi-year survey at 5.5).

Actually it seems that the biggest factor in favouring Black is short time limits, and the shorter the better from Black's point of view. At faster even than NHK style fast limits, Black's winning rate goes up to over 56%. The mice are dancing!

The NK has also often suggested that the strength of the players may be a factor - the stronger the player the more he knows he to use komi. Certainly there has been an edge in favour of Black the further into a tournament you go (i.e. in the leagues and title matches Black does better than in the preliminaries). This possibly dubious metric has, however, become otiose in recent years as the higher dans are no longer automatically seeded into the higher reaches of tournaments.

They have in the past also argued that figures may be slightly warped as players adapt their styles to a higher komi, but that putative effect must have worn away by now (6.5 komi s introduced in Japan in mid -2003).

But there does seem to be another factor totally ignored by the NK - country.

If you look at the GoGoD database figures for results in ALL countries combined (Japan, Korea, China, Taiwan, and ignoring a very small number of blips such as void games), we get a chiaroscuro picture rather than the mostly oscuro one of Japan.

Looking at almost 48,000 games since 2004 onwards, White mostly dominates, and if you take out the Japanese games and their bias towards Black, games elsewhere seem to favour White consistently. (The komis here cover 6.5, 3.25 zi and Ing komis.)

Year White %
---- -------
2004 52.6%
2005 50.3%
2006 51.5%
2007 50.5%
2008 50.5%
2009 48.2%
2010 49.5%
2011 52.1%
2012 50.0%
2013 50.6%
2014 50.7%
2015 50.6%
2016 50.6%
2017 51.8%
2018 52.8%

The 2017-2018 figures may indicate an upward tick because of AI, but then how do we explain 2011?

A more general argument may hold water better: that a change in playing style (of which copying AI is just one manifestation) is enough to affect overall results, and the go world is nothing if not fashion conscious!

And how do we explain Japan apparently being different from the rest of the go world? The most obvious difference may be that Japan retains longer time limits, but their own (admittedly small-scale) figures suggest contrasting things at the two extremes, as I recall: blitz limits strongly favour Black but Black also has quite a big edge in major title games.

To my untutored eye it all doesn't yet look much different from flipping a coin.

[jlt]

In John Fairbairn's data, assuming there were 3200 games each year, White's observed winrate was 50.8%. When we toss a fair coin 48000 times, the probability to get heads at least 50.8% of the time is about 1:6000, so 6.5 komi probably favors White.

When we look at series of 3200 games, assuming a mean of 50.8%, the standard deviation is 0.9%. The value in 2011 is 52.1% = 50.8% + 1.3%, still well within two standard deviations, so should be considered as normal fluctuation.

[John Fairbairn]

assuming there were 3200 games each year,

I don't understand much of the "statistics, lies and damned lies" business, but if it does matter, the annual data follows a rising trend in number of games (from 2130 to 4732 a year. For the last two years the figures are 4446 and 4732, so the proportion of games where White had a larger than usual winning percentage is a bit larger. I assume that might matter but I've no idea how.

[jlt]

It doesn't change much. If the rise is approximately linear, then White's winrate is still between 50.8% and 50.9%, and it is even less likely that 6.5 points komi is fair.

[moha]

Actually it seems that the biggest factor in favouring Black is short time limits, and the shorter the better from Black's point of view.

This is interesting. If faster games would simply equal lower playing quality, then the opposite would be less surprising, since fair komi is smaller for weaker players than for perfect play (what W gets is constant, what B can work against it is less).

Going against this effect may have to do with the playing style being largely different in faster games? But if such factors have this much effect then the original slight white advantage is not surprising either.

[mhlepore]

I'm mixing threads here a bit, but in the "What a crying shame" discussion, did anyone else wonder why Cho Hye-yeon, when posed with the question of how much clock time a single point is worth, did not say "depends, am I black or white?"

Does she think komi is so perfectly dialed in that it doesn't matter? Or did she not think things out completely? Or something else?

[Uberdude]

I think she was making a dramatic point without enough qualification of assumptions and context. Her answer is nonsense in a 1 hour sudden death game

[sorin]

Her answer is nonsense in a 1 hour sudden death game

... which is not something we need to worry about, because we play go, not chess

[BlindGroup]

Even if it turns out that 6.5 komi favors white, the trickier question is whether it can be fixed by reducing komi. To know the answer to that, one has to know the relationship between komi and the win rate for white. From what John writes, it looks like the initial increase in komi by 1 point (5.5 to 6.5) changed the win rate by 2%. So, if the win rate for white is only slightly above 50 -- say 50.8% as mentioned above -- changing komi by one point won't make the game more even. However, even this assumes that this single observation of pre-AI data still holds today.