Life In 19x19
http://www.lifein19x19.com/

Fair komi without ties
http://www.lifein19x19.com/viewtopic.php?f=45&t=12121		 Page 2 of 5
Author:  pookpooi [ Tue Oct 24, 2017 9:32 am ]
Post subject:  Re: Fair komi without ties
We've to get stats on how much 6.5 komi makes different outcome than 5.5 komi.

From stat we've now 5.5 komi has 2.84% deviate from 50% and 7.5 komi has 2.50%
If 6.5 komi affect only 0.1% than 2.84-0.1 = 2.74% which is still more deviate than komi 7.5. So komi 7.5 is still 'more fair' in this sense.

But of course, the real work (if Weiqi Association really want to change komi again) will be much more complicated than my math.
Author:  luigi [ Tue Oct 24, 2017 10:04 am ]
Post subject:  Re: Fair komi without ties
pookpooi wrote:
We've to get stats on how much 6.5 komi makes different outcome than 5.5 komi.

From stat we've now 5.5 komi has 2.84% deviate from 50% and 7.5 komi has 2.50%
If 6.5 komi affect only 0.1% than 2.84-0.1 = 2.74% which is still more deviate than komi 7.5. So komi 7.5 is still 'more fair' in this sense.

But of course, the real work (if Weiqi Association really want to change komi again) will be much more complicated than my math.

So no chance for 7 plus button, huh?
Author:  HermanHiddema [ Tue Oct 24, 2017 2:17 pm ]
Post subject:  Re: Fair komi without ties
billyswong wrote:
Herman's explanation further strengthen my suspicion that 6.5 komi may be a reasonable komi value for area scoring, while Luigi's explanation repeated what I learnt from the Sensei wiki I linked.

6.5 komi may be less "fair" than 7.5 komi, maybe not. But assuming 6.5 komi is exactly as fair/unfair as 5.5 komi sounds strange, as professionals fine tune their game play in accordance to komi value, maybe 6.5 komi games will contain seki more often? If statistics really shows that 5.5 and 7.5 are about equal, as said by Herman, there's a chance, however slight, 6.5 may be fairer.


The reason I said 5.5 and 7.5 are "about equally unfair", is because theoretically, if 7 is the perfect komi, then 7.5 and 6.5 are "exactly equally unfair", because they exactly award the perfect jigo games to one player or the other. However, the difference between 5.5 and 6.5 is so small under area scoring, that it doesn't really matter if you use 5.5 or 6.5, the difference between those is negligible.

So is 6.5 a reasonable value? Yes, as reasonable as 7.5.

Or 5.5 or 8.5 really. All of them swing the 7 point games one way or the other, without influencing the 5 point or 9 point games. I really don't think professionals fine-tune their games to the komi that much, and certainly not to an extent that it would influence the amount of sekis.
Author:  Bill Spight [ Tue Oct 24, 2017 2:33 pm ]
Post subject:  Re: Fair komi without ties
luigi wrote:
It's also not surprising that White's advantage with 7.5 komi under Chinese rules is more apparent in AlphaGo games than it is in human games, as AlphaGo is closer to perfect play.


I am not sure that we can draw any conclusion about komi from the published AlphaGo Master self play games, because nobody has said that they were randomly chosen from among the millions of such games. They may have been chosen because Fan Hui liked them, who knows?
Author:  Bill Spight [ Tue Oct 24, 2017 2:37 pm ]
Post subject:  Re: Fair komi without ties
HermanHiddema wrote:
I really don't think professionals fine-tune their games to the komi that much, and certainly not to an extent that it would influence the amount of sekis.


I agree. I haven't seen any evidence that they do such fine tuning. Sure, Black and White played differently when there was no komi, but the difference of around 7 pts. was huge.
Author:  John Fairbairn [ Tue Oct 24, 2017 3:04 pm ]
Post subject:  Re: Fair komi without ties
Quote:
I haven't seen any evidence that they do such fine tuning


There are quite frequent references in the commentary literature to the "large komi" and the effect one's play. There was also the famous case of Kobayashi Koichi refusing to play with the Ing komi because he said it was too large. Currently we are seeing Ke Jie favouring White (as quite a few other players seem to do).

This may not be the same as fine tuning, but at least it shows awareness of a potential effect of komi size.

Even so, as you imply, it doesn't seem to be an issue among pros. If it becomes one, komi bidding may be an (awful) solution. There is also the possibility of resurrecting the old Meijin rule of using an integer komi and declaring jigos a win for White, with the proviso that in an event like a league, where ties can occur in the final placing, jigo wins count less than other wins. Another extreme would be to accept jigos and have a replay (as occurred last week with a quad ko).
Author:  pookpooi [ Tue Oct 24, 2017 5:15 pm ]
Post subject:  Re: Fair komi without ties
Bill Spight wrote:
I am not sure that we can draw any conclusion about komi from the published AlphaGo Master self play games, because nobody has said that they were randomly chosen from among the millions of such games. They may have been chosen because Fan Hui liked them, who knows?


"In my experience and the experiments we've run, komi 7.5 is very balanced, we only observe a slightly higher winrate for white (55%)."

Source from DeepMind's Julian Schrittwieser

https://www.reddit.com/r/MachineLearnin ... ittwieser/
Author:  Bill Spight [ Tue Oct 24, 2017 5:51 pm ]
Post subject:  Re: Fair komi without ties
John Fairbairn wrote:
Quote:
I haven't seen any evidence that they do such fine tuning


There are quite frequent references in the commentary literature to the "large komi" and the effect one's play.


But that does not mean that the changes in play have affected the results.

We saw this before with the 4.5 and 5.5 komis, decades ago. With the 4.5 komi the median result on the board was between 6 and 7 for Black, and with the 5.5 komi the median result was between 6 and 7. No fine tuning.
Author:  jeromie [ Tue Oct 24, 2017 5:52 pm ]
Post subject:  Re: Fair komi without ties
I can understand the desire to eliminate ties in a tournament setting, but I wish more amateurs (and particularly online servers) use an integer Komi and allow a tie result. It seems to produce the most balanced game, and, unlike chess, go is unlikely to see a high number of results end in a tie. I think a tie game is an exciting outcome in a tense match.
Author:  Bill Spight [ Tue Oct 24, 2017 5:57 pm ]
Post subject:  Re: Fair komi without ties
pookpooi wrote:
Bill Spight wrote:
I am not sure that we can draw any conclusion about komi from the published AlphaGo Master self play games, because nobody has said that they were randomly chosen from among the millions of such games. They may have been chosen because Fan Hui liked them, who knows?


"In my experience and the experiments we've run, komi 7.5 is very balanced, we only observe a slightly higher winrate for white (55%)."

Source from DeepMind's Julian Schrittwieser

https://www.reddit.com/r/MachineLearnin ... ittwieser/


So they ran experiments and did not just rely upon the published games, which, IIUC, have an even higher win rate for White.

Anyway, calling 55% a slightly higher win rate suggests that Schrittwieser is not a go player. That's high enough, with human players, to suggest a komi of 5.5. I suppose, however, that in their experiments a 5.5 komi produced an even higher win rate for Black. :-| Button go, anyone? :cool:
Author:  moha [ Tue Oct 24, 2017 6:10 pm ]
Post subject:  Re: Fair komi without ties
Bill Spight wrote:
Button go, anyone? :cool:
Komi 7 simply.
Author:  luigi [ Wed Oct 25, 2017 1:31 am ]
Post subject:  Re: Fair komi without ties
Bill Spight wrote:
Button go, anyone? :cool:

I've mentioned this multiple times in this thread already. :D It would be the perfect solution, but no one seems excited about it.
Author:  John Fairbairn [ Wed Oct 25, 2017 2:03 am ]
Post subject:  Re: Fair komi without ties
Quote:
With the 4.5 komi the median result on the board was between 6 and 7 for Black, and with the 5.5 komi the median result was between 6 and 7. No fine tuning.


Bill: I'm not sure what you are saying here, but my first impression is that this only shows that pros tend to count out the game when the result falls within a certain band. You need to include resigned games to complete the picture.

On that point, I toss out some stats from GoGoD for some of the various komis (from year dot to 2017-10, jigos excluded):

No komi
10702 games
W+ 4032 (of which W+R = 1989, or 49.3% of W+)
B+ 5612 (of which B+R = 2932, or 52.3% of B+)
W+ 37.7% vs B+ 52.4%

2.5 Japanese komi
362 games
W+ 149 (W+R= 84 or 56.4%)
B+ 180 (B+R = 102 or 56.7%)
W+ 41.1% vs B+ 49.7%

3.5 Japanese komi
60 games
W+ 28 (W+R = 15 or 53.6%)
B+ 32 (B+R = 20 or 62.5%)
W+ 46.7% vs B+ 53.3%

4.5 Japanese/Korean komi
2685 games
W+ 1208 (W+R = 645 or 53.4%)
B+ 1477 (B+R = 928 or 62.8%)
W+ 45.0% vs B+ 55.0%

5.5 Japanese/Korean komi
21953 games
W+ 10347 (W+R = 5544 or 53.6%)
B+ 11563 (B+R = 6765 or 58.5%)
W+ 47.1% vs B+ 52.7%

6.5 Japanese/Korean komi
26521 games
W+ 13133 (W+R = 8814 or 67.1%)
B+ 13365 (B+R = 9080 or 68.0%)
W+ 49.5% vs B+ 50.4%

2.75 Chinese komi (said to = Japanese 5.5)
7471 games
W+ 3499 (W+R = 2355 or 67.3%)
B+ 3972 (B+R = 2902 or 73.1%)
W+ 46.8% vs B+ 53.2%

3.75 Chinese komi (said to = Japanese 6.5)
15496 games
W+ 8160 (W+R = 6425 or 78.7%)
B+ 7327 (B+R = 5792 or 79%)
W+ 52.7% vs B+ 47.3%

As regards jigo: when komi was 0 there were 274 jigos out of 10702 games (2.6%), and when komi was 5 there were 9 jigos out of 962 games (0.9%).

Others can interpret these data better than I, but it seems obvious that increasing the komi has increasingly benefited White (the slight blip for 3.5 may be ignored for too few games).

However, the sudden leap in White's favour for 3.75 Chinese komi seems potentially anomalous, at least when compared to Japanese/Korean 6.5, and also considering that JK 5.5 vs Chinese 2.75 gives similar results. I can offer no explanation for this, though it does seem to gel with what Chinese pros say about White's advantage. However, in my innumerate naivety I can't help but notice that 2 x 2.75 = 5.5 whereas 2 x 3.75 = 7.5 and not 6.6. In any case, Japanese/Korean 6.5 komi produces the closest B+ vs W+.

The small number of jigos surprised me.

It is noticeable that Black always seems more likely to resign, but Chinese players are very significantly more inclined to resign (with either colour) than JK players. As I've suggested before (on the basis of talking to Chinese players) this may have something to do with the tedium/inconvenience of counting up the Chinese way, but it's become more marked under the larger (?more unbalanced) komi so it may also have something to do with that.
Author:  HermanHiddema [ Wed Oct 25, 2017 2:14 am ]
Post subject:  Re: Fair komi without ties
John Fairbairn wrote:
3.75 Chinese komi (said to = Japanese 6.5)

...

However, the sudden leap in White's favour for 3.75 Chinese komi seems potentially anomalous, at least when compared to Japanese/Korean 6.5, and also considering that JK 5.5 vs Chinese 2.75 gives similar results. I can offer no explanation for this, though it does seem to gel with what Chinese pros say about White's advantage. However, in my innumerate naivety I can't help but notice that 2 x 2.75 = 5.5 whereas 2 x 3.75 = 7.5 and not 6.6. In any case, Japanese/Korean 6.5 komi produces the closest B+ vs W+.


I'm curious who's claiming Chinese 3.75 = Japanese 6.5. Chinese 3.75 should be equivalent to 7.5 Japanese, not 6.5 Japanese (which would be 3.25 Chinese).

Perhaps some people were confused by the fact that Japan went from 5.5 to 6.5 and China went from 2.75 to 3.75 at around the same time?
Author:  Bill Spight [ Wed Oct 25, 2017 8:20 am ]
Post subject:  Re: Fair komi without ties
John Fairbairn wrote:
Quote:
With the 4.5 komi the median result on the board was between 6 and 7 for Black, and with the 5.5 komi the median result was between 6 and 7. No fine tuning.


Bill: I'm not sure what you are saying here, but my first impression is that this only shows that pros tend to count out the game when the result falls within a certain band. You need to include resigned games to complete the picture.


I am including resigned games. The median (50-50 point) does that. :)

My guess is that the changes in strategy based upon changes in komi produce more resignations than small losses, which I think is something that you have in mind. However, that may well not change the median, if it is close to komi. The changes in strategy need to change small losses into wins to be effective.
Author:  dfan [ Wed Oct 25, 2017 11:56 am ]
Post subject:  Re: Fair komi without ties
pookpooi wrote:
"In my experience and the experiments we've run, komi 7.5 is very balanced, we only observe a slightly higher winrate for white (55%)."

Source from DeepMind's Julian Schrittwieser

https://www.reddit.com/r/MachineLearnin ... ittwieser/

That's a pretty high win rate! 55-45 is just about the same as White's advantage in chess (counting draws as a 0.5-0.5 tie), and everyone likes to point to that as one of the disadvantages of chess.
Author:  billyswong [ Fri Oct 27, 2017 4:13 am ]
Post subject:  Re: Fair komi without ties
luigi wrote:
Bill Spight wrote:
Button go, anyone? :cool:

I've mentioned this multiple times in this thread already. :D It would be the perfect solution, but no one seems excited about it.

I guess button go being unpopular is because upon introduction of a new object "button", people think of the game as a go-variant, not seeing it a 'proper' ruleset of go anymore, so it's a very perceptual thing.

So Luigi your first proposal "first passer wins ties" may be more likely to get considered seriously into common use. As long as a pass lifts ko bans, I guess it will be okay?
Author:  Bill Spight [ Fri Oct 27, 2017 5:47 am ]
Post subject:  Re: Fair komi without ties
billyswong wrote:
luigi wrote:
Bill Spight wrote:
Button go, anyone? :cool:

I've mentioned this multiple times in this thread already. :D It would be the perfect solution, but no one seems excited about it.

I guess button go being unpopular is because upon introduction of a new object "button", people think of the game as a go-variant, not seeing it a 'proper' ruleset of go anymore, so it's a very perceptual thing.


I don't think that button go is unpopular; people are just unaware of it.

The equivalent of button go has already been played in international competition. IIUC, go in the Mind Sport Olympiad uses area scoring but if White passes first White gets an extra point.

It is easy to modify AGA scoring to incorporate the equivalent of a button. Under current rules White must pass last, handing over a pass stone. Under the equivalent of button go White must pass last only if Black made the first pass, otherwise Black must pass last. OC, with the pass stones, territory counting is used.
Author:  luigi [ Fri Oct 27, 2017 6:18 am ]
Post subject:  Re: Fair komi without ties
Bill Spight wrote:
I don't think that button go is unpopular; people are just unaware of it.

The equivalent of button go has already been played in international competition. IIUC, go in the Mind Sport Olympiad uses area scoring but if White passes first White gets an extra point.

It is easy to modify AGA scoring to incorporate the equivalent of a button. Under current rules White must pass last, handing over a pass stone. Under the equivalent of button go White must pass last only if Black made the first pass, otherwise Black must pass last. OC, with the pass stones, territory counting is used.

But, as you explained in some other thread, neither of those is completely equivalent to Button Go. They're just logical implementations of territory scoring. Or am I missing something?
Author:  Bill Spight [ Fri Oct 27, 2017 7:46 am ]
Post subject:  Re: Fair komi without ties
luigi wrote:
Bill Spight wrote:
I don't think that button go is unpopular; people are just unaware of it.

The equivalent of button go has already been played in international competition. IIUC, go in the Mind Sport Olympiad uses area scoring but if White passes first White gets an extra point.

It is easy to modify AGA scoring to incorporate the equivalent of a button. Under current rules White must pass last, handing over a pass stone. Under the equivalent of button go White must pass last only if Black made the first pass, otherwise Black must pass last. OC, with the pass stones, territory counting is used.

But, as you explained in some other thread, neither of those is completely equivalent to Button Go. They're just logical implementations of territory scoring. Or am I missing something?


The first method, which uses area scoring, adds on average 0.5 pt. to White's score. It is true that the result is an integer score, but you can combine it with 6.5 komi. 6.5 + 0.5 = 7, so in effect you have a 7 pt. komi with a button. :)

The AGA modification is actually an implementation of double button go, as it is used with territory counting. It also adds on average 0.5 pt. to White's score, and can be combined with a 6.5 komi.

While you can regard button go as a form of territory scoring, it avoids the complications of territory rules, as it does not stop play when plays become negative by territory scoring, so that kos and life and death can be resolved by play. :) But you can also regard button go as a form of area scoring. Button go is a hybrid of both forms of scoring. :D
Page 2 of 5 All times are UTC - 8 hours [ DST ]
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group
http://www.phpbb.com/