Komi in the AI age

[golem7]

I read an older thread concerning the size of komi which seems to indicate (based on data from Japan) that with a value of 6.5 white has a tiny advantage (50.8% winrate).

Modern AI is trained exclusively with a value of 7.5 as far as I know, which leads to this value becoming ingrained into the neural network weights. At the beginning of the game e.g. Leela estimates white ahead with ~53.5% winning probability.

So does this mean that 7.5 is too much after all? Or should we change tradition by way of the weaker player taking white in the future?

Is anyone aware of efforts to create an AI based on 6.5 komi? I would like to see what winning probabilities it sees at the beginning of the game. Also how that might change certain evaluations in the early opening. Best of course would be a version that works with any value for komi, also for handicap games, although that would mean significant changes to the underlying algorithm and it's questionable if/how it can be done.

[mhlepore]

I thought bots score with area scoring, which traditionally comes with the extra komi point. This sometimes, but not always, produces equivalent results to territory scoring and 6.5 komi.

(edited for clarity)

[Bill Spight]

Statistics from four decades ago indicated the the proper komi for best human play at the time was around 7, by both area and territory scoring. It was known that statistically appropriate komi depended upon the skill level of the players, with komi increasing with skill. Some people thought that komi with perfect play might increase to around 9.

However, with AI players who are better than humans finding that 7.5 komi seems to favor White, it now seems unlikely that komi with perfect play will be as high as 9. If komi with perfect play is 7, then the best fractional komi with area scoring is surely 7.5.

[golem7]

Thanks for the information! Very enlightening.

Bill, do you know when fractional komi was introduced? And why it was thought necessary to determine a winner in every game?
I myself would welcome the possibility of jigo in official games.
It would still be a rather rare occurence (not like in chess) but I think it's a nice conclusion to a game, official or not and has its own drama attached to it.

[Uberdude]

Jigo is a pain in knockout tournament and other such formats in which you have scheduled time for 1 game and want a winner for future rounds.

[Bill Spight]

See https://senseis.xmp.net/?HistoryOfKomi

[ez4u]

Statistics from four decades ago indicated the the proper komi for best human play at the time was around 7, by both area and territory scoring. It was known that statistically appropriate komi depended upon the skill level of the players, with komi increasing with skill. Some people thought that komi with perfect play might increase to around 9.

However, with AI players who are better than humans finding that 7.5 komi seems to favor White, it now seems unlikely that komi with perfect play will be as high as 9. If komi with perfect play is 7, then the best fractional komi with area scoring is surely 7.5.

Since we did not start accumulating professional experience with 6.5 and 7.5 point komi until 2002, I would be interested to know what 'statistics' were used four decades ago. We did not need AI's to tell us that 7.5 komi favored White. We have tens of thousands of games between humans that tell us that already. The leap from statistical evidence to some theory of perfect komi is unwarranted of course. Under area scoring we have the problem that we get the same result with 6.5 komi and 5.5 komi so the issue becomes does 7.5 komi grant a bigger advantage to White than 5.5 komi grants to Black. We should be testing that with statistics not theories.

[Bill Spight]

Statistics from four decades ago indicated the the proper komi for best human play at the time was around 7, by both area and territory scoring. It was known that statistically appropriate komi depended upon the skill level of the players, with komi increasing with skill. Some people thought that komi with perfect play might increase to around 9.

However, with AI players who are better than humans finding that 7.5 komi seems to favor White, it now seems unlikely that komi with perfect play will be as high as 9. If komi with perfect play is 7, then the best fractional komi with area scoring is surely 7.5.

Since we did not start accumulating professional experience with 6.5 and 7.5 point komi until 2002, I would be interested to know what 'statistics' were used four decades ago.

In 1977 someone submitted an article to the American Go Journal claiming that correct komi is 7, by Japanese rules. He had used the results of 1400 Japanese pro games with 4.5 komi and 1400 games with 5.5 komi (5 pt. komi, White winning jig). Before accepting the article Terry Benson sent me a copy of the article plus the data for review, I checked his results and, to my surprise, each set of data suggested a komi of 6.5. (I had already reached that tentative conclusion based upon much less data about the value of handicap stones. The article included a footnote by me to the effect that we really ought to test 6.5 komi before drawing any firm conclusion.) A few years later I found out that Ing had decided on a komi of 7.5 for his rules. I suppose that he made that decision based upon some statistics, but I don't know.

Pros were using 7.5 komi for Ing rules in the early 1980s, and even earlier, for all I know. In the mid 80s some people thought that a 9.5 komi might be better.

[moha]

the issue becomes does 7.5 komi grant a bigger advantage to White than 5.5 komi grants to Black. We should be testing that with statistics not theories.

Statistics are nice but finding the latter case would be surprising theoretically. If komi for perfect play is 7, then awarding ties to black makes more sense (6.5 or 5.5) since in practice we expect imperfect play (thus slightly less komi).

[ez4u]

the issue becomes does 7.5 komi grant a bigger advantage to White than 5.5 komi grants to Black. We should be testing that with statistics not theories.

Statistics are nice but finding the latter case would be surprising theoretically. If komi for perfect play is 7, then awarding ties to black makes more sense (6.5 or 5.5) since in practice we expect imperfect play (thus slightly less komi).

Sorry but I think that all statements including, "If perfect play (komi) is 7", are just hand waving. We have a lot of actual practise with different komi. I think that should be the starting point rather than the various theories of the last forty years or so. Something that I think would be very interesting would be a strong bot trained from scratch with 5.5 komi. Would it find Black has the advantage on move 1 and by how much? Would it pay the opening differently than the current 7.5 bots and how?

[SpongeBob]

Under area scoring we have the problem that we get the same result with 6.5 komi and 5.5 komi

Could you please explain that?

[mhlepore]

Under area scoring we have the problem that we get the same result with 6.5 komi and 5.5 komi

Could you please explain that?

I don't know if I follow completely, but I think the logic is along these lines:

In area scoring, final raw score difference must be odd:

181-180 -> 1 point difference
182-179 -> 3 point difference
183-178 -> 5 point difference
184-177 -> 7 point difference
etc...

With a 5.5 komi, White adds 2.75 to her score at the end.
With a 6.5 komi, White adds 3.25 to her score at the end.
With a 7.5 komi, White adds 3.75 to her score at the end.

Edit: Since White wins if she gets 180.5 or more after komi, then White will win if she has 178 points and lose if she has 177 points (whether komi is 5.5 or 6.5)

[gowan]

I'm not sure that we can answer the question of what the "correct" komi is basing it on statistical analysis of results of actual games. The reason is that in human games (excluding AI games) history shows that people play differently depending on whether there is a komi and what it is. When 7.5 komi was introduced many pros preferred to play white and it was a bit surprising when Ryu Naiwei preferred black. In effect we adjust how we play depending on the komi. I doubt we can choose whether this adjustment takes place. When we play we aim to win but I don't think we aim to win by the largest possible margin.

[SpongeBob]

In area scoring, final raw score difference must be odd:

181-180 -> 1 point difference
182-179 -> 3 point difference
183-178 -> 5 point difference
184-177 -> 7 point difference
etc...

This already shows that 5.5 or 6.5 komi cannot change the result - thanks.