12

Komi is the number of points added to White's score in Go in order to compensate for his going second. Some fifty years ago, it was 5 1/2 points; then playing experience showed that it was insufficient; so that it was moved up to 6 1/2 points, then 7 1/2 points.

During the same period, computer Go moved from being a curiosity to being able to beat the best human players. That's because today's computers play the game more precisely than humans.

Has computer Go weighed in on the debate about the proper komi (to give Black and White equal chances), and if so, in favor of the status quo (7 1/2), or a different number?

  • It is not implausible that the ideal komi depends on the strength of the players, which one could investigate by making the machine play at different strengths! Of course, anything other than perfect play is hard to define, so maybe it is not really feasible. – PJTraill Nov 5 '17 at 21:41
  • 3
    One subtle point is that territory scoring (Japanese) with 6.5 komi is nearly equivalent to area scoring (Chinese) with 7.5 komi. Also for area scoring changing komi only has a significant effect if you change it by 2. – CodesInChaos Nov 6 '17 at 9:00
5

Something to note because many people are confused by this issue... The skill of the players is irrelevant to komi. Komi is defined as the number of points given to white such that perfect play will yield a tie. (For example, komi on a 3x3 board is known to be 9 using Chinese rules because perfect play will wipe white completely from the board, and it's 9 regardless of the skill of the players.)

Traditionally, an extra half point is given to white to avoid a tie. Therefore, white should have a moderately greater chance of winning than black. (New Zealand's official rules don't do this.)

AlphaGo Teach, which assumes a 6.5 komi (Japanese) shows the percent chance of winning with black having the first move (https://alphagoteach.deepmind.com). Note, that black's best chance is 47.1%. Since white is expected to have a slightly greater chance of winning (because of the extra half point,) AlphaGo backs up the komi.

For Chinese rules komi is necessarily an odd number (because in order for there to be a tie, the sum of both player's scores must be even and there are an odd number of points available on the board.)

Whatever komi is, for Chinese rules there is widespread agreement that it's not 9 points and historically 5 points seems too low which means 7 points is the most likely value. However, without perfect play, we can't know for sure.

-- New information --

I've learned a couple of new pieces of info that relates to this question. Go on a 5x5 board was solved using a computer (not "deep learning" but a computer program none-the-less) and it turns out that komi is 25 points for that size board in Chinese rules. Also, 7x7 Go seems to have been "weekly" solved and it seems that the komi is 9 points for this. I'm not entirely sure about the info about the 7x7 board because the original paper is in Chinese and the derivative English info about it doesn't seem to be in agreement about what the paper actually says. So computers are telling us things about the proper value of komi.

2

DeepMind provided a set of 50 games played by AlphaGo against itself, where black won 12 games. It could be an indication that a 7 1/2 komi (Chinese rules) is too high.

  • Nonsense. Fifty games is far too small a sample size to make that determination. – Daniel T. Apr 30 '18 at 15:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.