Hi,

I'm a bloody newbie and make my first Go steps with Drago + Gnugo on a 9x9 board. What Komi value should be set on a 9x9?

Thanks in advance
Michael

Opinions vary, but 6.5 is probably reasonable, the same as is standard on 19x19 boards. In handicap games, just 0.5 is used, which is the same as 'white wins draws'.

If you're good enough to play drago on even (or even if you aren't), you might be more interested in playing against other people. It's also likely a better way to learn against a range of opponents, though a computer is okay for getting the basics. KGS is the most popular English language go server, and has a great community. Of course you don't have to, but don't fall into the trap of thinking you're too new to do so...you aren't!

Thank you. I am too new because maybe I do not understand all rules! Also, I do not always see the ending of a match... And my playing is not good

For instance, what I don't understand. At the end of the match, if white has a territory on the boards side, with, say 10 points. Why doesn't black put his piece into it, so only 9 points are left for white?

If the black stone cannot live, white doesn't have to play to capture it.

_________________
Occupy Babel!

OK thank you, we discuss the invading issue in this thread, it's more ontopic: viewtopic.php?f=11&t=4463

there may be a setting that you can set to tell it not to resign. That way it'll play it out to the end and it'll be easier to understand how a game ends.

To attempt to answer the question of “What komi is appropriate for the 9x9 board?” I decided to run some simulations and attempt to find out. Included with the source-code for gnugo, was a python script called twogtp.py, that allows you to pit two instances of gnugo (level 10) against each other. I used this script to run lots of games (Japanese scoring) with three different values of komi (5.5,6,5,7,5), and then looked at the white win rates. Using a little statistics (approximating the distribution as normal) I was able to estimate the uncertainty in the white win probability based on the number of games played and a 95% confidence level. Here’s what I found.

9x9 Results (Japanese rules)
komi W-wins B-wins W-win-probability
5.5 427 573 0.427 ± 0.03
6.5 518 482 0.518 ± 0.03
7.5 564 436 0.564 ± 0.03

13x13 Results (Japanese rules)
komi W-wins B-wins W-win-probability
5.5 4685 5315 0.469 ± 0.01
6.5 5063 4937 0.506 ± 0.01
7.5 5223 4777 0.522 ± 0.01

I had to increase the number of games for the 13x13 board to 10,000 to reduce the uncertainty in white’s win probability estimate to about 1%. The white-win probabilities for the 13x13 board were much closer together, which means the komi is less important on the bigger board (bigger point spread).

In conclusion, it looked like 6.5 was the appropriate komi for white on both the 9x9 and 13x13 boards (according to gnugo). The 6.5 komi was the only one that gave a 95% confidence white-win-probability range that encompassed the desired 50%. I’m tempted to try running more simulations with other go engines (pachi) to see if I get similar results,

It took my poor old laptop about 5 days to play each batch of 10,000 13x13 games.

thanks, was useful for me as well

I recommend taking up the goquest amount of 7 KOMI, seems pretty balanced.... but slightly better for white in my case, but if i checked other id seems to have an opposite result.....

Not actually running simulations, just trusting KataGo's 40b neural net's distillation of its self-play data (which would be less accurate than running games, of course, but still using the net's experience from very strong self-play games):

Winrates for white on 9x9 with different komi, Japanese-like rules, searching from empty board with a few hundred K playouts:

16% 4.5
26% 5.5
67% 6.5
81% 7.5

Winrates for white on 13x13 with different komi, Japanese-like rules, searching from empty board with a few hundred K playouts:

29% 4.5
42% 5.5
54% 6.5
65% 7.5