Re: computer-go: komi

[David Fotland <fotland@xxxxxxxxxxxxxxx>]

Do you mean komi of 3 points?  There should be more variation in weaker 
player's games.

I have a book of pro-pro 9x9 games, played with 5.5 point komi, so the pros 
think
5.5 is correct (although they just switched to 6.5 on 19x19).

Many Faces is weak at 9x9, because it hasn't been tuned for it.  I do not 
test at 9x9.

Here are statistics from the book:  In 9x9 games with 5.5 komi, the results 
were:

B wins by:

0.5     5
1.5     2
2.5     4
3.5     7
5.5     2
6.5     2
9.5     2
Resignation     41
65 total


W wins by:

0.5     5
1.5     5
2.5     7
3.5     6
4.5     4
5.5     1
6.5     2
7.5     1
8.5     1
11.5    1
12.5    1
resignation:    51
85 total

This would seem to indicate that the correct komi is 3.5, but most of these 
games were huge fights from
start to finish, so it's unclear how a different komi would have affected 
the fights.

It's interesting that your match and the pro games find the same 3 point komi.

David Fotland

At 09:41 PM 5/26/2003 -0400, Don Dailey wrote:

> A few questions for the group or anyone that knows:
>
>    1. What do you believe is fair komi for 9x9 Go?
>
>    2. Would 2 very strong players need a different komi than 2 very weak 
> players in
>       order to equalize?
>
>    3. Does this relate to computers in any special way?
>
>    4. How strong is gnugo 3.0.0  compared to the latest versions?
>
>    5. Are some programs particularly weak/strong in 9x9 Go (in comparison 
> to other
>       programs that is?)
>
>    6. Does komi vary significantly with board size?
>
>
> These questions are  very interesting to me because  I'm working on my
> program again and focused primarly on 9x9 board sizes.  I have come up
> with a very interesting playing  algorithm that seems to scale up very
> well with compute power (and is not based on brute force searching.)
>
> In a 900 game match, a komi  of 3 stones was just about exactly enough
> to give white  equality on level 4 (a move  is made almost "instantly"
> in human perception on levels below 7 for my program.)
>
> I will  do the komi  test against much  stronger levels and  report my
> results here.  I am curious if others have done these tests with their
> programs and what the results show.
>
>
> Don