Re: [gnugo-devel] Three stone games

[Arend Bayer]

Hi Morten,

you wrote:
> In 9 x 9 Go I believe we could get something like  :                          
>                 
>       Black wins      70.00%                          
>       White wins      25.00%                          
>       Tie             5.00%                           
>                       100.00% 
> without komi offset.                  
> And this I'd like to investigate. When we know the numbers for 9 x 9,
> then we 
> can investigate if this is true for 19 x 19. 
There some statistics about winning probabilities in even games between
players of different ranks at the web site of the EGF ratings:

http://egf.posluh.hr/rating/gor.html

This might not be exactly what you are looking for, but it is s.th. similar
and interesting.

> I'd like to do this, if it is helpful for you. gnugo versus gnugo, and
> gnugo versus Morten Gulbrandsen. I only play as 13 kyu, so it could be
> an even match.                                                
We would certainly be interested to hear how you compete against GNU Go,
on different board sizes. GNU Go vs GNU Go might also be interesting.

> I also find it very difficult to understand, why gnugo does not increase
> its strength
> with computing speed?  If gnugo plays against itself with different time
> limits, 
> then how many minutes will it need to play 1 kyu stronger ?
The strength does increase with more computing speed, but the increase
will get slower and slower. Strength of chess programs (measured in ELO)
increases linearly with each duplication of computing power. This is,
as far as I know, not true for any Go program. 

Speaking about GNU Go, there are simply mistakes that GNU Go, in its
current design, would make with any time possibly used. E.g. it makes
connection mistakes, because there a dynamic connection analysis is not
yet fully implemented.

In fact, I would find it quite interesting to know how much GNU Go's
strength depends on the time spent. Especially, at the moment, I would
like to know whether it would be worth adding better support for higher
levels than 10.
With the program "twogmp" (included in the GNU Go download, in the directory
interface/gtp_examples) you can do this very comfortably. E.g. you could
try to run 

twogmp --black '../gnugo --quiet --mode gtp' \
        --white '../gnugo --level 11 --quiet --mode gtp' \
        --handicap 1 --games 20
(This will run 20 games without komi, black GNU Go at level 10, and white
GNU Go at level 10.)
(Of course, you can also similarly run one game at a time from cgoban.)

It would be interesting to know the results such a series, and the time
needed by black and white.

I will report the result of a similar match later, where I let one
version play with --owl-threats (which does more reading in some
situations, needing more time).

Arend