[computer-go] cgos komi (was: Odd experimental results)

[David Doshay]

These numbers make me think that setting komi to
6.5 for the next N00,000 games should be tried.

Cheers,
David



On 28, Aug 2006, at 8:48 PM, Don Dailey wrote:

> I did a little analysis from all the CGOS games.
>
> I considered all the games between pairs of players of roughly equal
> strength and made a table.
>
> komi on CGOS is 7.5
>
>    RANGE        GAMES        BLACK PERC        WHITE PERC
>  600 - 799      17546          41.47 %          58.53 %
>  800 - 999      10883          39.36 %          60.64 %
> 1000 - 1199      1515          43.37 %          56.63 %
> 1200 - 1399      5542          43.79 %          56.21 %
> 1400 - 1599      9774          48.23 %          51.77 %
> 1600 - 1799     14706          49.37 %          50.63 %
>
> In my table, I only consider games where BOTH players are in the same
> 200 point range and both players are well established.
>
>
> A similar table considering only opponents within 100 ELO of each  
> other:
>
>
>    RANGE        GAMES        BLACK PERC        WHITE PERC
> 1000 - 1099       803          41.84 %          58.16 %
> 1100 - 1199        55          43.64 %          56.36 %
> 1200 - 1299      2792          43.80 %          56.20 %
> 1300 - 1399       338          47.63 %          52.37 %
> 1400 - 1499       289          46.02 %          53.98 %
> 1500 - 1599      5984          48.85 %          51.15 %
> 1600 - 1699      1962          47.50 %          52.50 %
> 1700 - 1799      6044          50.15 %          49.85 %
> 1800 - 1899       863          46.93 %          53.07 %
>
> It appears that with 7.5 komi, white has an advantage for most of  
> these
> brackets of players except the 1700-1799 players.
>
> If one of the players who frequently plays on CGOS has a color bias or
> bug, that could upset these statistics.
>
> The table implies a trend for stronger players to require higher  
> komi's
> and also that 7.5 komi is pretty close.  This proves nothing
> conclusively of course.  These are computers and not humans.  The  
> really
> strong players, between 1800 and 1899 appear to need LESS komi, but
> there are only 863 games to sample from there.