[computer-go] Komi for 9x9 computers

[alain Baeckeroot]

Le Mardi 2 Mai 2006 22:56, Chris Fant a écrit :
> > > Does GnuGo make different moves during the game based on the komi?
> > >
> > Yes, but only when the victory is safe (more than 5 points if i remember)
> > So for less than 5 points victory it is very likely that changing the komi
> > by 1 point would change nothing to the game (except some pathologic case).
> 
> I don't think that reasoning is valid.  If GnuGo sees the game as
> being won by 5 points it takes action A, if it sees it as being won by
> 4 points it takes action B.  So changing the komi by one point could
> change the result.

I agree but :)
Let's be more precise, i just checked again and hope some GNU Go guru will
correct me if i m wrong.
gnugo always play the biggest move, (=> rather directly maximise score)
there are 2 moderators:
-1- if score > 20 value more connections
-2- if pessimistic_score >5 become paranoiac and consider dead dragons are 
not so dead before choosing move.

for (1) On a 9x9 board, it seems very unlikely that at one moment gnugo wins
by more than 20 => play slightly safer and later lose the game due to this.

for (2) the pessimistic score is really pessimistic (considers everything
is in favor of opponent) and this constraint is generally harder to reach
than the previous one.

-3- :) i had done some twogtp match with identical seeds and different komi
at the very begining of my investigations, and as far as i remember it
changed nearly nothing.

-4- I tried in the twin to "freeze" the game when winning, and play safe
move instead of big moves, but this does not worked as i hoped, the 
frozen-twin still makes lots of points and have huge victories against
beginners.

As a conclusion, i agree with you that in general changing komi should
change the game, but with GNU Go it seems a very good approximation
to consider that changing the komi by +/- 1 point will only change
the final result by the same amount.

Alain