Yes, we even call it dynamic komi as you just proposed. Many programs
implement it and there have been many discussion about it. It seems to
be very effective in high handicap games and in even games it may have
a small effect for some programs.
There are different implementations: the simplest starts with lets say
7 point extra komi for each handicap stone and this value is then
linearly decreased until 200 moves have been played or so.
More complex versions of dynamic komi adjust the komi during search so
that the winrate stay to close to 50% (higher or lower depending on
how aggressive one wants the program to play). This method is more
demanding for programmer because it is not so easy to handle cases
where the score fluctuates large in search or on the actual board.
Magnus
author of Valkyria
Quoting ds <[email protected]>:
Yes thanks for the answer. It is difficult to search the list for
detailed information.
Do you know, if anybody tried an approach as dynamic Komi with this.
In the beginning of the game take into account the stones (eg.
sqrt(sones) and reduce it later. In the beginning an estimated high
win should be probably more save than a 0.5 win?!
Thanks Detlef
Am Freitag, den 29.04.2011, 12:21 +0200 schrieb Oliver Lewis:
This has been the subject of many discussions and debates on this
list. Essentially, the evidence shows that although win/loss "loses"
information, it is the most correct evaluation that captures what the
go player should try to achieve. If you take into account the size of
the win, you are somehow saying that bigger wins are better than small
wins. However you weight the scores, you risk the programme favouring
some of the aggressive lines of attack which may also be riskier,
rather than playing safe for a half-point win.
On Fri, Apr 29, 2011 at 11:15 AM, ds <[email protected]> wrote:
If I read correctly, the playout results in Monte Carlo
usually used are
only win or loss.
Is there a good reason, why one should not take into account
"height
(number of stones)" of the win?
It feels like loosing information to me?!
Thanks a lot
Detlef
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
