[Pachi] [Computer-go] MCTS and perfect endgame

Jean-loup Gailly jloup at gailly.net
Sun Jul 3 16:49:06 CEST 2011


Leon,

Thanks for your suggestions. Pachi already supports dynamic komi, but we
don't
keep histograms of win ratio per komi value as you suggest. It is probably
feasible
but I'm not sure it's worth it. It should be possible to tune the current
dynkomi code
to get a similar effect but without keeping histograms.

Jean-loup



>
> Greetings,
>
> Solution is so simple it can be done by everybody.
> I am not that kind of programmer and have no idea
> how to apply patch. Anyway this idea apply to all MCTS.
>
> It is about playing perfect endgame not winning.
>
> Here is short version. Idea is in RANDOM KOMI.
>
>
> ----------------------------------------------------------------------
> MCTS and perfect endgame (short version)
>
> Endgame:
>     groups are stabilized, playout returns only result between A and B,
>     instead of Tromp-Taylor count use RANDOM KOMI (k) between a and b,
>     like a<A<B<b and a<k<b
>     winrate gives values in proportion
>     at the end of the game use real komi
>
> Example:
>     endgame results between -1 and 3
>     random komi between -4.5 and 5.5
>     0.1 in winrate diff is point equal pont diff
>     winrate gives values between 0.3 and 0.7 (depends color)
>     when you look at possible moves you get their value
>
> Why:
>     basic mathematics
>
> Plain terms:
>     optimal play has greater percentage wins
>     you get rid of noise (random play inside secure territory)
>     and you can choose between best moves since they are clearly
> differentiated
>     of course bad playout and weak machine woud make too much noise
>
> Leon.
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://rover.ms.mff.cuni.cz/pipermail/pachi/attachments/20110703/e7d7e7ea/attachment.html>


More information about the Pachi mailing list