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

Jean-loup Gailly jloup at gailly.net
Sun Jul 3 18:45:16 CEST 2011


> Suppose you have 5 valid point which lead to same result r.
> No matter how you play your winning percentage is  (r-a)/ (b-a)
> where a lower bound of random komi, b upper bound of random komi, r is
fixed result
> And (b-r)/(b-a) for opponent of course.

This is true only if a <= r <= b. You select moves with higher r, which is
also
what the parameter val_scale does. The cutoff b is similar to the
current parameter val_points. The question is which heuristic is better,
and this can only be determined by experiments.

Jean-loup

2011/7/3 Leon Matoh <leon.matoh at guest.arnes.si>

>
> Only endgame, because else there is too much noise. Exact constants are
> variable, but easily determined by  UTC (sure win/loss + somewthing for
> noise)
>
> Also playout engine must be sufficient strenght. Pachi has problem with
> life/death seki, nakade so if there are such situations then you have to
> adjust
> window.
>
> As of why it is obvius in next mail.
>
>
> Leon.
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://rover.ms.mff.cuni.cz/pipermail/pachi/attachments/20110703/e4714bfe/attachment.html>


More information about the Pachi mailing list