On Wed, Jun 23, 2010 at 05:32:01PM +0900, Darren Cook wrote:
> I had a real shock today, on a par with the shock of England losing 0-0
> to Algeria the other day.
>
> I'd always though playouts, even heavy ones, have a random element. But
> in fact they are a sequence of deterministic if/else statements, at
> least in Fuego [1] and Mogo [2].
>
> I'd always thought patterns and atari moves and the like were just
> weights. So if there was a chain in atari then e.g. saving it might get
> played 60% of the time, but 25% of the time a pattern move would be
> played, and 15% of the time a pure random move. But actually it is 100%
> of the time the save move will be played.
I have not done some testing recently but from my past experience with
Pachi, 90% is roughly the optimum, but 100% is not much worse!
In general, the shorter the thinking time and the larger the board,
the more weight heuristics should have since the initial speed of
convergence is much higher even though perfect convergence is much
harder; this seems to work similarly for noise in playouts and e.g.
weight of priors in the tree. Perhaps best option would be to tune
these values based on time and branching factor, but rigorous
analysis of the relationship seems difficult and experimentation
with long thinking times very time-consuming.
--
Petr "Pasky" Baudis
The true meaning of life is to plant a tree under whose shade
you will never sit.
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
