________________________________
From: Oliver Lewis <[EMAIL PROTECTED]>
On 11/18/08, Michael Williams <[EMAIL PROTECTED]> wrote:
It doesn't make any sense to me from a theoretical perspective. Do you have
empirical evidence?
> I agree that empirical evidence is required, but theoretically, if MC
> converges to something that is not perfect play, then as the number of
> playouts goes up, the probability of playing a different move (the perfect
> move) goes down, so the programme could get weaker.
Not converging to the best move does not imply that stopping sooner would have
found a better move. It might be looking in the wrong place entirely - like a
drunk who searches for his keys under the lamp post because the light is better
there, instead of near the car where he dropped them.
An argument might be made that more playouts fail to improve the quality, and
the time could be better spent elsewhere. I've noticed a trend to separate the
discussion of "quality of move selection in the tree" versus "quality of move
selection in playouts", for instance. One might argue for fast, light playouts
for evaluation purposes, coupled with more intensive top-level search for
evaluation candidates.
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
