James Pettit wrote:
> 9x9. 10k gives about a 5% winrate as black, 50k maybe 17%, and 100k a
> solid 20%,

Ok, that doesn't seem right.

Testing with a program that should be close to yours (uniform random
playouts except for simple ko and filling its own 'eyes', UCT + RAVE, no
exploration term, expanding each node on the second visit), taking black
with 6.5 komi against gnugo 3.8 level 6, I get a very neat 30% with 10k
playouts, 40% with 50k, 50% with 100k.

(Unlike what others are reporting, I do see a big improvement from RAVE
even with uniform random playouts.)


> but it's not so much the winrate as the fact that the lost games are
> almost all W+87.5.

I'm not sure that's anything to worry about. Monte Carlo programs don't
care whether they lose by a little or a lot, and on a 9x9 board it's
easy for everything to die.

My program resigns, so I can't give a direct comparison, but it does
look like lots of its losses would end up like this.

-M-
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go

Reply via email to