On Mon, Dec 14, 2009 at 11:45:44AM -1000, Mark Boon wrote:
> > I've found that AMAF gives very little boost with light playouts,
> > you really need something fairly meaningful already to get any kind
> > of real boost. To have at least 10% chance of beating GNUGo with
> > reasonable time per game (to be able to play-test your bot), I think
> > you can't avoid doing a lot more than plain UCT + AMAF + light playouts.
>
> I suppose it depends on what exactly you try to boost.
>
> As others already stated doing just light playouts, no tree-search,
> AMAF gives a huge boost.
I never tried that, I was talking about UCT case only, sorry if that was
unclear.
> With light playouts and UCT tree-search, AMAF
> still gives a considerable improvement. Maybe you can try to search
> the archives, I think I posted about it. My MCTS ref-bot has a
> parameter called 'useAMAF'. You can experiment with witching it on and
> off with different numbers of playouts if you want to. If I remember
> well, the difference was very obvious, even with high number of
> playouts.
Interesting, in my bot, AMAF really made only very small difference
(maybe 15% -> 25%, +-5% [*]). Perhaps I had some bug in my implementation
that caused these measurements.
[*] My test records do not go back enough and I cannot re-check easily
since apparently, in current Pachi non-AMAF UCT policy must be broken
since it never seems to win against GNUGo, and I really don't feel
motivated enough to debug plain UCB1. ;-)
So I'm sorry if I misled anyone, perhaps it is/was some Pachi bug
causing these observations of mine.
--
Petr "Pasky" Baudis
A lot of people have my books on their bookshelves.
That's the problem, they need to read them. -- Don Knuth
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
