David,
A while back, Don did a 9x9 scalability study (before this one
http://cgos.boardspace.net/study/index.html) comparing light versus heavy
playouts. The light playouts didn't?scale badly, they didn't plateau early,
they just?weren't as strong?as the heavy ones. There was nothing to suggest
that (enough) light playouts, in a MCTS, couldn't produce a strong engine. This
is consistent with what you report.
("Scalability" is another good keyword for those interesting in keywords to
guide search in this list.)
It sounds like you are passing up low hanging fruit here. I predict that a
version of Many Faces incorporating, for instance, the heavy playout policy
described in the (first?) Mogo paper would be significantly stronger than the
current one. If it wasn't, I would be quite surprised and curious to learn why
not.
Naturally, I'm handicapped by not knowing the details of the current Many Faces
playout policy. I bet I'm right, though.
- Dave Hillis
"Often in error, never in doubt."
-----Original Message-----
From: David Fotland <[EMAIL PROTECTED]>
To: 'computer-go' <[email protected]>
Sent: Tue, 14 Oct 2008 11:17 pm
Subject: RE: [computer-go] Go/Games with modified scores
I use Many Faces' knowledge in the search, but not the playouts. I made
them as light and fast as I could.
I don't have a way any more to just do playouts without the uct search. In
your position after a few thousand playouts, it gets over 95% win for white
and 5% win for black (depending on who I make move first).
If I limit the uct search to 1 ply, it thinks white wins 49.5% and black 49%
(of the best move for each, always D9)
I guess this confirms your assertion (and Don's as well) that light playouts
without UCT search can't play well. With UCT search, it seems that light
playouts are quite effective.
David
> -----Original Message-----
> From: [EMAIL PROTECTED] [mailto:computer-go-
> [EMAIL PROTECTED] On Behalf Of Darren Cook
> Sent: Tuesday, October 14, 2008 7:42 PM
> To: computer-go
> Subject: Re: [computer-go] Go/Games with modified scores
>
> > Many Faces uses quite light playouts, and is 1 kyu 19x19 on KGS when
> run on
> > 32 cores. So I think you can make a fairly strong program using
> light
> > playouts. My playouts are certainly far lighter than Crazystone or
> Mogo.
>
> Hi David,
> "quite light" is a bit vague, and I got the impression you were using
> the classical many faces knowledge, at least the tactical search
> information, in your playouts?
>
> What do your program's playouts think when presented with the board
> position in the article? This is a terminal position, both players have
> passed, a comfortable white win, yet pure random playouts think black
> will win more often.
>
> Darren
>
> >>
> http://dcook.org/compgo/article_the_problem_with_random_playouts.html
>
>
> --
> Darren Cook, Software Researcher/Developer
> http://dcook.org/mlsn/ (English-Japanese-German-Chinese-Arabic
> open source dictionary/semantic network)
> http://dcook.org/work/ (About me and my work)
> http://dcook.org/blogs.html (My blogs and articles)
> _______________________________________________
> computer-go mailing list
> [EMAIL PROTECTED]
org
> http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
