-----Original Message-----
From: [EMAIL PROTECTED]
To: computer-go@computer-go.org
Sent: Wed, 28 Feb 2007 11:09 PM
Subject: [computer-go] Re: Big board
This seems to have gotten stuck in various email delays, so I am resending.
Sorry in advance if you get 2, but I did not see it get through.
Cheers,
David
Begin forwarded message:
> From: David Doshay <[EMAIL PROTECTED]>
> Date: 27, February 2007 6:29:01 PM PST
> To: computer-go <computer-go@computer-go.org>
> Subject: Re: [computer-go] Big board
>
>
> So, I would say that they are all still on the hot side, but the > one in the
> middle is closest to being "eerily reminiscent" of an > Ising system near the
> critical temperature.
>
> Are you willing to discuss what is involved in your heavy playout? > I am
> racking my brain trying to think about energy functions and > other things
> that are close to physics, but I have had no good > ideas yet.
>
> Cheers,
> David
----------------------
Chris Fant opened a door by demonstrating how easy it is to generate a
decent sized image from a go game using fast playout games. (Easy conceptually,
memory management is another matter!) It seems kind of obvious in hindsight but
I sure hadn't seriously considered it. I've wanted to apply image processing
techniques to go from the first time I saw a go board. But there's not a lot to
be done with a 19x19 grid.
Designing the rules for a heavy playout function is somewhat analogous to
designing an iterated function system (IFS) to draw a fractal image. (A fractal
fern is the most recognizable example of an IFS. A few minutes at wikipedia is
enough to see how to draw the image from the equations. Generating the
equations in the first place is more involved but Barnsley's Collage Theorem
allows one to turn it into a search problem.) To get the behavior I want, I
need to adjust the rules synergistically. I can do it by hand, but ultimately,
I intend to tune the rules automatically.
But I'm not drawing a picture. The rules in AntIgo's playouts, which is
what I used to generate the images, are strongly influenced by Mogo, with some
machine learned patterns and a few extra go-heuristic rules added in.
And I agree with you that the middle image is getting close to a proper
fractal but not yet there.
Here's the part most likely to interest you. I have a clustering rule that
discourages moves where the stone would be outnumbered by enemies in a local
region and encourages moves where the balance of color in a local region is
near balance. Changing parameters in this rule strongly influences how clumpy
the resulting board becomes. But to make it look right, I really have to adjust
the other rules too, even the pattern sensitivity.
Interestingly, when I added that last rule, the playout games became
stronger. By this I mean, if I run a bunch of playout games where black and
white use the heavy rules but only black includes the clustering rule, black
wins more often then it would otherwise. The margin of improvement is only
about 7% for 9x9 boards but increases with board size and becomes quite high
for very large boards. Others on this list have mentioned that this kind of
improvement does not automatically translate into real improvement when
incorporated into an MC engine. I've had pretty good luck in this regard so far
though, but time will tell.
Dave Hillis
________________________________________________________________________
Check Out the new free AIM(R) Mail -- 2 GB of storage and industry-leading spam
and email virus protection.
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
