[computer-go] Parallel lookahead with GNU Go

[david doshay <ddoshay@xxxxxxxxxxx>]

On Monday, December 8, 2003, at 04:16  PM, Arend Bayer wrote:

> I suppose
> what he meant by this statement is that many go programs do _not_ use
> full board lookahead. I happen to know that such is the case for GNU 
> Go :)
> And it's probably right for most of the programs that took part in 
> Graz,
> only exceptions I am aware of are Smart-Go and Aya.
>
> As you have explained, Many Faces does use full-board lookahead, and
> the other top programs probably use it, too. And I am convinced we have
> to add it to GNU Go some time, too.
Last week I reported that I am part of a group working on a parallel Go
project, and that we would report if and when we had something. We
now have a little something. Emphasis on *little.*

We have completed some infrastructure for spreading problems over a
networked cluster of workstations, in this case Apple G4s. We are
developing this infrastructure for working on a class of physics 
problems,
but are testing it in the context of Go playing. We are using GNU Go as
our move generator and have implemented parallel lookahead.

Playing white against GNU Go 3.2 with no handicap, we used GNU Go
3.2 to generate possible moves. We selected 4 lookahead paths and ran
them out 8 moves. We used a weighted linear combination of GNU Go's
estimate of the value of the original moves and GNU Go's estimate of the
overall influence in the final positions to select one move from the 4. 
We
beat GNU Go almost all of the time, usually by a substantial margin. The
small number of times we lost, the margin was small.

I could write about my opinions of what I saw, but only if folks care 
enough
to ask me. I would guess that it is only of interest to GNU Go 
developers.

We were just starting to gather statistics and tune the weighting in the
linear combination when we began an update to GNU Go 3.4.

GNU Go 3.4 does not have the influence function we had been using, so
we played our modified 3.2-with-lookahead against 3.4. 3.4 trounced 3.2
with lookahead by over a 100 point margin. We only did the one run.

We are integrating our lookahead code with GNU Go 3.4 and will report
the results.

Cheers,
David

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://computer-go.org/mailman/listinfo/computer-go