Re: [computer-go] Time constraints for computer Go

[David G Doshay <ddoshay@xxxxxxx>]

On 15, Feb 2005, at 3:31 AM, Mark Boon wrote:

> At the moment SlugGo is the only software that scales well.
I would not put it quite that way.

I would say that SlugGo is an experiment in a particular
form of scaling, and it demonstrates the way that GNU Go
evaluations scale under that scenario.

It is true that SlugGo uses hardware in a way that scales
such that increasing lookahead depth results in linear
increase in wall clock time regardless of width of search
(as long as we don't run out of CPUs) so SlugGo meets
that definition of "scales well."

There also is the general trend that short lookahead depth
has limiting behavior of playing like GNU Go and longer
lookahead depth plays considerably stronger, so that fits
the general description of "scales well."

But I believe that even longer lookahead depths would
NOT continue to result in stronger play. I think that there
is a range of scalability of SlugGo's method and that
range is limited. I think it has a peak, and after that further
lookahead depth will clearly take longer, but will result in
weaker play ... an example of "does not scale well."

After the round of bug fixes we are now in, we will be
running experiments to determine the strength v.s.
lookahead depth curve. Some preliminary results were
odd, so we are going to have to gather a very large
number of games to get good statistics.

Cheers,
David


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