Re: [computer-go] Parallel lookahead with GNU Go

[david doshay <ddoshay@xxxxxxxxxxx>]

On Tuesday, December 16, 2003, at 02:28  PM, Don Dailey wrote:

> That  sounds impressive,  but you  did not  report anything  about the
> number of  games played.  It's easy  to be extremely  mislead by small
> statistical samples.
I am a physics guy, and I know exactly what you mean.

> Can you give more stats?
I wish I could. As I said, we were about to start collecting real 
statistics
when we instead switched to 3.4. Any attempt at real statistics over the
games we played was further compounded by variations in the methods
for selecting the 4 lookahead paths and by different weighting between
the original value of the move and the value assigned to the influence
at the end of the lookahead.

As I also said, we had a "little" something to report. I really do not 
think it
is particularly impressive yet.

But roughly, we played about a dozen games and lost 2. We won by
between 35 and 140 points, and lost by 10 or less.

Integration with 3.4 is nearly done. We will fiddle a bit and then start
collecting real statistics. We will save the games, both in standard
format and with more info about the lookahead paths, for analysis.

For now it is still easier for us to play our GNU Go with lookahead
against GNU Go as it normally is. I do think that this causes something
of a "mirror" effect, in that when we choose a path with great estimated
strength, it must similarly be evaluated exactly the same way by the
same move generator, which will not necessarily attack something it
evaluates as that strong. When we are further along we can get to
playing against other programs to judge the real overall effect.

David

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