Re: [gnugo-devel] level of play

[david doshay]

On Saturday, January 10, 2004, at 12:15  AM, Evan Berggren Daniel wrote:

> On Fri, 9 Jan 2004, david doshay wrote:
> we win 4 out of 8 doing a 16 move lookahead against a 4 stone 
handicap.

> Any chance we could see the game records?  Just the games themselves
> would be interesting, though seeing what moves were looked at would 
> too.

We save all game records and look-ahead branch points (but not the
lookahead paths) along with the evaluation of the board value at the end
of each lookahead. We will be happy to share all records and our code,
but would prefer to make a few more improvements first. If you really 
want
to see it sooner we can do that too. We also will be gathering better 
statistics.

> Do you know if these results generalize at all?  ie, is your machine 
> that
> much stronger when played against human or other computer opponents?  
> My
> own metamachine experiments suggest that the effect is far more 
> pronounced
> against gnugo than against other opponents,

We have not run against any other program or human. We will run against
Many Faces of Go 11.0 in the near future.

Does anyone have statistics on how 3.4 has done against Many Faces?

For some reason I am presuming that the effect is more pronounced 
against
itself than other programs. But it is worth mentioning that in earlier 
tests using
this method in a machine learning context with professional games for 
input,
we were able to correctly select more of the professional moves than 
normal
3.2 (what we were running at that time). We have since dropped the 
machine
learning efforts because we did not see much in the way of convergence 
from
any automated methods we tried, but we did hand tune some parameters as
a result of what we observed.

Thanks,
David