[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [computer-go] Statistical Significance (was: SlugGo v.s.ManyFaces, newest data)

On Sep 7, 2004, at 10:56 AM, Mark Boon wrote:
I would agree with you if the posted results concerned versions A and B of
the same program.
When we play against GNU Go at 5 stones we win 70% of the time,
with the GNU Go engine playing at level 15 and SlugGo using the
GNU Go engine at level 10.


Since the most recent results are versus ManyFaces, an
alltogether different program, I think the outcome should have more
credibility. Unless the program was specifically tuned to play against
ManyFaces beforehand, something I understood wasn't the case.

The program is not really tuned at all. There are probably 20 or so
numerical values in SlugGo that are not in GNU Go that could be
"tuned" to see if it matters. We made a wild guess and then played.
Note of the parameters has ever been adjusted. Not for play against
GNU Go or Many Faces.
Given the results posted I think it clearly shows a 4 stones difference over
ManyFaces. Maybe further testing will adjust it down to three stones, but my
gut-feeling says that's going to be very unlikely. And even when it does,
the results are very impressive.
Thanks, we are pleased. But rather than tune the existing parameters
or gather more statistics I am going to make improvements in the logic.
As noted by some others, SlugGo makes some obvious endgame blunders,
due to some algorithms being implemented quite differently than I had
intended. I feel sure that the problems that are so obvious in the endgame
are also a problem, although more subtile, in the middle game.

Once these logic problems are fixed and we see how that changes
things, then I intend to work on intelligent branching, rather than just
doing the once-branched then linear lookahead. I think that the power
of the cluster will come to light then. Right now we use 10 cpus to play
these games. I'd like to use all 72 for one game rather than than for
playing multiple games.

Cheers,
David


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