May I suggest a fundamental limit to the utility of this scalability study?
We are comparing three programs to each other, IIRC - Fatman, Mogo, and Gnugo.
All three are known to have certain odd little quirks. The two MC programs, in
particular, are
known to be deficient when addressing certain L&D situations. They may continue
to get
better with respect to each other for some time yet. They will reach a sort of
local optimum
with respect to each other.
At the end of the day, a halfway clever 9 kyu player will play the vital point
on a bulky five,
or induce the programs to create a square four, or some other common strategem,
and the
standard of play may be shown to have some serious gaps. The programs will be
"perfect"
or "near genius level" only against other programs. This is a variation of the
problems faced by Sluggo --
if I recall correctly, it could give Gnugo a 6 or 7 stone handicap, which was a
very impressive achievement - but it fell flat against the average human player.
I think this study is useful - that's why I'm lending four cores - but I also
think we should create
some test cases and see whether more playouts will cause the programs to
resolve these test cases
properly, or not. When a group is provably dead, then it's contribution to the
final estimation should be that of a 100% dead group, not 50% or 80%. I know
this is difficult in the general case ( there's a paper demonstrating that it
is NP-hard, if I recall correctly ), but let's try to test the 20-kyu and
10-kyu cases,
at least.
Otherwise, I suspect that it may be a very long time before these fairly common
corner cases are handled correctly.
Terry McIntyre <[EMAIL PROTECTED]>
____________________________________________________________________________________
Never miss a thing. Make Yahoo your home page.
http://www.yahoo.com/r/hs
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
