For testing Suzie on 9x9 we (Peter Woitke and Chrilly) use Gnu-Go Level
16.
Is there something stronger around /available?
Yes there is cgos 9X9 ;)
I am used to another development method. I watch the engine-engine-matches
and I change the programm when I see a move which really hurts. I trust more
my feelings than statistics. E.g. for measuring an 10 Elo difference one
needs 1000 games. One can not wait after each change for 1000 games. The
games have to be played close to the usual tournament settings. Usuall
changes are in the <= 10 Elo range and for bigger improvements (or step
backs) one needs no statistic at all. One sees it after a few moves.
Its not only the "hurt effect". It is also very important to see how long a
programm needs to find a good move. The pattern: Opponent plays a move, the
eval of the own programm goes up, it thinks that the opponent has played a
blunder, but then the eval fails low is very important to watch. E.g. if the
programm searches to 7 plies and it finds the problem/correct answer at this
depth, the opponent sees 1 Ply further. Although the programm plays the
correct answer, one has to work that the programm fails low at least at ply
6.
In server matches this kind of information is usually lost. One can write
the traces into a file, but practically one never checks it in the same way
than when sitting infront of the screen and seeing it "live". The immediate
pain is not there. Avoiding this pain is in my experience the most important
factor/stimulus for improving a programm.
To generate this pain one needs a slightly stronger opponent. The pain-level
of Gnu-Go is for Suzie on 9x9 already too low.
Chrilly
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
