Remi,
It's not so much a question, just a thought about the use of power.
Not every MC program is the same. And that's assuming you leave all the
UCT/RAVE/etc. variations completely out of the comparisons. In fact, in
my reading this list over the last 24 months, I would say it appears
that the similarities between MC implementations are coarse grained.
When you get down to the implementations, each MC program seems to be
varying in the details. Some are light, others are heavy in their
playouts. Some are including hand coded evaluators some of the time,
others all the time thereby playing out fewer games per second. Some are
playing deeper (auto stop on 9x9 at 200 moves, 300 moves, when b/w score
ration exceeds a ratio threshold, etc.), some are playing broader
(degree of distribution deepening particular nodes). If you were to try
and map all of this into some genomic analogy, there would be lots of
similarity in morphology (directly observable appearances), but huge
differences at the genes. And the optimal effectiveness/performance
ratio devil is in the details.
So, when two programs play each other and they both assert they are "MC"
programs, I am not sure much is really be said in terms of the
comparison, other than MC has become Go's equivalent of the marketing
term "multi-media" in the PC world. It now includes so much, it's hard
to call it a distinction any more.
That said, it means that one "MC" program might be using it's power much
more effectively (require substantially less CPU cycles to produce a
similarly skilled result) than another, even though both assert they are
"MC".
Perhaps it would be better said that a Go program has an MC
characteristic rather than saying it is an MC program; association
versus identity.
Jim
Rémi Coulom wrote:
Jim O'Flaherty wrote:
Remi,
I find it interesting that he won with the slowest hardware. I am
still wondering how much performance is still a low influencer. In
other words, a really fast poor algorithm won't be a better algorithm
on slower hardware (or slower software).
Jim
Hardware has a lot of influence on playing strength. I noticed an
improvement of about 100 Elo points for a doubling of computing power
(19x19). This is against other MC programs. Against humans, the
improvement is less.
I am not sure if this was your question. A poor algorithm won't be a
better algorithm on slower hardware, but I am not sure what it means.
Rémi
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/