> I think that is UCT and it's happening now. UCT is the most promising
> for 19x19 progress that we have now.
yes, it's new, and it's doing quite well. my hunch is simply that a few
thousand more ELO are not going to happen in hardware with this method.
> I think this statement is more or less true. Didn't you see the
> scalability data for 19x19? In fact didn't you help me produce it?
we tested some very low ELO ranges. speculating about how that
scales up to the "upper stratosphere" of ELO is pretty difficult for
me. it wasn't straight enough for me to believe that it doesn't go
"log" at some point nearby and start to cripple the doubling of cpu
advantage.
> This is a hard paragraph to understand. I wish you had included the
> statement that you are referencing here and who said it.
oh, this is just in response to the quote from the paper. my point
was simply that doing very well is easy for a new method, and doing
as well as the state-of-the art, or even doing slightly better, is interesting,
new, and exciting, but that we shouldn't extrapolate to imagine
that it will solve future problems.
> In what sense could this be the wheel re-invented?
in the sense that 19x19 is still brutally difficult, and that these
methods haven't improved the state of the art by more than a stone
or two, if that. so we should definitely not extrapolate, or expect
them to perform, any better than we already have evidence for.
> It appears to be the case that 19x19 UCT MC programs are "better than"
> the traditional program now. At least from the email from Rémi.
i'll let david fotland field this. a stone or two is a good thing, but i'm
not yet convinced that it scales, in a real-world way, to the additional
15 stones or so required to beat top humans.
> I'm not sure what you imply by lack of evidence, but the evidence is
> getting pretty strong in favor of UCT type programs being superior if
> that's what you mean. It isn't even a question at smaller boards.
i agree that on smaller boards UCT-type programs are superior.
without trying too hard to sound like an apologist/traditionalist, i will
mention that
boardsize isn't merely a "scaling factor" in this problem. things change
in a fundamental way inbetween 9x9 and 19x19 that direct search can't
recognize. (this is essentially what monte carlo methods are doing, as
they are somewhat carefully sampling from the move distribution).
> I'm sure some will believe this observed scalability is short lived but
> I know of no reason to believe that other than superstition.
i hate to do this, but i'll give you an analogy that i think is relevant.
if you crawl at 1/2 mph across the desert for 7 years, encounter a
tiny hill, and manage to scale it, you may say to yourself that you've
made a massive accomplishment. and you have. but it doesn't
imply, entail, or otherwise suggest that all future obstacles will be
of similar size.
honestly, 9x9 doesn't even leave *room* for some of the important
problems that are critical on a 19x19 board. those problems don't exist on
a small board because it's a full-on tactical fight from the get-go. this
is a different kind of problem than being willing to trade 40% of the board
for a 51% likelihood of getting 41% of the board in exchange. 9x9 is
about getting a 100% likelihood of winning as soon as possible.
> There is no indication whatsoever than computers are even close to topping out
> in computer chess.
don't get me wrong -- computers are remarkably good at chess. much
better than people. all at once, the world's best player lost a few games
to a computer, and then the whole thing skyrocketed. it was pretty amazing
to watch.
right now we're talking about computers possibly playing go at the level of
someone who thinks about the game but who would be absolutely crushed in a
simultaneous demonstration by a professional. is that around 1500 in chess?
how little hardware could you get away with and still play at the 1500 level
in chess these days? could my phone do it? could my pc, as a screensaver?
> That's why I believe a super hardware gizmo could easily be built that
> would be in the DAN range somewhere at 19x19, at least low Dan. I'm
> not so bold as to predict that it will be at top human levels any time
> soon though.
i think that we're likely in agreement here. crazy hardware could get you into
the 1 dan range, but professional play is way, way out of bounds at this point.
to see why i think this, watch a 7d game on kgs and listen to the 1d kibitz.
note how ridiculously out-of-touch they are with the game that is going on
in front of them. pro play is yet another magnitude or two of "out of touch"
from
amateur play.
s.
____________________________________________________________________________________
Be a better Heartthrob. Get better relationship answers from someone who knows.
Yahoo! Answers - Check it out.
http://answers.yahoo.com/dir/?link=list&sid=396545433
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
