Le dimanche 8 avril 2007 03:05, Don Dailey a écrit :
> A few weeks ago I announced that I was doing a long term
> scalability study with computer go on 9x9 boards.
>
> I have constructed a graph of the results so far:
>
> http://greencheeks.homelinux.org:8015/~drd/public/study.jpg
Thanks for th
Paper 1 in the list below states:
Numbers were originally implemented in Lisp I as a list of atoms.
and the Lisp 1.5 manual states: Arithmetic in Lisp 1.5 is new
Could you give an example how the number 3 was implemented in Lisp-1 and how
2+1?
So far I have found only this remarks but not
Thanks dons for producing these fascinating results. I hope that
when you have finished the study, you will show us not just this
graph, but also the game results (number of wins) that it is
derived from.
At 02:05 08/04/2007, you wrote:
A few weeks ago I announced that I was doing a long term
According these results the slope is considerable greater than in chess. In
the classical experiment of Ken Thompons searching 1 ply deeper is worth
about 200 Elo. 1 ply corresponds to 5-6 times longer/faster. In 9x9 already
a factor of 2 gives the same improvement. This is really remarkable. An
The discussion here http://senseis.xmp.net/?EloRating suggests that
the difference between beginners and top players in go is about 3000
ELO on a 19x19 board. This difference is very dependent on the board
size. I can
think of a naive argument that this difference should scale linearly
with t
On Sat, Apr 07, 2007 at 04:03:47PM -0400, Don Dailey wrote:
>
> On Sat, 2007-04-07 at 14:36 -0400, Matt Kingston wrote:
> > What I want from a commercial go playing program is one that I can use
> > to learn to be a better go player. [...]
>
> I have heard this argument before, but I personally d
On Sun, 2007-04-08 at 00:44 -0400, [EMAIL PROTECTED] wrote:
> I question whether it's valid to make this kind of comparison when
> Gnugo scales so differently from UCT. If you froze one of the UCT
> prgrams at 1 million playouts/move and then tried to scale gnugo until
> it matched that level of s
On Sun, 2007-04-08 at 09:43 +0200, alain Baeckeroot wrote:
> Could'nt the inflexion of heavy curve also mean that the advantage of
> heavy play-out disappears when the number of simulation is very high ?
> With huge number of simulation the heavy player could become weaker
> than
> the light player
On Sun, 2007-04-08 at 09:56 +0200, Chrilly wrote:
> Is it just enough to make a 2 million playouts version
> to beat the top-Dans in 9x9? Is it that easy?
Of course the ELO numbers are arbitrary. I assigned GnuGo 3.7.9
a level of 2000 but on CGOS it is 1800.But CGOS numbers are
arbitrary t
On Sun, 2007-04-08 at 09:36 +0100, Tom Cooper wrote:
> Thanks dons for producing these fascinating results. I hope that
> when you have finished the study, you will show us not just this
> graph, but also the game results (number of wins) that it is
> derived from.
I have all games and all data
On Sun, Apr 08, 2007 at 08:32:48AM -0400, Don Dailey wrote:
> Eventually, the curve would have to close back up as both versions
> approach perfect play. So even if the gap widens for a while, this
> cannot stay forever because in the end they will touch.
Aren't you being a bit optimistic here? I
On Sun, 2007-04-08 at 11:24 +0200, Heikki Levanto wrote:
> > In fact this is how beginners think about the game. It doesn't
> > seem to me like a good learning aid to try to get the computers
> > to "emulate" the losing strategy weaker players use.
>
> Weaker players can not estimate the scor
On Sun, 2007-04-08 at 14:44 +0200, Heikki Levanto wrote:
> Aren't you being a bit optimistic here? It is quite conceivable that
> the
> curves will flatten out and reach a maximum level somewhat below
> perfect
> play. I don't see how we can predict the difference between them at
> that
> time.
U
On Sun, Apr 08, 2007 at 08:48:03AM -0400, Don Dailey wrote:
>
> On Sun, 2007-04-08 at 11:24 +0200, Heikki Levanto wrote:
> > Weaker players can not estimate the score until very late in the game.
> > Not with enough precision, anyway. Thus, most of the time they have no
> > idea if they are winnin
The question here is not about UCT(yes, it gaves the same rusults as
alpha-beta). It's about MC scoring. It has not been proved that MC score will
generate the optimum play with large enough simulation.
Now the best super computer uses about 500,000 CPUs, which is 2 to the 18th
power of comput
The factorial of 81 is about 10^140. The number of legal positions may be some
roots of that. It's still a huge number. The number of simulations used in UCT
grograms is several hundred thousand, which is tiny compared to the total
possible number of plays. How can they still be so successful? I
On 4/8/07, [EMAIL PROTECTED] <[EMAIL PROTECTED]> wrote:
The factorial of 81 is about 10^140. The number of legal positions may be
it may be 103919148791293834318983090438798793469
regards,
-John
___
computer-go mailing list
computer-go@computer-go.or
Hello Don,
A few weeks ago I announced that I was doing a long term
scalability study with computer go on 9x9 boards.
I have constructed a graph of the results so far:
Your work and insight keeps on amazing me.
If I understand is correctly the playouts are made all from the
root position?
I
On 4/7/07, Don Dailey <[EMAIL PROTECTED]> wrote:
Of what learning purpose is it if you are losing the game and
the computer gets you to focus on a dramatic battle somewhere
that is totally irelevant?
I think the issue is that once the computer has a won position, all
moves seem almost equally
You don't have to see to the end of the game to play well.
They have done studies in computer chess which show the
number of times a deeper search changes it's mind. It
becomes smaller and smaller with depth - presumably because
most of the moves are already optimial.
Also, it is clear (in
On Sun, 2007-04-08 at 18:11 +0200, Edward de Grijs wrote:
> Hello Don,
>
> >A few weeks ago I announced that I was doing a long term
> >scalability study with computer go on 9x9 boards.
> >I have constructed a graph of the results so far:
>
> Your work and insight keeps on amazing me.
>
> If I u
I'm not really fanatic about this either way. If I knew how to
easily "fix" this, I probably would provide a mode to make it
work either way.
I see it as aesthetically MORE pleasing when the program appears
to be playing stupid, but then you realize that YOU are the one
that doesn't understa
On Sun, 2007-04-08 at 10:09 -0400, [EMAIL PROTECTED] wrote:
> The question here is not about UCT(yes, it gaves the same rusults as
> alpha-beta). It's about MC scoring. It has not been proved that MC
> score will generate the optimum play with large enough simulation.
MC is obviously wrong as a
Just a warning -
Tomorrow, I'm taking the old CGOS down and replacing it with
the NEW cgos.
Also, the CGOS test I'm running from my home computer will
go away. So the 2 minute server will also go away!
The new cgos on boardspace will be 5 minute time control
instead of the previous 10 minut
I will try to explain it better:
What had puzzled me until recently is how to combine two facts:
a. As Jason says, "A single MC playout corresponds to a Bernoulli
trial with probability p" and therefore, p-hat is distributed as a binomial
which converges to the normal.
b. The playout itself is
I think you are right, because MC score becomes precise when only a few
available moves left. However, do you search to the depth of the end of the
game, or to the extent that MC score becomes precise?
-Original Message-
From: [EMAIL PROTECTED]
To: [EMAIL PROTECTED]
Cc: computer-go@co
Don Dailey wrote:
I have this idea that perhaps a good evaluation function could
replace the play-out portion of the UCT programs.
I thought about something similar but only for initializing the
counters: introduce 10 fake playouts and estimate the number of
wins by a function returning somet
> I have this idea that perhaps a good evaluation function could
> replace the play-out portion of the UCT programs.
I thought about something similar but only for initializing the
counters: introduce 10 fake playouts and estimate the number of
wins by a function returning something in [0, 10]. A
On 08/04/07, Jacques Basaldúa <[EMAIL PROTECTED]> wrote:
I will try to explain it better:
What had puzzled me until recently is how to combine two facts:
a. As Jason says, "A single MC playout corresponds to a Bernoulli
trial with probability p" and therefore, p-hat is distributed as a binomi
> A few weeks ago I announced that I was doing a long term
> scalability study with computer go on 9x9 boards.
> http://greencheeks.homelinux.org:8015/~drd/public/study.jpg
Hi Don,
This is very interesting. I envy your persistent energy and enthusiasm.
If you'd stopped at level 6 or 7 the concl
> ...XCell Journal. Search on the net for "Lorenz, Donninger, Hydra" and
> format "pdf". But in this papers the concept is only mentioned ...
Thanks Chrilly. For anyone else interested, it is here:
http://www.xilinx.com/publications/xcellonline/xcell_53/xc_pdf/xc_hydra53.pdf
But, as you say, the
On Mon, 2007-04-09 at 09:48 +0900, Darren Cook wrote:
> > A few weeks ago I announced that I was doing a long term
> > scalability study with computer go on 9x9 boards.
> > http://greencheeks.homelinux.org:8015/~drd/public/study.jpg
>
> Hi Don,
> This is very interesting. I envy your persistent
32 matches
Mail list logo
