On Fri, Dec 31, 2010 at 3:49 PM, David Fotland fotl...@smart-games.comwrote:
Monte Carlo go was around for a long time. See Bouzy's papers for example.
The UCT formula for balancing exploration and exploitation came from
research on the one-armed bandit problem, not related to go.
Mogo and
Dear all,
the original MCTS paper is by Rémi Coulom (to the best of my knowledge at
least...). It's clear for us that we did not invent MCTS
and always referenced Remi's paper.
*1) For UCB-like formula:*
- On the theoretical side, the consistency proof of MCTS without the
UCT-like exploration
Hi Jacques
I got a lot of improvement from Rémi's Bradley-Terry
ideas in move prediction (although with some
overlearning which I didn't care much about as
predicting moves is not my interest.) But neither
the naif values (times played/times seen) nor the
improved Bradley-Terry values are
Hi,
I'd like to advise against using the exact algorithm I described in my 2006
paper. I compared it to UCT at that time, and UCT performed better. I am sorry
I don't have a reference to my data any more. I posted the results to the
mailing list. It used to be archived at that link:
On 12/30/2010 08:20 PM, Aja wrote:
Hi Jeff,
When, do you think, did Mogo started dominating all the KGS computer
events and CGOS, and also was the first to extend that dominance from
9x9 to 19x19.?
Hello Aja,
Here I quote from the computer-go archives, unless otherwise noted:
Dec. 31, 2006:
Hi Petr
I guess so. If you want to also to move up to 19x19, try adding
priors to your moves based on the Common Fate Graph distance to the last
move; I think your program should get to at least 5k KGS at that point
if you have everything well debugged.
CFG's is on my list of things to
On 31.12.2010 13:31, Rémi Coulom wrote:
being less selective at the root may improve strength
So what about P; P ca. 3 ply, at each of these plies, a) firstly filter
obviously bad moves, b) secondly consider each still available move's
children by MC / UCT? Or, more generally, dynamically
Hi Olivier,
Thanks for your clarification. Apologize for my wrong description about the
invention of RAVE. I show the mail next again,
Sorry, I might be wrong at RAVE. Maybe it should be: Sylvain proprosed the
idea of RAVE and David Silver proposed a new formula for RAVE.
Aja
-
On 12/31/2010 07:31 AM, Rémi Coulom wrote:
I'd like to advise against using the exact algorithm I described in
my 2006 paper. I compared it to UCT at that time, and UCT performed
better. I am sorry I don't have a reference to my data any more. I
posted the results to the mailing list. It used
Hi Jeff,
Here I quote from the computer-go archives, unless otherwise noted:
Yes, I agree Mogo was so strong in that period (2006/12/31-2009/3/17 or
more). I think it is just similar to the situation when Crazy Stone got gold
medal in Computer Olympiad 2006. Can't we also say this?
Oh!
Hi Jeff,
Maybe I have to reviewed your messages to get your point. I really don't
want to raise any debate in such a happy new year time. :)
Aja
Crazy Stone was strong relatively in that period. It's the same with
Mogo. I don't understand why you speically indicate the period when Mogo
was
Hello all,
... I really don't want to raise any debate in such a happy
new year time. :)
In central Europe it is still old year for many hours.
So, no direct reason to become romantic.
Here are some facts from the ICGA
Computer Olympiads in 2005, 2006, 2007.
September 2005 Taipei
-- Forwarded message --
From: Rémi Coulom remi.cou...@free.fr
Date: Tue, Jul 25, 2006 at 8:22 AM
Subject: [computer-go] Experiments with UCT
To: computer-go computer...@computer-go.org
Hi,
I mentioned UCT in one of my previous messages to the list:
-- Forwarded message --
From: Rémi Coulom remi.cou...@free.fr
Date: Tue, Jul 25, 2006 at 8:22 AM
Subject: [computer-go] Experiments with UCT
To: computer-go computer...@computer-go.org
Hi,
I mentioned UCT in one of my previous messages to the list:
i think that don has best made this point in the past[1], but
championship events are relatively poor predictors of skill because of
their limited number of sample points. something like cgos ranking
over time (among those who participate) is a pretty good way to
compare computer go playing
Hi Remi,
Thanks for the suggestions. Sorry for inaccuracies in my previous
statements. Now I have read your paper more carefully, I find in the
appendex many discussions related to improvements on playout move
selections. On another note, I find that formula you gave on variance
calculation very
That was such an amazingly crazy period. I am sure some day, some one will
write a book (or two) about all of this, so that latecomers get to know all
the details.
Best Wishes!
Fuming
On Sat, Jan 1, 2011 at 12:22 AM, Ingo Althöfer 3-hirn-ver...@gmx.dewrote:
... but
championship events are
Hi Francois,
You mean all the 3x3 patterns? I'm only using 3x3
patterns that occur a number of times in my training
collection.
If 3x3 is the only pattern size you use, that probably does
not apply. But if your pattern system supports other sizes
then you must have implemented some kind of
Hi Fuming,
Sigma is the sum of playout values. If you score 1 for a win, and 0 for a loss,
then, yes, it is the number of wins. At the time of the paper, I was using
territory, so it was not the number of wins. Sigma_2 is the sum of squares. If
outcomes are either 0 or 1, then Sigma and
It would help immensely if the various servers would agree on an interoperable
standard.
Terry McIntyre terrymcint...@yahoo.com
Unix/Linux Systems Administration
Taking time to do it right saves having to do it twice.
From: Olivier Teytaud olivier.teyt...@lri.fr
To: Aja ajahu...@gmail.com;
... basically I think that there would be not so much work around MCTS
without the MC part with a strong influence
of the last move. This is probably the main contribution of MoGo, and *all*
computer-go MCTS are influenced by this, I think.
Most importantly, MCTS is very efficient, for many
They do provide an interoperable standard, GTP. The NNGS competition
provided a script that connected using GTP. KGS provides a client to uses
GTP. The CGOS client uses GTP. Of course some programs choose to implement
only the subset of GTP used by one server, and thus limit themselves.
Brilliant methods almost always arrive after a long period of trial-and-error
stumbling. We call this the hindsight is 20-20 rule.
Terry McIntyre terrymcint...@yahoo.com
Unix/Linux Systems Administration
Taking time to do it right saves having to do it twice.
- Original Message
Hi Olivier,
I fully agree with your words. We were talking exclusively about the
breakthrough of MCTS and UCT. Or I think everyone must admit that you, Yizao,
Remi...etc, all of us, stand in the shoulder of the “giants” to make progress.
I was in a team of Chinese chess program for over 3
24 matches
Mail list logo