As I have spent a lot of time trying to guess what could be done for
Quasi-Monte-Carlo
or other standard forms of Monte-Carlo-improvements in computer-go, I
write below
my (humble and pessimistic :-) ) opinion about that.
Let's formalize Monte-Carlo.
Consider P a distribution of probability.
There is no restriction on how many mogo bots can run. However, there
is not much of a point if everyone is just running the same bot unless
they are running at different levels and we can see exactly how they are
set up.
We have launched 4 mogos, and I explain here what is tested:
-
I'd connect Crazy Stone to CGOS if Many Faces is there.
Mogo will be there also; a 19x19 Cgos would be very interesting
in my humble opinion. The only drawback of Cgos for me is that
we have no idea (at least, I have no idea) of the equivalence
with human standards (kgs rankings are much easier
If needed I can also offer an account on a server with space, webserver
and an account for you to manage the CGOS server, but I don't have the
time to manage it myself.
We (the mogo-team) can install the cgos 19x19 server; we will
keep and admninister it if everything works well with our
Ok for 30 minutes after the testing phase (for the tests
I guess that 10 minutes is too long :-) ).
For the moment I am trying to get the authorization
of opening a port for socket connection -
for the moment I guess only
people in the same laboratory as me can connect to
the machine, what is
I'm not sure what the status of the 19x19 server is, if it looks like
it isn't going to happen I have another option.
Technically it works, but an authorization (for opening the ports
for computers out of the laboratory) is still missing.
But, if someone else wants to install it, no problem
Technical Report on MoGo link from http://www.lri.fr/~gelly/MoGo.htm
http://www.lri.fr/%7Egelly/MoGo.htm eventually leads to
http://hal.inria.fr/docs/00/12/15/16/PDF/RR-6062.pdf
http://hal.inria.fr/docs/00/12/15/16/PDF/RR-6062.pdf
First-play urgency is described in the bottom paragraph on
Thanks to GNU-people who successfully
connected their bot to the server.
The server seemingly works.
cgos.lri.fr, port 6919.
http://www.lri.fr/~teytaud/cgosStandings.html
19x19, 10 minutes per side (for the moment, to be increased).
Olivier
___
The cgos 19x19 server is seemingly ok,
the port 6919 is now opened for all the universe.
The name of the machine is cgos.lri.fr (and not pc5-120.lri.fr as
previously).
The port is 6919. It is 19x19, 10 minutes per side for testing; I will
move to something longer later (depending on what people
Sorry for the trouble for downloading the SGF files
on the 19x19 server; it is seemingly ok now.
Olivier
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
19x19 server:
I have changed 10 minutes to 30 minutes per side.
I have modified the anchors (but the --positional-superko option
is seemingly not recognized; I'll correct that later).
Olivier
___
computer-go mailing list
computer-go@computer-go.org
The anchors are:
/usr/games/gnugo --mode gtp --score aftermath --capture-all-dead
--chinese-rules --level 0
/usr/games/gnugo --mode gtp --score aftermath --capture-all-dead
--chinese-rules --level 10
The numbers (1200 and 1800) are arbitrary; all suggestions welcome,
as for the command-line
The web site displays the wrong time-control. That will be confusing
to people. Can you fix that?
Unfortunately, I can not change things from where I am until wednesday.
I'll fix all I can, starting from November 1st (or before if I can find a
stable internet connection).
Olivier
I don't see Mogo on the server?Where is Mogo?
we will come :-)
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
I have re-launched the cgos 19x19 web-updater for
http://www.lri.fr/~teytaud/cgosStandings.html.
I suggest that bug-reports and comments are made with an
explicit subject or I might miss many of them.
Best regards,
Olivier
___
computer-go mailing list
So Olivier will need to restart when he gets a chance.
Done!
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
Could Sylvain (or anyone who knows) talk about MoGo's pondering
strategy? Does it just build the tree as usual or does it speculate
on some number of moves and hope that the opponent choses one of
those?
MoGo just builds the tree as usual.
Olivier
I was waiting till someone restarts, but nobody seemed to notice.
CGOS was hanging yesterday morning (European time) with 3 games
4849..4851 where no black engine placed any stone.
If black restarted (one of the black bots was mine) it lost on time because
the 30 minutes had been used. Black lost
Unfortunately I do not manage the 19x19 server or I would kill and restart.
I kill and restart in a few minutes.
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
2. Mogo (and CrazyStone) are using lots of intelligence in their
playouts, and that is the cause of the nakade weakness. They are good
players, but they have preconceptions. They consider the moves required
to discover the difference between a nakade and
dead-stones-in-a-definitely-alive-group as
It looks like the server is down again. It's too bad since there were so
many strong programs connected.
I hope it comes back up soon.
I have tried to solve that, but this is seemingly due to a general
failure of the network there (or no more electricity perhaps...),
what is beyond what I can
Due to electricity shutdowns in our university, I will wait a few
consecutive hours
with constant electricity before starting the 19x19 cgos server again.
Sorry for that. Be sure I am in bigger trouble
than you with these electricity shutdowns :-)
Olivier
Cgos 19x19 is back.
I hope electricity is stable :-)
Olivier
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
I will ask Olivier if he wants to incorporate my scripts into the 19x19
server once I have it set up in cron.
The server is now hopefully working properly.
Don, I'll include your script soon; well, I hope so, when I have
a few minutes :-)
The trouble was reported with a title emphasizing the
Any progress on CGOS 19x19?
It appears to have been stuck for 2 weeks now?
Hum, I have not seen these emails before now;
I launch the Cgos19x19 right now, sorry for that.
I have received Don's email about the script, but
I do not use it yet. Minutes are very expensive :-)
Olivier
The event was notable for unjustified resignations.
Seems strange so many bots had resignation trouble. Looks like beta
code. Probably in response to the Please have your bot resign, for
your own good thread?
For mogo, this is a bug in the parallel version of the pondering.
We do all the
It is my opinion that 30 minutes per side is common for human tournament
games, and thus makes sense for 19x19 CGOS. I think 10 minutes is rather
restrictive, so maybe 20 minutes makes sense.
I'll change the time settings to what people want (I'll count the choices
on the mailing list and move
CGOS 19x19 is back.
Following current discussions,
I have temporary moved the time settings to 15 minutes and increased the
gift to 1s per move; when the discussion about time settings will be over, I'll
set the time settings that most people want.
Olivier
[game against MoGo and a professional player]
It will be the MPI version of mogo, and in various board-sizes.
What handicap, if any?
I don't know yet.
___
computer-go mailing list
computer-go@computer-go.org
[game between mogo and a professional player]
What will be the hardware for Mogo ?
The hardware should be provided by Bull
(cluster of multi-core machines, with good
communications - this last point is
very important).
Olivier
___
computer-go
[game between mogo and a professional player]
What will be the hardware for Mogo ?
The hardware should be provided by Bull
(cluster of multi-core machines, with good
communications - this last point is
very important).
By the way I point out that the hardware
for public games is provided by
It's not very clear to me how strong Mogo is at 19x19. I have no idea.
Can't we estimate from KGS games?
k
Due to hardware and library constraints,
mpi-mogo is not on KGS.
Olivier
___
computer-go mailing list
computer-go@computer-go.org
Has this estimation been done?Does anyone here have a sense of how
strong the very best Mogo versions play?
We often compare a version X to a version Y,
and have notable successes for this criterion, for all boardsizes;
but we have only statistics of self-play, and usually with shorter
time
Does someone have positive results
for non-symetrical bandit-based planning, e.g.
using a bandit with more exploration for the opponent
than for oneself.
Best regards.
___
computer-go mailing list
computer-go@computer-go.org
Mogo will just be one data point in the experiment, but an important one
because we will benchmark the same exact version on CGOS.
--nbTotalSimulations 11000 (not high level -- 20 is of course much
stronger but requires more time) instead of --time
no pondering, as you want fixed level
Will Mogo with nbThreads=4 and --nbTotalSimulations 11 yield the
same results as nbThreads=1 and --nbTotalSimulations 11,
presumably in approximately 1/4 the time?
--nbTotalSimulations gives the number of simulations for the first thread;
the others are stopped by time. As threads are
How much memory does mogo require if I crank up the number of
simulations pretty high? Does it allocate dynamically or work from a
fixed pool? What happens if there is not enough memory?
I think you won't have any troubles with that, unless the
hardware is very old.
There is a pruning
Any estimates of when this problem is likely to surface? Is a version
available which is more suitable for greater numbers of simulations?
We can compile that easily, but
I don't know if I can distribute it
(administrativly). To be checked...
Olivier
The crosstables are back, but the sgf archives ar not.
Sorry, many troubles since the maintenance of the website... i'm
on that.
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
[... discussion about bounded size tree...]
Olivier seemed to think it would work acceptably well.
Yes, I think so. We trust the pruning method. If we
are wrong, it's a good piece of news for us - we can improve
the algorithm just by increasing the constants :-)
Olivier
I am trying to connect a bot to a KGS private room.
If we set computer-go as the room, the bot comes and plays against
its opponents.
But if we set the name of the room, the bot comes and can be seen in the
room, but does not play.
Someone has already solved such an issue ?
stupid question :
I have never tested this, but is your bot allowed in this room ?
Yes, and it enters the room.
I've been told that the solution is as follows:
- get in the room under the name of the bot (manually)
- and then, launch the bot with the name of the room in the config file
(the
We have tried many things, using the following documentation:
http://www.weddslist.com/kgs/how/outline.html
but we still fail... someone has an example of kgsconfig which works ?
we have no trouble for connecting a bot to the computer go room.
Olivier
Le lundi 21 janvier 2008, Olivier
Have you selected the room with bot's name as a member?
Yes. Seemingly only public rooms are possible for bots.
I'm interested in if someone has a solution for private rooms.
Olivier
___
computer-go mailing list
computer-go@computer-go.org
I can provide a new release with double instead of float.
(unless the other mogo-people reading this mailing-list do not agree for
this; Sylvain, no problem for you ?).
I don't know exactly when it begins to do bad moves. However, I know that
after several hours, the estimated winning rate
Cgos 19x19 is down, I'm trying to fix
that (more serious trouble
than usually, it is seemingly a
trouble on the machine, which
is the only one allowed
to be connected from outside
the university).
Olivier
___
computer-go mailing list
Cgos 19x19 needs some work due to the change in our website.
It will probably not work before a few days (monday or tuesday
it should be ok).
Olivier
Same error here.
Terry McIntyre [EMAIL PROTECTED]
ÿÿWherever is found what is called a paternal government, there
Please correct me if I'm wrong, but I thought:
Only once you combine the two components does the influence of UCB1 and
UCB1-tuned become less obvious. If you look at just the RAVE success ratio
component, or just the success ratio component, I believe the UCB1-Tuned
formula is still present.
Why are m and n different? Isn't every playout used both to update the UCT
win rate and the RAVE values for the same nodes? Won't the number of UCT
simulations and the number of RAVE simulations be the same?
Each playout is used both to update the UCT win rate and the RAVE
values for the
For people requesting mogoRelease3 without the bug for long computation
times due to a float instead of a double:
http://www.lri.fr/~teytaud/mogo (32 bits version, with double instead of
float)
http://www.lri.fr/~teytaud/mogo64 (64 bits version, double also)
I'm just surprised to hear that the program that introduced UCT (and got
so many others to use it), isn't using UCT any more. Combining RAVE and
UCT as described in the PDF still sounds like UCT to me, but with no
sqrt(log) term, it no longer is. I'll certainly have to think about the
trades
- at each (or every n) iteration you add one node.
As far as I see, new nodes are created only if new nodes are visited;
if
score(visited nodes) score(unvisited nodes)
why would mogo visit new nodes ?
But (before the recent PDF file) I never understood completly
the bandit in mogo, so
I can't tell if you mean the float version or the double version. Using the
float version (since it was all I had), I did a fairly extensive analysis of
the losing move from the MoGo game that Fotland added comments to. My
results were posted to this list on 2/1/08 under the subject, UCT and
I can't tell if you mean the float version or the double version.
Using the float version (since it was all I had), I did a fairly
extensive analysis of the losing move from the MoGo game that Fotland
added comments to. My results were posted to this list on 2/1/08
under the subject, UCT
easiest way to use the power of the gpu) the result should be more meaningful
too. Has anyone done any test like that (like use gnugo level 0 instead of an
MC-playout) ? Does anyone have a minimalistic non-MC go engine I could look
at ? One more thing - has anyone tried using quasi-MC for go ?
Sylvain wrote:
Thinking a little more about it, I think we have to add an hypothesis
which is that, for a given move, the number of AMAF updates if alpha
(nb total UCT updates), with alpha 1. That seems to hold for most of
the updates (with alpha close to 0.5), but there may be cases where
A new position is always visited unless the leaf of the tree is the
end of the game. In that case, one player always win, so the other
always win. Then, the losing player will explore all the other moves
to avoid the sure loss. If all moves are still loosing, that will
propagate to the move
46 E5 C5
By the way, mogoRelease with moderate computation times often plays
E5-C5-C6, and this third move C6 is very weak (at least I've
been told so :-) ).
___
computer-go mailing list
computer-go@computer-go.org
We don't (at least up to the release, I don't know everything they are
doing now). Using the floats was for flexibility, because we did try
mogo still has 0-1 reward (at least, our version of mogo).
Olivier
___
computer-go mailing list
That assumes creating a situation where the nakade is misevaluated once
you're behind is easy. Writing code that exploits this weakness may be harder
than improving your program through other means. And the 'gain' could be
short-lived when MC programs fix the weakness.
By the way,
For information on the mogo/pro challenge:
- during preliminary tests, mogo has won 4/0 against a very high level
human; at that time we were just very very very happy :-)
- some other humans, supposed to be weaker, have however
won some games at that time (before the nakade correction);
-
How well does the nakade improvement perform on 13x13?
no idea on 13x13, but it does not work on 19x19 (seemingly,
perhaps we just need tuning...).
Also, it works only, in terms of success rate against the old
mogo, for sufficiently large number of simulations per move.
Olivier
Saturday:
3/23/08 3:00 PM
Saturday:
3/22/08 3:00 PM
is right?
Hi; it's saturday 22.
Olivier
(stress++)
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
will this be with komi 7.5?
Yes. Previous records against Guo Juan, as far
as I know:
- 1/3 wins with komi 7.5
- 9/14 wins with komi 0.5 (mogo black,
i.e. komi in favor of mogo)
Best regards,
Olivier
___
computer-go mailing list
What computing power did have that MoGo at its disposal?
4 cores, 2.4 GHz.
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
Has the program become that much stronger on 9x9 recently?
(Compared to the version was trying?)
*Parallelization: MPI == ~80% vs no mpi in 9x9 (for same number of
cores).
*Monte-Carlo improvement == strongly depends on number of simulations
and number of cores (as the multi-core reduces
The million dollar question: How well does Mogo scale on this number
of processors?Can you give us at least some kind of generalization?
unfortunately, using more than 10 nodes is probably not very very useful
in 9x9, for the moment - but we have not tested that sufficiently,
and we have
Would you guess that mogo is 2 or 3 ranks stronger at 19x19 with all
this hardware?
I just claim that mpi-mogo wins with very high probability against
sequential-mogo in 19x19. But I'm afraid that the improvement is
disappointing against humans.
I hope better improvements are possible
It was 2 cores 2.6GHz. (intel core2 duo).
sorry, I believed it was the tipi.
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
Hi,
I have got a lockless hash table to work, and I thought I'd share the
results.
[...]
Great! For networks of 4-cores, it is not very useful,
but for highly smp machines it could be great - with your
grid5000 account, you might run crazystone on a
16-core machine and have a very impressive
congratulations to mogo on its performance today!
it was an excellent result (1-2) versus a professional,
Thanks a lot. MoGo has also played some games against other
players (including a professional player) and I hope people
will accept that we publish the results (they are better than
I have put a report of the weekend's challenge games between MoGo and Catalin
Taranu 5p at http://www.computer-go.info/tc/
mainly to make it easier for people to find the game records.
Thanks a lot for that.
Some points are wrong however, below some informations about the errors.
The
The hardware in case of trouble, which has been used for two games, is
provided by Université Paris-Sud.
Precisely: LRI, Université Paris-Sud.___
computer-go mailing list
computer-go@computer-go.org
I am now wondering if scalability could be unaffected by playouts (just
adding a constant offset) and only depend on the UCT/search implementation.
From the publications of the MoGo team it seems likely that the programs are
very similar there.
Leela and mogo are probably quite similar.
On
simulations is perhaps not always a good idea.)
Is the patch in some way parameterized by the number of simulations?
No. Perhaps it should :-)
___
computer-go mailing list
computer-go@computer-go.org
light-playout variant of leela, but perhaps the
nakade-patch version of mogo and maybe even some third
no problem for the nakade-patch version of mogo, but results
are only known in 9x9, no idea for 13x13. Maybe it is better,
maybe it is worse :-)
___
Good to find out, no?
we have validated that:
- it is good in 9x9;
- it is bad in 19x19 (unless perhaps for very large number of
simulations).
we did not have a look at 13x13.
___
computer-go mailing list
computer-go@computer-go.org
In any case, later versions of MoGo used (still use?) an exploration constant
of zero, which means that the UCB formula is no longer used.
Still use. But I'd like to check that it's a good idea, in particular
for the parallel version.
For example, most of our openings are useless for the
It looks like we have a clear trend now. Light play-outs do not scale
as well as heavy play-outs.
I don't know if the data are sufficient for this conclusion, but another
element is that heavy playouts are seemingly easier to parallelize than
light playouts. This is tested clearly in
Just a small contribution to the discussion
on the scalability of various forms of
playouts:
for the success rate of our new playouts in MoGo,
the percentage of wins against the baseline version increases with the
number of sims/move:
10 000 49.3 +- 1.2 \%
100 000 73.7 +- 2.9\%
200
The results confirm that Valkyria still benefits from using confidence bounds
with UCT, although the effect is really small.
The standard deviation is a bit large for concluding.
I'll try to get similar numbers for mogo. For the moment
everything leads to 0 as the best constant, but perhaps
it
One reason of this discount is that the MoGo bros running on cgos are
the _big versions. By my obserbation (they are running on my pcs and
both are Q6600/3GHz with different mother boards), mogo_big_4core's
perallelism is around 300% (by top command), perhaps due to its
heavier uct part (just my
Personally, I think that much of the really high quality issues are not
that important for MC Go right now. I think that other things like not
having a reasonable distribution function (which UCT does a remarkable
job of smoothing over) completely overwhelm the effects of a poor PRNG.
I agree
1) does not hold in mogo anymore, because there's no upper-confidence terms
in MoGo anymore.
I just want to add that this is also true in many codes (either no
upper-confidence term, or with a very small term and no statistical
advantage on the null case).
I had read previously that Mogo used MPI, but I didn't know
if could work on clusters without sharing the whole tree. I have been
formulating such an algorithm for the past week or so, so I would like
to make sure that it hasn't already been done. Has anything been
written about Mogo's parallel
If I understand correctly what it is about, I do have something in my thesis
about that.
p153: 4.4.3 Mathematical insights: strength VS accuracy, and more
precisely Theorem 4.4.1 (Accurate simulation player without memory is
strong)
It is certainly not of direct practical application though.
Thanks. Can you tell us how many CPUs were used?
32 octo-cores.
As far as I remember, it was infiniband, and the machines were 3GHz.
I point out that the switch is important. My speed-up curves
in a previous email are probably pessimistic for the 9x9: some
networks are better than those
Yes, the 19x19 server is down.
It's up and running now.
Olivier
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
Kim :-) ).
Best regards,
Olivier
--
=
Olivier Teytaud (TAO-inria) [EMAIL PROTECTED]
Tel (33)169154231 / Fax (33)169156586
Equipe TAO (Inria-Futurs), LRI, UMR 8623(CNRS - Universite Paris-Sud),
bat 490 Universite Paris-Sud 91405 Orsay Cedex
Yes, and then 19x19 with handicap.
On Aug 25, 2008, at 10:47 PM, Olivier Teytaud wrote:
Just for information, mogo will play in a few minutes (on Kgs /
computer-go) some games
against high level humans.
MogoTitan is playing 9x9 against nutngo ?
Christoph
and the player :-)
Best regards,
Olivier for the mogo-team
Yes, and then 19x19 with handicap.
On Aug 25, 2008, at 10:47 PM, Olivier Teytaud wrote:
Just for information, mogo will play in a few minutes (on Kgs /
computer-go) some games
against high level humans.
MogoTitan is playing 9x9
By my recent experiments, 8~9 * (threads - 1) ELO is lost. This
matches my earlier result well.
Do you have tricks for avoiding redundancies between simulations ?
I suggest simple tricks like do not go to node X if there is a thread
currently in node X
(simply by setting the score of the
Yes. I use Sylvain's fpu and decrease it a little before starting a
simulation, say, fpu *= 0.99. This is very simple and fast.
Ok. Perhaps I'm wrong (I might misunderstand your solution and I might be
wrong
whenever I've understood :-) ); but
- I think that this does not avoid
Although I'm parallelizing in not SMP systems but a cluster of loosely
coupled (small) computers connected through moderate speed networks
using broadcasting positions, this may not change the vlaue of
avoiding redundancies. I'll study more when implementing
pre-knowledge or some. Thanks.
MC is playing most goal-directed (zielgerichtet
in German) when the position is balanced or when
the side of MC is slightly behind. However, when
MC is clearly ahead or clearly behind it is playing rather
lazy.
At some point we were investigating that here, but only on small sets of
games
I made a change over the weekend, which looks like it makes 9x9 150 ELO
weaker and 19x19 over 200 ELO stronger.
We have plenty of size-dependent parameters and plenty of if
(boardsize==19) in MoGo for things like that :-)
___
computer-go mailing list
mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
--
=
Olivier Teytaud (TAO-inria) [EMAIL PROTECTED]
Tel (33)169154231 / Fax (33)169156586
Equipe TAO (Inria-Futurs), LRI, UMR 8623(CNRS
The bright side here is that 9x9 is not really important but just
a test bed. If it works for 19x19, that's good.
People moderately intested in Go could also claim that both 9x9 and 19x19
are
just testbeds for power plant management :-)
In my humble opinion, both are intesting, both as
- There had been a TV program of professional 9x9 Go for years (some
member of this list have the records of the games played in this
program). Takemiya 9p and Yuki 9p were the strongest.
I'm afraid the answer is no, but:
are these records free and available somewhere ?
Thanks for your
Dear all,
the results of the 9x9 computer-go event in Taiwan (including a 9x9
competition and games between humans and computers)
can be seen at
http://go.nutn.edu.tw/eng/main_eng.htm
(see news)
These games were organized by the National University of Tainan and the
Chang Jung Christian
1 - 100 of 188 matches
Mail list logo