Your point is obvious but that's a horrible proof since there are usually more
than one legal moves from which to chose (that means it takes more time).
steve uurtamo wrote:
As for other things we'd like to see improved, we could build a list. My pet
peeve is the KGS "score estimator", which is
Even though KGS is not "open", you can still reverse engineer it, right? Why
not create an accessible web interface to KGS?
terry mcintyre wrote:
If "accessibility" is the only criterion, a client would do the trick;
it would need an open protocol.
It's been a bit of an inconvenience that KG
Never mind, the problem was on my side.
Michael Williams wrote:
I'd like to read it, but that link doesn't work for me. Is it correct?
Brian Sheppard wrote:
I recommend the paper
http://hal.inria.fr/docs/00/36/97/83/PDF/ouvertures9x9.pdf by the Mogo
team,
which describes how to
I'd like to read it, but that link doesn't work for me. Is it correct?
Brian Sheppard wrote:
I recommend the paper
http://hal.inria.fr/docs/00/36/97/83/PDF/ouvertures9x9.pdf by the Mogo team,
which describes how to use a grid to compute Mogo's opening book using
coevolution.
Brian
___
Darren Cook wrote:
GoChild is designed for learning how to play Go efficiently.
Here is the English link for Michael :-)
http://www.gochildgame.com/en/index.htm
Thanks! I know how his wife feels -- I'm also still trying to figure out what
to do after Ko is created :/
___
Can you tell us English speakers anything about it?
幼儿围棋 wrote:
Hi All,
GoChild is designed for learning how to play Go efficiently.
Web App - http://gochild2009.appspot.com
Official site - http://www.gochildgame.com
Regards,
gosharplite
Jacques,
It has been some time since you made this. Did you have to make changes to any of the original Fuego files? I'm asking because I'm trying to figure out what
would go wrong if I dumped current Fuego files into the Windows-buildable source that you provided.
Jacques Basaldúa wrote:
I
Very nice. Do you (or anyone) have the details of "(Boon, 2009)" concerning
arbitrarily shaped patterns?
Petr Baudis wrote:
Hello!
Today I held a presentation on the "complete" state of art in
Computer Go, showing hopefully all the current widely applicable
results, most successful strate
If you thought finding Go(game)-related stuff on the web was hard before...
Ben Shoemaker wrote:
Has anyone tried programming Go in the "Go" Programming Language?
http://golang.org/
___
computer-go mailing list
computer-go@computer-go.org
Very interesting stuff. One glimmer of hope is that the memory situations
should improve over time since memory grows but Go boards stay the same size.
Christian Nentwich wrote:
Mark,
let me try to add some more context to answer your questions. When I say
in my conclusion that "it's not wo
Can someone refresh me on the concept of fill board?
David Fotland wrote:
I finally found time to try fill_board, and it made the program a little
weaker.
I tested against gnugo level 10, 19x19 board, vs Many Faces with 10K
playouts per move.
Before fill_board it won 1183 of 1411, 84%, 1.9% 95
e the output actually changed, it
would still be nice to see a few more lines and see the trend.
Thanks for helping ... debugging for someone else must not be the best
use of your time,
René
On Tue, Aug 25, 2009 at 8:03 PM, Michael Williams
mailto:michaelwilliam...@gmail.com>> wrote:
ng "8*m + n" with "8*m + (32-n)" on the last line may do the
trick for you, but it breaks my local copy). So, now I am positively
baffled.
Help?
René
On Tue, Aug 25, 2009 at 7:25 PM, Michael Williams
mailto:michaelwilliam...@gmail.com>> wrote:
Debug build:
t; m << "\n";
}
and add a line something like this in the main program:
if (test && sandbox) test_sandbox (); /* new line */
if (test && verify) test = test_bitmap_t::test (); /* already there */
That does not solve anything, but it may tell me if those functi
unsigned long n;
_BitScanForward (&n, board.m128i_u32[m/4]);
return 8*m + n;
}
René
PS. x64 ... new playground, you could use __lzcnt64 ... so many things
to try!
On Tue, Aug 25, 2009 at 5:43 PM, Michael Williams
mailto:michaelwilliam...@gmail.com>> wrote:
(AMD Phenom CPU)
(AMD Phenom CPU)
Michael Williams wrote:
It came through fine the first time. Gmail probably hid the duplicate
text from you for convenience since it saw that it was text that you
already sent out. Or something.
I can compile the original (9x9) bitmap Go files in Windows 7 x64, but
when I
It came through fine the first time. Gmail probably hid the duplicate text from you for convenience since it saw that it was text that you already sent out.
Or something.
I can compile the original (9x9) bitmap Go files in Windows 7 x64, but when I
run it I get this:
test-error [bitmap_t::low
For many people on this list, Computer Go is a hobby and that means that it is important to do whatever you find interesting and motivating, event if it may not
be the best or most promising direction in order to become the next world champion program. There is room for pushing limits in all dire
If an engine (or an engine's playout policy) needs to check the legality of
every move before making a selection, this could still be a benefit.
René van de Veerdonk wrote:
David,
Confession: I have not tested 19x19. As you have noted, and others
before you over the years, a 19x19 board does
ted pretty well, but there's no chance on 19x19 to do so.
Am 13.08.2009 um 05:14 schrieb Michael Williams:
After about the 5th reading, I'm concluding that this is an excellent
paper. Is anyone (besides the authors) doing research based o
After about the 5th reading, I'm concluding that this is an excellent paper.
Is anyone (besides the authors) doing research based on this? There is a lot
to do.
David Silver wrote:
Hi everyone,
Please find attached my ICML paper with Gerry Tesauro on automatically
learning a simulation po
Unpack to C:\ETC\Fuego in order for it to compile. Or modify the list of
include folders in the project configuration.
Jacques Basaldúa wrote:
I have made a native Windows Fuego build compiled with MS Visual Studio
2008.
Thanks to the Fuego team for making such a nice program free software
Very nice! Thank you.
Jacques Basaldúa wrote:
I have made a native Windows Fuego build compiled with MS Visual Studio
2008.
Thanks to the Fuego team for making such a nice program free software !!
I will use it to measure some tree metrics to tune my own program and
for a validation
experim
Although I've never done a RAVE implementation (soon, very soon), this is
related in the sense that AMAF is related to RAVE:
I have recently (yesterday, actually) done some experiments on AMAF with 5-vertex patterns (normal AMAF can be considered 1-vertex patterns). I was not able to
observe a
You should set the limit to whatever yields the highest ELO in YOUR program.
Harry Fearnley wrote:
Darren Cook wrote:
The largest nakade shape is the rabbity six. My wild guess would be to
outlaw self-atari for groups of 7+ stones.
The fun thing about computer go is how hard it is to make har
If your weights are all between 0 and 1:
do
double r = rand(0 to 1)
int i = rand(0 to weightCount - 1)
until weight[i] > r
I think that's right.
Mark Boon wrote:
When using patterns during the playout I had improvised some code to
select patterns randomly, but favour those with higher w
n SGF file showing the dominant variations
with the number of wins and losses? It'd be interesting to see what the
bot considers to be the best sequences are...
Sent from my iPhone
On Jul 10, 2009, at 10:10 PM, Michael Williams
wrote:
Now that I have this system of generating really big
The CGOS page should have a list of the GTP commands required by CGOS.
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
So far, E5 and E6 are neck-and-neck for the favored opening move (E5 was leading almost the whole time until recently). The deepest part of the tree is 35 ply.
About 18% of run time is spent swapping to disk. 40% is traversing the tree. 14% is Libego.
Michael Williams wrote:
Now that I have
Now that I have this system of generating really big game trees, what sort of interesting things could I do with it? The exact number of nodes I can store is
not exact because I'm doing various things to reduce each node's footprint when it goes to disk. I'm currently building a tree that is bus
Prediction:
First, developers and algorithm gurus will expend huge amounts of effort to
parallelize code to take advantage of multi-core chips.
Then, hardware engineers and physics gurus will give us light-based CPUs and
orders of magnitude improvement in single-core performance.
other words, I would perform a mark-and-sweep garbage
collection.
Peter Drake
http://www.lclark.edu/~drake/
On Jul 6, 2009, at 8:02 PM, Michael Williams wrote:
If you are talking about 128k nodes, I don't think traversing them
will take
If you are talking about 128k nodes, I don't think traversing them will take
very long.
Peter Drake wrote:
On Jul 6, 2009, at 9:26 AM, Isaac Deutsch wrote:
If you're hashing by chaining, you presumably go to the appropriate
slot in the table, then traverse the (short) linked list of nodes
h
That's the beauty of MC! It really is a beautiful system. Initially, they were totally random playouts. But the more powerful MC Go engines do not do
completely random playouts; they do "heavy" playouts. But if you were to look at a heavy playout, it would still look like a very weak Go player
Suicide? Do you mean self-atari? But there must be more to it than that
because you don't have a rule that prevents all self-atari, right?
David Fotland wrote:
X can't fill J6 because that would be suicide. For the moment, H5 is an
eye.
David
-Original Message-
From: computer-go-b
nterested in that sort of feature, I'll put it in.
p.s. server crashes and disconnects are already handled by the clients.
Severe engine crashes are a different story. Some of those can't be
handled, hence the heartbeat messages idea.
On 24/06/2009 17:52, Michael Williams wrote:
A wh
A while back, I wrote a general purpose monitor tool that checks a specified URL for a specified RegEx and takes a specified action based on the result. The
action can be play a sound or send an email or flash the tray icon. Something like that might work (it would be much easier if the CGOS pag
Turn off Windows update or put the CGOS connect script in the startup folder
and set an automatic login.
David Fotland wrote:
I can have a reduced version of Many Faces up all the time on an old
computer, but I don't monitor it, so someone would have to email and remind
me when it goes down (u
If it were me, I'd run all anchor candidates against the current CGOS to
determine the anchor value to use for that anchor candidate.
Hideki Kato wrote:
I'm running Fatman1, GNU Go and GNU Go MC version for 9x9 and two
instances of GNU Go for 13x13, five programs in total, on a dual-core
Athl
Fuego gets an advantage because when it plays the anchor, it is playing a version of itself. That's bad for the same reason that it's bad to test against a
version of your own program -- inflated results. But I don't think it's a big deal. What about using both Fuego and Mogo as anchors? Don't
ext/pattern for that position at
that node.
- Dave Hillis
-Original Message-
From: Michael Williams
To: computer-go
Sent: Mon, Jun 22, 2009 3:02 pm
Subject: Re: [computer-go] 1x1 patterns?!
I had never considered using AMAF with larger pattern. That's an
interesting idea. Perhap
I had never considered using AMAF with larger pattern. That's an interesting idea. Perhaps a 5-vertex cross-shaped pattern or a 3x3 pattern. Has anyone tried
this?
Magnus Persson wrote:
Probably 1x1 patterns implies that different priorities are assigned to
the absolute position of empty mo
Are the games archived? Does the public have access to those archives?
Ingo Althöfer wrote:
Hello,
some time ago I had asked if discussions on
"computer Havannah" were welcome here in the list.
The reactions were positive, but (by different
reasons) actors preferred not to use the opportunit
Is it pronounced few-go or fway-go?
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
Section 3.2 describes a pair of tests that took about 4.2 minutes each (if my calculations are correct). Why not play more games and have each game contain
more simulations? Writing the code and the paper is the hard part, waiting for a computer to run your code is easy.
Peter Drake wrote:
An
I vote for 2 venues, each optional. Separate rating pools is a must.
Łukasz Lew wrote:
Maybe we could agree that 1 day out of 7 in a week would be played on
6 times faster time controls.
The same bots, connections, logins, the same number of games per week.
Different rating of course.
This wou
Cool, but what is the cost? The website doesn't even have a word about how to
obtain thier product. Insane, IMO.
http://www.google.com/products?q=tilera returns ONE (useless) hit.
Or is this chip not yet released?
David Fotland wrote:
Lots of simpler cores is possible, but only for running
I know I wasn't very clear with what exactly I tested, but I tested several
variations and could find nothing that worked. And I was ignoring the runtime
cost.
Michael Williams wrote:
I tried it in my previous engine, which while probably littered with
bugs, at least had the character
I tried it in my previous engine, which while probably littered with bugs, at least had the characteristic that more playouts lead to better play. This test
was extremely slow and produced a weaker bot. YYWV.
I'm 95% sure Don mentioned it on this list a couple of years ago, but I don't
know i
Another strategy to be considered is to not allow the thinking to cease until the maximum win rate and the maximum visit count agree on the same move.
Obviously this requires some extra code to make sure you don't lose on time, etc.
Brian Sheppard wrote:
When a UCT search is completed, the usua
i always buy at the low end because you get so much better of a deal. But I'd
guess the i7 is as excellent at Go as it is at everything else.
Łukwasz Lew wrote:
Hi
I have few days to buy a computer for Monte-carlo Go program.
There is not enough money for a multi processor, so I have to decide
Łukasz Lew wrote:
On Wed, Jun 3, 2009 at 00:56, Michael Williams
wrote:
Two things: Firstly, I'm storing (only in RAM) the precalculated Winrate
and InvSqrtVisits and keeping them updated.
So my UCT formula went from
Wins / Visits + sqrt(lnParentVisits / Visits)
to
Wi
There will be multiple time controls, but they will be in sync, so that your
program can always play in a shorter time control game without missing a
game at the longer time control.The idea is to keep your bot busy while
waiting for future rounds.You play in the longest time control, but
Jason House wrote:
On Jun 2, 2009, at 6:56 PM, Michael Williams
wrote:
Two things: Firstly, I'm storing (only in RAM) the precalculated
Winrate and InvSqrtVisits and keeping them updated.
So my UCT formula went from
Wins / Visits + sqrt(lnParentVisits / Visits)
to
Wi
2047/2048. Anytime a proper UCT loop occurs, the likely-best reference is updated (about
90% of the time there is no change, so I think it's safe).
Jason House wrote:
That sounds like a good optimization. What did you do?
Sent from my iPhone
On Jun 2, 2009, at 3:16 PM, Michael Williams
I mean "SSD".
Michael Williams wrote:
Update: After concentrating on tightening the UCT loop, I've optimized
myself back into needing the SDD :/
But now I should be able to get to 20B nodes in just one day.
(still only doing 7x7 Go)
Michael Williams wrote:
Yes, when me
Update: After concentrating on tightening the UCT loop, I've optimized myself
back into needing the SDD :/
But now I should be able to get to 20B nodes in just one day.
(still only doing 7x7 Go)
Michael Williams wrote:
Yes, when memory is full, I save and free all leaf nodes (which i
Yes, when memory is full, I save and free all leaf nodes (which is the vast
majority). Nodes are loaded as needed.
Don Dailey wrote:
On Mon, Jun 1, 2009 at 4:57 PM, Michael Williams
mailto:michaelwilliam...@gmail.com>> wrote:
I've optimized my disk access to the poin
I've optimized my disk access to the point where I'm mostly CPU limited now, even when using a standard hard disk instead of an SSD. I can now create trees of
up to about 30 billion nodes, which would take about a week. The simulation rate is continuously going down because so much time is spent
Back in the days of Don's reference bot, I made some similar modifications to
them and posted the results (slight improvement). The postings were around
10/28/08.
Ingo Althöfer wrote:
Hello,
maybe this is old stuff for the insiders.
In my lecture today a clever student (Hagen Riedel)
brou
Generally, the playout stops when there are no more "valid" moves remaining. Where "valid" means not playing inside 1-point eye. You can also terminate a
playout if one player is winning by a large margin (known as the mercy rule). And you must set a max playout length because some games will g
MoGo was inspired by Crazy Stone? I've never heard that before.
Ian Osgood wrote:
On May 23, 2009, at 3:17 AM, Joshua Shriver wrote:
I know with the Chess community, it's looked down upon to use others
code w/ respect to competing in tournaments. I'm curious, how is it
with Go?
Even more
I'm currently concentrating on the opening. I can now build a UCT tree of about 14 billion nodes before I run out of storage. Concentrating so much on one
position only makes sense where it can be reused, which is at the start of the game. On 7x7, creating that many nodes takes about a day. I
What was the consensus on 7x7 komi? It was discussed back during Don's
scalability study, but I couldn't find the number itself. Was it 9.0?
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-
Cool idea.
Magnus Persson wrote:
Valkyria computes AMAF win rates for all moves including those that are
pruned or illegal in the position. What I noticed is that in cases of
critical semeais the AMAF values of moves that are for example illegal
can get very high since they only get legal when
Michael Williams wrote:
Ben Shoemaker wrote:
Success! I was able to build on WinXP using Scons and minGW (with
gcc4.3.3). Here's what (finally) worked for me:
1. Install Python 2.6.2
http://www.python.org/ftp/python/2.6.2/python-2.6.2.msi
2. Install minGW (using TDM's installe
Ben Shoemaker wrote:
Success! I was able to build on WinXP using Scons and minGW (with gcc4.3.3).
Here's what (finally) worked for me:
1. Install Python 2.6.2
http://www.python.org/ftp/python/2.6.2/python-2.6.2.msi
2. Install minGW (using TDM's installer on empty minGW directory)
http://down
-go.org] On Behalf Of Michael Williams
Sent: Thursday, May 14, 2009 7:08 AM
To: computer-go@computer-go.org
Subject: Re: [computer-go] Implications of a CPU vs Memory trend on MCTS
C# does. It should only take 30 bytes per node to store the information I
need to have. But somehow that turns into
It's on my list of things to improve.
Michael Williams wrote:
C# does. It should only take 30 bytes per node to store the information
I need to have. But somehow that turns into 50 bytes. Byte alignment
plus class overhead, I guess.
Matthew Woodcraft wrote:
Michael Williams wrot
C# does. It should only take 30 bytes per node to store the information I need to have. But somehow that turns into 50 bytes. Byte alignment plus class
overhead, I guess.
Matthew Woodcraft wrote:
Michael Williams wrote:
I want to correct that last statement. With about 350M nodes
little worse as the tree gets bigger, but
probably not much.
Michael Williams wrote:
Those numbers are the average after the tree has grown to 1B nodes. I'm
sure the cache hates me. Each tree traversal will likely make several
reads from random locations in a 50 GB file.
Don Dailey wro
Plus, they are both made out of the same stuff so one would expect them to
scale at about the same rate.
Mark Boon wrote:
I'm a bit late reading this thread and the discussion seems to have
veered from the original topic a bit.
As to the CPU vs. memory discussion, it's not so clear-cut to me t
more than it did
other moves. I'm not saying keeping the tree is a huge benefit. I'm just saying that I don't think your 99% number is fair.
Dave Dyer wrote:
At 02:13 PM 5/12/2009, Michael Williams wrote:
Where does your 99% figure come from?
1/361 < 1%
by endgame ther
Where does your 99% figure come from?
Dave Dyer wrote:
Storing an opening book for the first 10 moves requires
331477745148242200 nodes. Even with some reduction for symmetry,
I don't see that much memory becoming available anytime soon, and you still
have to evaluate them somehow.
Ac
1897
*From:* Michael Williams
*To:* computer-go
*Sent:* Tuesday, May 12, 2009 1:09:41 PM
*Subject:* Re: [computer-go] Implications of a CPU vs Memory trend on MCTS
That's basically what I'm doing. Except that there is no depth limit
and only the parts of the tree that you
erley Robinson, 1897
--------
*From:* Michael Williams
*To:* computer-go
*Sent:* Tuesday, May 12, 2009 10:18:28 AM
*Subject:* Re: [computer-go] Implications of a CPU vs Memory trend on MCTS
In my system, I can retrieve the children of a
ng if most of your reads and writes are
cached.What happens when your tree gets much bigger than available
memory?
- Don
On Tue, May 12, 2009 at 1:18 PM, Michael Williams
mailto:michaelwilliam...@gmail.com>> wrote:
In my system, I can retrieve the children of any node a
y off.
My guess is that this is impractical, but it's fun to think about how it
might be done. I'm not sure how to do it without having a caching
nightmare.
- Don
On Tue, May 12, 2009 at 12:41 PM, Michael Williams
mailto:michaelwilliam...@gmail.com>> wrote:
ve uurtamo wrote:
is the ssd fast enough to be practical?
s.
On Tue, May 12, 2009 at 12:41 PM, Michael Williams
wrote:
Don Dailey wrote:
On Tue, May 12, 2009 at 12:16 PM, Michael Williams
mailto:michaelwilliam...@gmail.com>> wrote:
I have a trick ;)
I am currently creating MCTS
Don Dailey wrote:
On Tue, May 12, 2009 at 12:16 PM, Michael Williams
mailto:michaelwilliam...@gmail.com>> wrote:
I have a trick ;)
I am currently creating MCTS trees of over a billion nodes on my 4GB
machine.
Ok, I'll bite.What is your solution?
I use an SSD.
I have a trick ;)
I am currently creating MCTS trees of over a billion nodes on my 4GB machine.
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
The Projects link (http://fuego.sourceforge.net/projects.html) on the Fuego
site (http://fuego.sourceforge.net/) is broken.
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
When Mogo runs on the supercomputer with long-ish time limits, how big does the
search tree get?
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
Does anyone know how to interpret the data at each node of a SGF created by the
Fuego command uct_savetree?
It looks like this:
MoveCount 35955
PosCount 35963
Mean 0.50
Rave:
D4 0.50 (31635.70)
B2 0.50 (27049.71)
C2 0.50 (28288.84)
D2 0.50 (29368.09)
E2 0.50 (27690.41)
F2 0.50 (27246.91)
G2 0
The Fuego documentation doesn't seem to contain any information about
command-line options (if there are any).
Ian Osgood wrote:
On May 8, 2009, at 11:57 AM, Ech Naton wrote:
> Hello
> I made a binary of Fuego 0.3.2 for Intel Mac. It is build on OS X 10.4.
> So I don't know if it runs als
After starting Fuego and issuing these commands:
boardsize 4
komi 1.5
book_clear
go_param timelimit 600
uct_param_search number_threads 1
go_param_rules ko_rule pos_superko
uct_param_search max_nodes 3000
reg_genmove_toplay
I get this result:
SgRandom::SetSeed: 1241728335
SgDefaultTimeContr
I found an error in the documentation:
===
void GoUctCommands::CmdParamPlayer ( GtpCommand & cmd )
===
Get and set GoUctPlayer parameters.
This command is compatible with the GoGui analyz
only white stones adjacent. Komi
of board is taken into account.
Michael Williams wrote:
What is the meaning of the value returned by uct_score?
Martin Mueller wrote:
How do you set the number of threads that you want Fuego to use?
E.g. for 4 threads
uct_param_search number_threads 4
What is the meaning of the value returned by uct_score?
Martin Mueller wrote:
How do you set the number of threads that you want Fuego to use?
E.g. for 4 threads
uct_param_search number_threads 4
This can go e.g. in a config file, or you can set it in GoGui in the Uct
Param Search dialog.
I found it right after sending that (of course):
"uct_param_player max_games N"
Michael Williams wrote:
Is there a way to get Fuego to do a certain number of playouts per turn?
"uct_param_search number_playouts N" looked promising, but that is
something different I thin
Is there a way to get Fuego to do a certain number of playouts per turn?
"uct_param_search number_playouts N" looked promising, but that is something
different I think.
Martin Mueller wrote:
How do you set the number of threads that you want Fuego to use?
E.g. for 4 threads
uct_param_sear
Thanks. Yes, it works in Windows.
Martin Mueller wrote:
How do you set the number of threads that you want Fuego to use?
E.g. for 4 threads
uct_param_search number_threads 4
This can go e.g. in a config file, or you can set it in GoGui in the Uct
Param Search dialog.
You could also play
How do you set the number of threads that you want Fuego to use? Does the
windows version support multiple threads?
Michael Williams wrote:
Very nice. I do have one suggestion for Fuego. Make use of the ability
to filter certain root moves out of the UCT tree to remove symmetrically
Very nice. I do have one suggestion for Fuego. Make use of the ability to filter certain root moves out of the UCT tree to remove symmetrically equivalent
moves. Or if it is not cost-prohibitive, throughout the tree.
Martin Mueller wrote:
Our technical report describing the Fuego framework
Michael Williams wrote:
See the April 30 2009 posting: http://www.tobiaspreis.de/
The CUDA SDK also comes with a sample called "Monte-Carlo Option Pricing"
___
computer-go mailing list
computer-go@computer-go.org
http://www.compu
See the April 30 2009 posting: http://www.tobiaspreis.de/
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
I wish I was smart :(
David Silver wrote:
Hi Remi,
I understood this. What I find strange is that using -1/1 should be
equivalent to using 0/1, but your algorithm behaves differently: it
ignores lost games with 0/1, and uses them with -1/1.
Imagine you add a big constant to z. One millio
David Silver wrote:
Hi Michael,
But one thing confuses me: You are using the value from Fuego's 10k
simulations as an approximation of the actual value of the position.
But isn't the actual
value of the position either a win or a loss? On such small boards,
can't you assume that Fuego is ab
I will ignore Magnus's comments about AMAF while I respond directly to your
comments.
If you do one or two simulations from a leaf node and they happen to result in losses, you would never simulate that node again? And never expand it into it's
child nodes? It is very possible that the winnin
My favorite part:
"One natural idea is to use the learned simulation policy in Monte-Carlo search, and
generate new deep search values, in an iterative cycle."
But one thing confuses me: You are using the value from Fuego's 10k simulations as an approximation of the actual value of the position.
1 - 100 of 290 matches
Mail list logo
