Am Samstag, den 20.12.2014, 09:43 +0100 schrieb Stefan Kaitschick:
Great work. Looks like the age of nn is here.
How does this compare in computation time to a heavy MC move
generator?
One very minor quibble, I feel like a nag for even mentioning it: You
write
The most frequently
Hi,
I am still fighting with the NN slang, but why do you zero-padd the
output (page 3: 4 Architecture Training)?
From all I read up to now, most are zero-padding the input to make the
output fit 19x19?!
Thanks for the great work
Detlef
Am Freitag, den 19.12.2014, 23:17 + schrieb Aja
Hi,
as I want to by graphic card for CNN: do I need double precision
performance? I give caffe (http://caffe.berkeleyvision.org/) a try, and
as far as I understood most is done in single precision?!
You get comparable single precision performance NVIDA (as caffe uses
CUDA I look for NVIDA) for
Hi,
I am just trying to reproduce the data from page 7 with all features
disabled. I do not reach the accuracy (I stay below 20%).
Now I wonder about a short statement in the paper, I did not really
understand:
On page 4 top right they state In our experience using the rectifier
function
Am 31.12.2014 um 14:05 schrieb Petr Baudis:
Hi!
On Wed, Dec 31, 2014 at 11:16:57AM +0100, Detlef Schmicker wrote:
I am just trying to reproduce the data from page 7 with all features
disabled. I do not reach the accuracy (I stay below 20%).
Now I wonder about a short statement
Hi,
I am working on a CNN for winrate and territory:
approach:
- input 2 layers for b and w stones
- 1. output: 1 layer territory (0.0 for owned by white, 1.0 for owned
by black (because I missed TANH in the first place I used SIGMOID))
- 2. output: label for -60 to +60 territory leading
architecture did you use? Can you give us some details?
On Sun, Feb 8, 2015 at 5:22 AM, Detlef Schmicker d...@physik.de
mailto:d...@physik.de wrote:
Hi,
I am working on a CNN for winrate and territory:
approach:
- input 2 layers for b and w stones
- 1. output: 1 layer
I also set up a 13x13 client. Seems to work fine, but rating is off I think.
I will let it up for a while, hopefully some anchors coming up:)
Thanks a lot for setting it up again, Detlef
Am 16.01.2015 um 17:21 schrieb Christoph Birk:
On Jan 16, 2015, at 1:51 AM, valky...@phmp.se wrote:
I
Seems a good idea to me. It is a quasi standard in publishing, so why
not set Gnugo-3.7.10 at level 10 to 1800ELO on every board size?!
Am 16.01.2015 um 23:17 schrieb Christoph Birk:
On 01/16/2015 12:03 PM, David Doshay wrote:
cgos.boardspace.net http://cgos.boardspace.net says:
At the
You are right, I too often read 3.7 in the past, but actually the papers
using 3.8 now:)
Am 17.01.2015 um 12:08 schrieb Urban Hafner:
On Sat, Jan 17, 2015 at 10:38 AM, Detlef Schmicker d...@physik.de
mailto:d...@physik.de wrote:
Seems a good idea to me. It is a quasi standard
schrieb Detlef Schmicker:
You are right, I too often read 3.7 in the past, but actually the
papers using 3.8 now:)
Am 17.01.2015 um 12:08 schrieb Urban Hafner:
On Sat, Jan 17, 2015 at 10:38 AM, Detlef Schmicker d...@physik.de
mailto:d...@physik.de wrote:
Seems a good idea to me
What driver is loaded before suspend?
My guess: your distro does not reload the corresponing kernel module
after suspend...
On Ubuntu the driver does not seem to be loaded as module, therefore I
can not check...
if you know what module, check lsmod to see if it is loaded
after suspend I
Hi,
I am planing to play around a little with CNN for learning who is
leading in a board position.
What would you suggest to represent the komi?
I would try an additional layer with every point having the value of komi.
Any better suggestions:)
By the way:
Todays bot tournament nicego19n
Am 11.01.2015 um 22:41 schrieb Aja Huang:
2015-01-11 15:59 GMT+00:00 Detlef Schmicker d...@physik.de
mailto:d...@physik.de:
By the way:
Todays bot tournament nicego19n (oakfoam) played with a CNN for
move prediction.
It was mixed into the original gamma with some quickly
Thanks,
I will, but it will take some time.
It is a problem of resources:
As I have holidays in the summer I think of visiting Advances in
Computer Games 2015 in Leiden.
And I do not want to go there with nothing!
So I need my graphic card for CNN training as I think of doing some
research
Hi Hiroshi,
I try to layout my approach:
- expandLeaf: expands a leaf node, after being visited 10 times (a
parameter, but increasing it usually did not harm playing strength
significantly)
it contains a
if (!expandmutex.try_lock())
return false;
which was in the code anyway to avoid,
This is great, especially as the next KGS tournament is 13x13 :)
Is there a way to get the CGOS archives back online? Or does anybody
have a copy which he can offer?
Would be really great
Thanks Detlef
Am 05.03.2015 um 08:23 schrieb valky...@phmp.se:
At least the 13x13 server is working
now.
Same with 9x9 and 19x19, sadly.
-Josh
On Fri, Mar 6, 2015 at 1:47 PM, Detlef Schmicker d...@physik.de wrote:
This is great, especially as the next KGS tournament is 13x13 :)
Is there a way to get the CGOS archives back online? Or does anybody have a
copy which he can offer?
Would be really
Hi,
I wonder which ideas are around for liberty races in playouts.
What nicego does: it reweights the random moves in the playout to make
sure, that each point is played with roughly the same probability.
This approach tries to solve the problem, that local playout rules
modify this
above
50% prediction rate), therefore I am using the net without last move
feature.
But it may be, that last move features come from our original gammas
anyway, which are mixed with CNN values...
Am 28.04.2015 um 18:26 schrieb Aja Huang:
On Mon, Apr 27, 2015 at 2:12 PM, Detlef Schmicker d
oakfoam has BOARDSIZE_MAX set to 25, but it seems it is only used to say
unsupported board size at the moment :)
I think the reason was gtp, but it was set long before I joined the
project
I dont see a reason, why there should be any problems using it with DNN
on 19x19 trained network.
];
if (result[i]0.1) result[i]=0.1;
}
delete[] data;
delete b;
}
Am 27.04.2015 um 13:44 schrieb Petr Baudis:
On Mon, Apr 27, 2015 at 12:35:05PM +0200, Detlef Schmicker wrote:
I dont see a reason, why there should be any problems using it with
DNN on 19x19 trained network. If a 25x25
Hi,
I set up a CGOS server at home. It is connected via dyndns, which is not
optimal of cause :(
physik.selfhost.eu
Ports:
8080 (webinterface)
8083 (19x19, GnuGo 3.8 set to ELO 1800 as anachor)
This is mainly for testing, if I get CGOS up correctly, what to do to
have it permanently
$gid $over
return
}
This definitly looks like a jigo is not possible. I am afraid, I will
probably not go into this. I still hope for a future CGOS replacement :)
Detlef
Rémi
- Mail original -
De: Detlef Schmicker d...@physik.de
À: computer-go@computer-go.org
}
This definitly looks like a jigo is not possible. I am afraid, I will
probably not go into this. I still hope for a future CGOS replacement :)
Detlef
Rémi
- Mail original -
De: Detlef Schmicker d...@physik.de
À: computer-go@computer-go.org
Envoyé: Samedi 2 Mai 2015 14:21:05
8084 for 25x25 with GnuGo 3.8 as ELO 1800 anachor is up for a while now :)
9x9 too, (13x13 I will not set up, original cgos is running fine on 13x13)
Detlef
Am 02.05.2015 um 07:21 schrieb Detlef Schmicker:
Hi,
I set up a CGOS server at home. It is connected via dyndns, which is
not optimal
big reason I've had trouble getting 9x9 and 19x19
back up.
I'll have to check the crontab but it should show a tcl script that
fires off a sqlite3 dump and feeds bayeselo.
Did you make any changes to the git source to get it to run ok?
-Josh
On Sat, May 2, 2015 at 3:28 AM, Detlef Schmicker d
to repackage the tclkit server
by copying a local built sqlite.so. file
Lots of kinks and issues, but to be honest you are doing a better job
at me at understanding/running it.
So I can at least offer dedicated space.
On Sat, May 2, 2015 at 12:18 PM, Detlef Schmicker d...@physik.de wrote:
8084
of lightsql. Then I can use php
for the backend webwork.
I'm going to re-use some of the webcode I wrote years ago for OICS as
well so that will kickstart that portion a bit.
Will update more as things progress.
-Josh
On Mon, Apr 6, 2015 at 6:13 AM, Detlef Schmicker d...@physik.de wrote:
What about
What about just start the project on github or https://bitbucket.org/
(is not bad at forking and merging)
Open an issue for the discussion and off we go:)
When I was thinking of a quick solution I was thinking about gogui,
which supports most of the game handling already.
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Oakfoam uses caffe library.I did not ask, as I considered it the same
as using e.g. boost lihrary to not write special kind of maps, you do
not want to write your self.
Of cause the net definition and training is our own. Most of the code
would be
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Hi,
I'd like to add the bayes rating to 9x9 and 19x19 intermediate server
(physik.selfhost.eu:8080) and wonder, if the bayeselo scripts for go
are around some where?
I did not find them in the original cgos source code:(
Detlef
Am 23.05.2015 um
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Ubunut 12.04 (64 bit) self compiled with
gcc -v
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/usr/lib/gcc/x86_64-linux-gnu/4.8/lto-wrapper
Target: x86_64-linux-gnu
Configured with: ../src/configure -v --with-pkgversion='Ubuntu
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
After my ISP crashed, I do not get up 9x9 at the moment.
Immediatly myCtest tries to connect from within the middle of a game i
think and DODs the server
Am 26.05.2015 um 18:56 schrieb Christoph Birk:
On 05/26/2015 02:41 AM, Detlef Schmicker
are right. If you have any trouble with this, please
use attached script which will send time_settings TIME 0 0
upon new game.
Best Hideki
Detlef Schmicker: 55910719.3050...@physik.de:
Is it correct, that only time_left is deliverd to the gtp engine?
No time setting before? (therefore my
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
http://blog.physik.de/?page_id=788
should be up again, sorry when I took my comuter to the olympiad in
Leiden (which was a great event) it changed the dyndns ip to leiden :(
Detlef
Am 03.07.2015 um 20:33 schrieb Olivier Teytaud:
Hello; we would
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
OK,
I moved it to an old Nokia N800 tablet (I was bored, and did not have
an idea to improve my bot:)
The anchor is not running on the tablet, since I was afraid of it
loosing on time, if a opponent plays a lot of moves (gnugo on 9x9
takes about 10s
Birk:
On May 22, 2015, at 10:46 AM, Detlef Schmicker d...@physik.de
wrote:
I wonder, if it would help to put it up once a week or so, with
announcement, and take it down again, if the number of bots falls
below 5 or so?
I am not actively developing a bot, but IMHO without being up 24/7
CGOS
this would bring more bots connect at the same time?!
I do not have the resource (and do not want to spend the energy costs)
to keep a number of strong open source bots up, so bots from 900 to
2700ELO could use cgos with some sense.
Detlef
Am 02.05.2015 um 18:18 schrieb Detlef Schmicker:
8084
/
Hideki
Detlef Schmicker: 558d5159.1000...@physik.de: Is there a client
to watch the games on NNGS go servers?
Thanks a lot
Detlef
___ Computer-go
mailing list Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Is there a client to watch the games on NNGS go servers?
Thanks a lot
Detlef
-BEGIN PGP SIGNATURE-
Version: GnuPG v1
iQIcBAEBAgAGBQJVjVFZAAoJEInWdHg+Znf4f3gP/0FwX4fu5g2AXJ6KO1fIMgQS
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Thanks a lot,
mine did not run stable, hope you have more luck! I think I had to
little RAM (128MB)
Detlef
Am 10.11.2015 um 14:11 schrieb Hiroshi Yamashita:
> Hi,
>
> I have started CGOS on my VPS(Virtual Private Server). 19x19 and
> 9x9 are
illegal patterns
> (surrounded middle stone). So I'd hint it's close to 2.
>
> On 03/11/2015 18:17, Detlef Schmicker wrote: I could not find the
> number of 3x3 patterns in Go, if used all symmetrie s.
>
> Can anybody give me a hint, were to find. Harvestin
; <alvaro.be...@gmail.com> wrote:
>
>> I get 1107 (954 in the middle + 135 on the edge + 18 on a
>> corner).
>>
>> Álvaro.
>>
>>
>>
>> On Tue, Nov 3, 2015 at 2:00 PM, Detlef Schmicker <d...@physik.de>
>> wrote:
>>
> Tha
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
I could not find the number of 3x3 patterns in Go, if used all symmetrie
s.
Can anybody give me a hint, were to find. Harvesting 4 games I get
1093:)
Thanks, Detlef
-BEGIN PGP SIGNATURE-
Version: GnuPG v2.0.22 (GNU/Linux)
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Hi,
I would like to train features like in
http://www.remi-coulom.fr/Amsterdam2007/
but using DCNN probabilities as an additional not trained gamma, which
is always present. Did anybody try using an additional not trained
gamma (not necessarily
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
If I compare hardware specs in the KGS tournaments I usually use
http://spec.org/cpu2006/results/rint2006.html
(Multithread Integer operations are the ones most important for
computer go programs I think)
Detlef
Am 08.10.2015 um 05:48 schrieb
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Hi,
I have a probability table of all possible moves. What is the fastest
way to pick with probability, possibly with reducing the quality of
probability?!
I could not find any discussion on this on computer-go, but probably I
missed it :(
Thansk a
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Hi,
all the guys which are using caffe for DCNN:
If you want to see how bad a bot is playing in CPU mode: have a look
at the last results of NiceGo :)
Obviously the caffe library changed between December 2014 and August
2015 an now every thread
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Thanks for the very detailed report! SO good to see, that stronger
programs start using DCNN.
We should ask Nick, if he DCNN gets an exception from the KGS rules.
At the moment I would interpret them as not allowing multiple bots
using the same CNN,
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Hi,
as somebody ask I will offer my actual CNN for testing.
It has 54% prediction on KGS 6d+ data (which I thought would be state
of the art when I started training, but it is not anymore:).
it has:
1
2
3
> 4 libs playing color
1
2
3
> 4 libs
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Am 06.12.2015 um 16:24 schrieb Petr Baudis:
> On Sat, Dec 05, 2015 at 02:47:50PM +0100, Detlef Schmicker wrote:
>> I understand the idea, that long term prediction might lead to a
>> different optimum (but it should not lead to on
omparison is difficult.
>
> Cumulative accuracy Detlef44%
> http://computer-go.org/pipermail/computer-go/2015-October/008031.html
>
> Regards, Hiroshi Yamashita
>
>
> - Original Message - From: "Detlef Schmicker"
> <d...@physik.de> To: <comput
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Yes, the published
http://computer-go.org/pipermail/computer-go/2015-December/008324.html
I think, you can not win this with "normal" good moves :)
You have to exploit mfgo
Detlef
Am 29.12.2015 um 15:18 schrieb "Ingo Althöfer":
> Hi Detlef,
>
>>
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
I gave pure DCNN 54% a try against the 15 kyu version:)
http://files.gokgs.com/games/2015/12/29/mfgo15kyu0-NiceGo19N.sgf
There was no pass handling, therefore filled an eye, without would
have been 133.5 loss or so :)
Detlef
Am 29.12.2015 um 11:20
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Hay,
I wonder if somebody has the same program (with the same settings)
rated on cgos 19x19 and kgs?
I am still fighting with resigning in the case of value-network and
playouts disagree, so I can not run oakfoam on kgs, but would like to
have a
t;> positions are 445693 from 2156 games. All games are shuffled
>> in advance. Each position is randomly rotated. And memorizing
>> 24000 positions, then shuffle and store to LebelDB. At first I
>> did not shuffle games. Then accuracy is down each 61000 iteration
>> (one epoch
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
> One possibility is that 0=loss, 1=win, and the number they are
quoting is
> sqrt(average((prediction-outcome)^2)).
this makes perfectly sense for figure 2. even playouts seem reasonable.
But figure 2 is not consistent with the numbers in section
2.2016 um 20:10 schrieb Hideki Kato:
> Detlef Schmicker: <56b385ce.4080...@physik.de>: Hi,
>
> I try to reproduce numbers from section 3: training the value
> network
>
> On the test set of kgs games the MSE is 0.37. Is it correct, that
> the results are repre
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Hi,
I try to reproduce numbers from section 3: training the value network
On the test set of kgs games the MSE is 0.37. Is it correct, that the
results are represented as +1 and -1?
This means, that in a typical board position you get a value of
schrieb Álvaro Begué:
> The positions they used are not from high-quality games. They
> actually include one last move that is completely random.
>
> Álvaro.
>
>
> On Thursday, February 4, 2016, Detlef Schmicker <d...@physik.de>
> wrote:
>
> Hi,
>
>
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Hi Ingo,
I think you are not alone: When I started computer go 4 years ago I
ask a good friend of mine, who did his PhD in Neural Networks back in
the 90s, if I have any chance to use them instead of pattern matching
and he said, they will probably
training time is roughly proportional to the number of
> neurons in the net?
>
> Thanks,
>
> David
>
>> -Original Message- From: Computer-go
>> [mailto:computer-go-boun...@computer-go.org] On Behalf Of Detlef
>> Schmicker Sent: Tuesday, February 02, 201
otated. And memorizing
>> 24000 positions, then shuffle and store to LebelDB. At first I
>> did not shuffle games. Then accuracy is down each 61000 iteration
>> (one epoch, 256 mini-batch). http://www.yss-aya.com/20160108.png
>> It means DCNN understands easily the differenc
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Thanks a lot for sharing this.
Quite interesting that you do not reach the prediction rate 57% from
the facebook paper by far too! I have the same experience with the
GoGoD database. My numbers are nearly the same as yours 49% :) my net
is quite
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Thanks,
would be great if we could get the hardware info on darkforest :)
(it is in the tournament rules as I understood, so Facebook should
release...)
Thanks Detlef
Am 10.01.2016 um 16:18 schrieb Nick Wedd:
> Congratulations to Zen19S, winner of
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Interesting, my value net does the same, even it was trained totally
different from 7d+ games :)
Am 13.03.2016 um 09:54 schrieb Darren Cook:
> From Demis Hassabis: When I say 'thought' and 'realisation' I just
> mean the output of #AlphaGo value net.
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
What are you using for loss?
this:
layers {
name: "loss4"
type: EUCLIDEAN_LOSS
loss_weight: 2.0
bottom: "vvv"
bottom: "pool2"
top: "accloss4"
}
?
Am 04.03.2016 um 16:23 schrieb Hiroshi Yamashita:
> Hi,
>
> I tried to make Value
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
You are right, but from fig 2 of the paper can see, that mc and value
network should give similar results:
70% value network should be comparable to 60-65% MC winrate from this
paper, usually expected around move 140 in a "human expert game" (what
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Am 13.03.2016 um 11:28 schrieb Josef Moudrik:
> How well do you think the mcts-weakness we have witnessed today is
> hidden in AG? Or, how can one go about exploiting it
> systematically?
>
> I think it might be well hidden by the value network
; } } } layers { name:
>> "relu10" type: RELU bottom: "conv10" top: "conv10" }
>>
>> layers { name: "conv11_3x3_128" type: CONVOLUTION blobs_lr: 1.
>> blobs_lr: 2. bottom: "conv10" top: "conv11" convolution_param
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
you can use caffe with time on the command line.
It gives you forward and backward time for a batch.
In my tests the batch size was not too important (I think, because the
net is quite large)...
cuDNN helps a lot in training, I did not test
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Hi,
thanks a lot for sharing! I try a slightly different approach at the
moment:
I use a combined policy / value network (adding 3-5 layers with about
16 filters at the end of the policy network for the value network to
avoid overfitting) and I use
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
OK, this thread is quite long, and I am not sure I saw all posts :)
My suggestion, rate the bots on CGOS before the tournament and take
this rating for McMahon or for handicaps. I think we can thrust the
bot authors to take the correct rating and
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
You are right, usually they do quite well, but e.g. liberty races with
large dragons are quite difficult.
And there must be a reason, why the value net was so wrong in the game
alphgo lost:)
Am 21.04.2016 um 13:51 schrieb Erik van der Werf:
> On
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Seriously, this was all on game 4?
Am 30.06.2016 um 07:43 schrieb "Ingo Althöfer":
> Hi,
>
> the organizers taped it on video. I will let you know as soon as I
> learn where it is put online.
>
> The event was: 10 minutes honorings 35 minutes
Hi,
I'd like to start a discussion on what zen might do being so strong on
CGOS with only one core and no graphic card :)
The version actual playing (and therefore best comparable to the
programs actual playing) is
RankNameElo + − Games
12
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Hi,
good to start this discussion here. I had the discussion some times,
and we (discussion partner and me) were not sure, against which test
set the 57% was measured.
If trained and tested with kgs 6d+ dataset, it seems reasonable to
reach 57% (I
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Hi Nick,
you created the game with 4 min each on KGS?
Detlef
Am 01.10.2016 um 16:26 schrieb Nick Wedd:
> The October KGS bot tournament will be on Sunday, October 9th,
> starting at 16:00 UTC and ending by 23:55 UTC. It will use 9x9
> boards, with
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
sorry, in your announcement you wrote five, but everything is good
Am 01.10.2016 um 22:27 schrieb Nick Wedd:
> Hi Detlef,
>
> On 1 October 2016 at 20:18, Detlef Schmicker <d...@physik.de>
> wrote:
>
> Hi Nick,
>
&g
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Hi Hiroshi,
thanks a lot for your info.
You did not try reinforcement learning I think. Do you have any idea,
why this would make the policy network 250ELO stronger, as mentioned
in the alphago paper (80% winrate)?
Are pros playing so bad?
Do you
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Hi Hiroshi,
> Now I'm making 13x13 selfplay games like AlphaGo paper. 1. make a
> position by Policy(SL) probability from initial position. 2. play a
> move uniformly at random from available moves. 3. play left moves
> by Policy(RL) to the end. (2)
You are absolutely right, as I was in understanding RL policy network
mode I thought, everything is about this, sorry
Am 21.11.2016 um 15:22 schrieb Gian-Carlo Pascutto:
> On 20-11-16 11:16, Detlef Schmicker wrote:
>> Hi Hiroshi,
>>
>>> Now I'm making 13x13 selfplay gam
I want to share some experience training my policy cnn:
As I wondered, why reinforcement learning was so helpful. I trained
from the Godod database with only using the moves by the winner of
each game.
Interestingly the prediction rate of this moves was slightly higher
(without training, just
Hi Erik,
as far as I understood it, it was 250ELO in policy network alone ...
section
2Reinforcement Learning of Policy Networks
We evaluated the performance of the RL policy network in game play,
sampling each move (...) from its output probability distribution over
actions. When
Very interesting,
but lets wait some days for getting an idea of the strength,
4d it reached due to games against AyaBotD3, now it is 3d again...
Detlef
Am 10.01.2017 um 15:29 schrieb Gian-Carlo Pascutto:
> On 10-01-17 15:05, Hiroshi Yamashita wrote:
>> Hi,
>>
>> Golois5 is KGS 4d.
>> I think
Hi,
what makes you think the opening theory with reverse komi would be the
same as with standard komi?
I would be afraid to invest an enormous amount of time just to learn,
that you have to open differently in reverse komi games :)
Detlef
Am 05.01.2017 um 10:50 schrieb Paweł Morawiecki:
>
Hi,
this sounds interesting! AlphaGo paper plays only with RL network, if I
understood correctly. If we start this huge approach we should try to
carefully discuss the way (and hopefully get some hints from people
tried with much computational power :)
If I understood correctly you would try to
> * Which of the currently three top bots will show up in the
> European Go Congress in Oberhof in July/August 2017?
just set up one of the top open source bots:
on moderate hardware
ray: http://www.dragongoserver.net/userinfo.php?uid=97868
and if this is too strong for europe
oakfoam:
oakfoam value network does exactly this, we have 6 komi layers -7.5 -5.5
-0.5 0.5 5.5 7.5 (+ and - due to color played) and trained from 4d+ kgs
games with this:
if (c_played==1):
if ("0.5" in komi):
komiplane=1;
if ("6.5" in komi or "2.75" in komi or "5.5" in komi):
#komi 6.5
I looked into this too:
oakfoam would not benefit a lot from more cpu power at the moment, with
4 cores I mix 10 playouts with the value net in the ratio (3:7) at the
moment.
In case of buying a Ryzen: take care the board allows two GTX1080 Ti
(wait till end of march to buy them) and buy a power
e thread hijacking, everyone.
>
>
> On Sat, Mar 4, 2017 at 4:29 AM, Detlef Schmicker <d...@physik.de> wrote:
>
>> I looked into this too:
>>
>> oakfoam would not benefit a lot from more cpu power at the moment, with
>> 4 cores I mix 10 playouts with the
Hi Nick
best info I have is:
http://computer-go.org/pipermail/computer-go/2016-June/009444.html
http://computer-go.org/pipermail/computer-go/2016-February/008638.html
Detlef
Am 18.07.2017 um 18:20 schrieb Nick Wedd:
> Hi Magnus,
>
> Thank you for the information. I don't know how to
Hi,
might be a little impolite, but I wonder about the strength of alphago.
The version playing Ke Jie seems to be about as strong (or stronger) as
Ke Jie is. I have the feeling the playing strength is carefully chosen
not to be too strong.
In the press conference it was told, alphago is
I thought it might be fun to have some games in early stage of learning
from nearly Zero knowledge.
I did not turn off the (relatively weak) playouts and mix them with 30%
into the result from the value network. I started at an initial random
neural net (small one, about 4ms on GTX970) and use a
how quickly their strength
> is improving?
>
> s.
>
> On Nov 6, 2017 4:54 PM, "Detlef Schmicker" <d...@physik.de> wrote:
>
>> I thought it might be fun to have some games in early stage of learning
>> from nearly Zero knowledge.
>>
>>
This is a quite natural approach, I think every go program which needs
to play with different komi does it in one way.
At least oakfoam does :)
Detlef
Am 26.10.2017 um 15:55 schrieb Roel van Engelen:
> @Gian-Carlo Pascutto
>
> Since training uses a ridiculous amount of computing power i
97 matches
Mail list logo
