The evaluation is always at least as deep as leaves of the tree.
Still, you're right that the earlier in the game, the bigger the inherent
uncertainty.
One thing I don't understand: if the network does a thumbs up or down,
instead of answering with a probability,
what is the use of MSE? Why not jus
Opent to intepretation if this method is brute force. I think it i. Uses
huge amounts of CPU power to run simulations and evaluate NN's. Even in
chess it was not just about tree search, it needs evaluationfunction ot
make sense of the search
2016-02-24 6:52 GMT+02:00 muupan :
> Congratulations, p
Congratulations, people at DeepMind! Your paper is very interesting to read.
I have a question about the paper. On policy network training it says
> On the first pass through the training pipeline, the baseline was set to
zero; on the second pass we used the value network vθ(s) as a baseline;
bu
On Wed, Jan 27, 2016 at 1:46 PM, Aja Huang wrote:
> We are very excited to announce that our Go program, AlphaGo, has beaten a
> professional player for the first time. AlphaGo beat the European champion
> Fan Hui by 5 games to 0.
It's interesting to go back nearly a decade and read this 2007 art
-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 3
I just want to see how to get 0.5 for the initial position on the board
with some definition.
One possibility is that 0=loss, 1=win, and the number they are quoting is
sqrt(average((prediction-outcome)^2)).
On Thu, Feb 4, 2016 at 3:40 PM, Hideki Kato wrote:
> I think the error is defined as th
I think the error is defined as the difference between the
output of the value network and the average output of the
simulations done by the policy network (RL) at each position.
Hideki
Michael Markefka:
:
>That sounds like it'd be the MSE as classification error of the eventual
>result.
>
That sounds like it'd be the MSE as classification error of the eventual result.
I'm currently not able to look at the paper, but couldn't you use a
softmax output layer with two nodes and take the probability
distribution as winrate?
On Thu, Feb 4, 2016 at 8:34 PM, Álvaro Begué wrote:
> I am no
I am not sure how exactly they define MSE. If you look at the plot in
figure 2b, the MSE at the very beginning of the game (where you can't
possibly know anything about the result) is 0.50. That suggests it's
something else than your [very sensible] interpretation.
Álvaro.
On Thu, Feb 4, 2016 a
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
>> Since all positions of all games in the dataset are used, winrate
>> should distributes from 0% to 100%, or -1 to 1, not 1. Then, the
>> number 70% could be wrong. MSE is 0.37 just means the average
>> error is about 0.6, I think.
0.6 in the range
Detlef Schmicker: <56b385ce.4080...@physik.de>:
>-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?
Looks cor
I re-read the relevant section and I agree with you. Sorry for adding noise
to the conversation.
Álvaro.
On Thu, Feb 4, 2016 at 12:21 PM, Detlef Schmicker wrote:
> -BEGIN PGP SIGNED MESSAGE-
> Hash: SHA1
>
> Thanks for the response, I do not refer to the finaly used data set:
> in
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Thanks for the response, I do not refer to the finaly used data set:
in the referred chapter they state, they have used their kgs dataset
in a first try (which is in another part of the paper referred to
being a 6d+ data set).
Am 04.02.2016 um 18:11 s
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 wrote:
> -BEGIN PGP SIGNED MESSAGE-
> Hash: SHA1
>
> Hi,
>
> I try to reproduce numbers from section 3: traini
-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
1-s
r Rosenthal" wrote:
>>
>>> ~~
>>> Robert: "Hey, AI, you should provide explanations!"
>>> AI: "Why?"
>>> ~~~~~~~~~~~~~~~~~~~~~~~~~~
>>>
>>> Cheers,
>>> Rainer
&g
~~
>> Robert: "Hey, AI, you should provide explanations!"
>> AI: "Why?"
>> ~~
>>
>> Cheers,
>> Rainer
>>
>>> Date: Mon, 1 Feb 2016 0
Cheers,
Rainer
Date: Mon, 1 Feb 2016 08:15:12 -0600
From: "Jim O'Flaherty" mailto:jim.oflaherty...@gmail.com>>
To: computer-go@computer-go.org
<mailto:computer-go@computer-go.org>
Subject: Re: [Computer-go] Mastering
On 02.02.2016 20:21, Olivier Teytaud wrote:
On the other hand, they have super strong people in the team (at the pro
level, maybe ? if Aja has pro level...)
Ca. 5d amateur in the team is enough, regardless of whether Myongwan Kim
thinks that only 9p can understand. Not so. Kim's above 5d amate
What? You have mixed up things.
http://www.europeangodatabase.eu/EGD/Player_Card.php?&key=17374016
2016-02-02 20:21 GMT+01:00 Olivier Teytaud :
>>> If AlphaGo had lost at least one game, I'd understand how people can have
>>> an upper bound on its level, but with 5-0 (except for Blitz) it's hard
>
> If AlphaGo had lost at least one game, I'd understand how people can have
>> an upper bound on its level, but with 5-0 (except for Blitz) it's hard to
>> have an upper bound on his level. After all, AlphaGo might just have played
>> well enough for crushing Fan Hui, and a weak move while the po
2016-02-01 12:24 GMT+01:00 Olivier Teytaud :
> If AlphaGo had lost at least one game, I'd understand how people can have
> an upper bound on its level, but with 5-0 (except for Blitz) it's hard to
> have an upper bound on his level. After all, AlphaGo might just have played
> well enough for crush
On 02.02.2016 19:07, David Fotland wrote:
consider some of this as the difference between math and engineering. Math
desires rigor.
Engineering desires working solutions. When an engineering solution is being
described,
you shouldn't expect the same level of rigor as in a mathematical proof.
On 02.02.2016 17:29, Jim O'Flaherty wrote:
AI Software Engineers: Robert, please stop asking our AI for explanations.
We don't want to distract it with limited human understanding. And we don't
want the Herculean task of coding up that extremely frail and error prone
bridge.
Currently I do not
, 1 Feb 2016 08:15:12 -0600
>> From: "Jim O'Flaherty"
>> To: computer-go@computer-go.org
>> Subject: Re: [Computer-go] Mastering the Game of Go with Deep Neural
>> Networks and Tree Search
>> Message-ID:
>> <
>> cakx5gkjc7
> Without clarity, progress is delayed. Every professor at university will
> confirm this to you.
>
IMHO, Petr contributed enough to academic research
for not needing a discussion with a professor at university
for learning how to do/clarify research :-)
--
=
On 02.02.2016 11:49, Petr Baudis wrote:
you seem to come off as perhaps a little too
aggressive in your recent few emails...
If I were not aggressively critical about inappropriate ambiguity, it
would continue for further decades. Papers containing mathematical
contents must clarify when some
Hi Robert,
maybe it's just me, but you seem to come off as perhaps a little too
aggressive in your recent few emails...
On Tue, Feb 02, 2016 at 09:35:14AM +0100, Robert Jasiek wrote:
> On 01.02.2016 23:01, Brian Cloutier wrote:> I had to search a lot of papers
> on MCTS which
> > mentioned "t
On 01.02.2016 23:01, Brian Cloutier wrote:> I had to search a lot of
papers on MCTS which
> mentioned "terminal states" before finding one which defined them.
> [...] they defined it as a position where there are no more legal
> moves.
On 01.02.2016 23:15, Brian Sheppard wrote:
You play until n
omputer-go] Mastering the Game of Go with Deep Neural Networks
and Tree Search
If anything, the other great DCNN applications predate the application of these
methods to Go. Deep neural nets (convnets and other types) have been
successfully applied in computer vision, robotics, speech recognitio
If anything, the other great DCNN applications predate the application of
these methods to Go. Deep neural nets (convnets and other types) have been
successfully applied in computer vision, robotics, speech recognition,
machine translation, natural language processing, and hosts of other areas.
The
[mailto:computer-go-boun...@computer-go.org] On Behalf Of
Brian Cloutier
Sent: Monday, February 1, 2016 5:02 PM
To: computer-go
Subject: Re: [Computer-go] Mastering the Game of Go with Deep Neural Networks
and Tree Search
> One thing that is not explained is how to determine that a game is o
> One thing that is not explained is how to determine that a game is over
You'll find that very little of the literature explicitly covers this. When
I asked this question I had to search a lot of papers on MCTS which
mentioned "terminal states" before finding one which defined them.
Let me see i
Aja,
I read the paper with great interest. [Insert appropriate praises here.]
I am trying to understand the part where you use reinforcement learning to
improve upon the CNN trained by imitating humans. One thing that is not
explained is how to determine that a game is over, particularly when a
p
Hi Hideki,
you put it wonderfully into two lines:
**
**
******
*** Much more economical methods should be de
er-go.org
Subject: Re: [Computer-go] Mastering the Game of Go with Deep Neural
Networks and Tree Search
Message-ID:
Content-Type: text/plain; charset="utf-8"
Robert,
I'm not seeing the ROI in attempting to map human id
Ingo Althofer:
:
>Hi Hideki,
>
>first of all congrats to the nice performance of Zen over the weekend!
>
>> Ingo and all,
>> Why you care AlphaGo and DCNN so much?
>
>I can speak only for myself. DCNNs may be not only applied to
>achieve better playing strength. One may use them to create
>play
The next type of event could be a new 'Pair Go'
Where a human and a program make up a pair, like Mark Zuckerberg and his
facebook
program against a Google VP and alphaGo. :-)
Thomas
On Mon, 1 Feb 2016, John Tromp wrote:
For those of you who missed it, chess grandmaster Hikaru Nakamura,
rated
For those of you who missed it, chess grandmaster Hikaru Nakamura,
rated 2787, recently played a match against the world's top chess program
Komodo, rated 3368. Each of the 4 games used a different kind of handicap:
Pawn and Move Odds
Pawn Odds
Exchange Odds
4-Move Odds
As you can see, handicaps
Hi!
On Mon, Feb 01, 2016 at 01:38:28PM +, Aja Huang wrote:
> On Mon, Feb 1, 2016 at 11:38 AM, Petr Baudis wrote:
> >
> > That's right, but unless I've overlooked something, I didn't see Fan Hui
> > create any complicated fight, there wasn't any semeai or complex
> > life&death (besides the
Robert,
I'm not seeing the ROI in attempting to map human idiosyncratic linguistic
systems to/into a Go engine. Which language would be the one to use;
English, Chinese, Japanese, etc? As abstraction goes deeper, the nuance of
each human language diverges from the others (due to the way the human
On 01.02.2016 15:15, Jim O'Flaherty wrote:
I'm not seeing the ROI in attempting to map human idiosyncratic linguistic
systems to/into a Go engine. Which language would be the one to use;
English, Chinese, Japanese, etc? As abstraction goes deeper, the nuance of
each human language diverges from t
On 01.02.2016 14:38, Aja Huang wrote:
AlphaGo may do much better in tactical
situations than Crazy Stone and Zen.
Judging very quickly from the Fan Hui games, AlphaGo's group-local
"reading" is very deep and accurate but I'd need to read for myself
equally deeply and carefully before I would
Hi Hideki,
first of all congrats to the nice performance of Zen over the weekend!
> Ingo and all,
> Why you care AlphaGo and DCNN so much?
I can speak only for myself. DCNNs may be not only applied to
achieve better playing strength. One may use them to create
playing styles, or bots for go va
Olivier Teytaud:
:
>Ok, it's not blitz according to http://senseis.xmp.net/?BlitzGames
>(limit at 10s/move for Blitz). But really shorter time settings.
>
>I've seen (as you all) many posts guessing that AlphaGo will lose, but I
>find
>that hard to know. If Fan Hui had won one game, I would say t
Ok, it's not blitz according to http://senseis.xmp.net/?BlitzGames
(limit at 10s/move for Blitz). But really shorter time settings.
I've seen (as you all) many posts guessing that AlphaGo will lose, but I
find
that hard to know. If Fan Hui had won one game, I would say that AlphaGo is
not ready fo
Olivier Teytaud:
:
>If AlphaGo had lost at least one game, I'd understand how people can have
>an upper bound on its level, but with 5-0 (except for Blitz) it's hard to
No, the other five are not blitz games. Quoting from the
paper (pp. 28):
Time controls for formal games were 1 hour main tim
On Mon, Feb 01, 2016 at 12:24:21PM +0100, Olivier Teytaud wrote:
> If AlphaGo had lost at least one game, I'd understand how people can have
> an upper bound on its level, but with 5-0 (except for Blitz) it's hard to
> have an upper bound on his level. After all, AlphaGo might just have played
> we
If AlphaGo had lost at least one game, I'd understand how people can have
an upper bound on its level, but with 5-0 (except for Blitz) it's hard to
have an upper bound on his level. After all, AlphaGo might just have played
well enough for crushing Fan Hui, and a weak move while the position is
sti
Hi!
On Mon, Feb 01, 2016 at 09:19:56AM +, Darren Cook wrote:
> > someone cracked Go right before that started. Then I'd have plenty of
> > time to pick a new research topic." It looks like AlphaGo has
> > provided.
>
> It seems [1] the smart money might be on Lee Sedol:
>
> 1. Ke Jie (wor
On Mon, Feb 1, 2016 at 10:19 AM, Darren Cook wrote:
> It seems [1] the smart money might be on Lee Sedol:
In the DeepMind press conferences (
https://www.youtube.com/watch?v=yR017hmUSC4 -
https://www.youtube.com/watch?v=_r3yF4lV0wk ) Demis Hassabis stated,
that he was quietly confident.
I assume
> someone cracked Go right before that started. Then I'd have plenty of
> time to pick a new research topic." It looks like AlphaGo has
> provided.
It seems [1] the smart money might be on Lee Sedol:
1. Ke Jie (world champ) – limited strength…but still amazing… Less than
5% chance against Lee Se
On 01.02.2016 07:30, Petri Pitkanen wrote:
Explaining why the move is good in human terms is useless goal. Good chess
programs cannot do it nor it is meaningful. As the humans and computers
have vastly different approach to selecting a move then by the definition
have reasons for moves. As an ex
Explaining why the move is good in human terms is useless goal. Good chess
programs cannot do it nor it is meaningful. As the humans and computers
have vastly different approach to selecting a move then by the definition
have reasons for moves. As an example your second item 'long-term aji', For
h
Robert Jasiek jas...@snafu.de:
>On 31.01.2016 20:28, Peter Drake wrote:
>> pick a new research topic.
>
>[a bunch of topics]
I have another topic suggestion. Deep learning needs tons of data. Humans reach
top performance after seeing far, far fewer examples than AlphaGo sees.
Whatever method hum
On 31.01.2016 20:28, Peter Drake wrote:
pick a new research topic.
- explain by the program to human players why MC / DNN play is good in
terms of human understanding of the game
- incorporate the difficult parts, such as long-term aji
- solve the game: prove the correct score, prove a weak s
Hi Peter,
> I'm due for a sabbatical next year. I had been joking, "It sure would be good
> timing if someone cracked Go right before that started. Then I'd have plenty
> of time to pick a new research topic." It looks like AlphaGo has provided.
you are not the only one in such a situation
Let me add my congratulations to the chorus. Well done!
I'm due for a sabbatical next year. I had been joking, "It sure would be
good timing if someone cracked Go right before that started. Then I'd have
plenty of time to pick a new research topic." It looks like AlphaGo has
provided.
On Wed, Jan
Ingo and all,
Why you care AlphaGo and DCNN so much? Surely DeepMind team did
a big leap but the big problems, such as detecting double-ko and
solving complex positions are left unchanged. Also it's well
known that to attack these weakpoint of MCTS bots, the
opponents have to be strong enoug
ic was the worst case so far...
> >
> > Greetings from the bottom, Ingo.
> >
> >
> >
> > Gesendet: Donnerstag, 28. Januar 2016 um 16:41 Uhr Von: "Lucas,
> > Simon M" An: "computer-go@computer-go.org"
> > Betreff: Re: [Computer-go] Mast
from the bottom, Ingo.
>
>
>
> Gesendet: Donnerstag, 28. Januar 2016 um 16:41 Uhr Von: "Lucas,
> Simon M" An: "computer-go@computer-go.org"
> Betreff: Re: [Computer-go] Mastering
> the Game of Go with Deep Neural Networks and Tree Search
>
> Indeed
erstag, 28. Januar 2016 um 16:41 Uhr
Von: "Lucas, Simon M"
An: "computer-go@computer-go.org"
Betreff: Re: [Computer-go] Mastering the Game of Go with Deep Neural Networks
and Tree Search
Indeed – Congratulations to Google DeepMind!
It’s truly an immense achievement. I’m st
On 1/27/16 12:08 PM, Aja Huang wrote:
2016-01-27 18:46 GMT+00:00 Aja Huang mailto:ajahu...@google.com>>:
Hi all,
We are very excited to announce that our Go program, AlphaGo, has
beaten a professional player for the first time. AlphaGo beat the
European champion Fan Hui by 5 ga
On Thu, Jan 28, 2016 at 3:14 PM, Stefan Kaitschick
wrote:
> That "value network" is just amazing to me.
> It does what computer go failed at for over 20 years, and what MCTS was
> designed to sidestep.
Thought it worth a mention: Detlef posted about trying to train a CNN
on win rate as well in F
I always thought the same. But I don't think they tackled the decomposition
problem directly.
Achieving good(non-terminal) board evaluations must have reduced the
problem.
If you don't do full playouts, you get much less thrashing between
independent problems.
It also implies a useful static L&D ev
From: Computer-go [mailto:computer-go-boun...@computer-go.org] On Behalf Of
Olivier Teytaud
Sent: 27 January 2016 20:27
To: computer-go
Subject: Re: [Computer-go] Mastering the Game of Go with Deep Neural Networks
and Tree Search
Congratulations people at DeepMind :-)
I like the fact that
I think such analysis might not bee too usefull. At least chess players
think it is not very usefull. Usually for learning you need "wake-up" your
brains so computer analysis without reasons probabaly on marginally useful.
But very entertaining
2016-01-28 13:27 GMT+02:00 Michael Markefka :
> I t
I think many amateurs would already benefit from a simple blunder
check and a short list of viable alternatives and short continuations
for every move.
If I could leave my PC running over night for a 30s/move analysis at
9d level and then walk through my game with that quality of analysis,
I'd be
Congratulations!
What I find most impressive is the engineering effort, combining so many
different parts, which even standalone would be a strong program.
I think the design philosophy of using 3 different sources of "go
playing" strength is great in it self (and if you read the paper there
: Computer-go [mailto:computer-go-boun...@computer-go.org] On Behalf Of
Jason Li
Sent: Wednesday, January 27, 2016 3:14 PM
To: computer-go@computer-go.org
Subject: Re: [Computer-go] Mastering the Game of Go with Deep Neural Networks
and Tree Search
Congratulations to Aja!
A question to the
Congratulations to the researchers!
On 27.01.2016 21:10, Michael Markefka wrote:
I really do hope that this also turns into a good analysis and
teaching tool for human player. That would be a fantastic benefit from
this advancement in computer Go.
The programs successful as computer players mo
Really nice result! Congratulations to the team.
Now off to study the paper instead of the blogs ...
René
On Wed, Jan 27, 2016 at 3:13 PM, Jason Li wrote:
> Congratulations to Aja!
>
> A question to the community. Is anyone going to replicate the experimental
> results?
>
>
> https://www.quora
Congratulations to Aja!
A question to the community. Is anyone going to replicate the experimental
results?
https://www.quora.com/Is-anyone-replicating-the-experimental-results-of-the-human-level-Go-player-published-by-Google-Deepmind-in-Nature-in-January-2016
?
Jason
On Thu, Jan 28, 2016 at 9:
Congratulations people at DeepMind :-)
I like the fact that alphaGo uses many forms of learning (as humans do!):
- imitation learning (on expert games, learning an actor policy);
- learning by playing (self play, policy gradient), incidentally generating
games;
- use of those games for teaching a
Wow, excellent results, congratulations Aja & team!
I'm surprised to see nothing explicitly on decomposing into subgames (e.g.
for semeai). I always thought some kind of adaptive decomposition would be
needed to reach pro-strength... I guess you must have looked into this;
does this mean that the
Congratulations Aja.
Do you have a plan to run AlphaGo on KGS?
It must be a 9d!
Yamato
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go
Sorry for a typo. I meant
> Hello Aja,
>
> congratulations to the success of you and the other team memberS!
So, not singular, but plural.
Ingo.
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer
2016-01-27 18:46 GMT+00:00 Aja Huang :
> Hi all,
>
> We are very excited to announce that our Go program, AlphaGo, has beaten a
> professional player for the first time. AlphaGo beat the European champion
> Fan Hui by 5 games to 0. We hope you enjoy our paper, published in Nature
> today. The pape
Congratulations Aja and David!
What an interesting idea to train the value network and surprising power
of the cloud!
Then, when you will get +400 Elo? :)
Hideki
Aja Huang:
:
>Hi all,
>
>We are very excited to announce that our Go program, AlphaGo, has beaten a
>professional player for the f
puter-go.org
Betreff: [Computer-go] Mastering the Game of Go with Deep Neural Networks and
Tree Search
Hi all,
We are very excited to announce that our Go program, AlphaGo, has beaten a
professional player for the first time. AlphaGo beat the European champion Fan
Hui by 5 games to 0. We hope
I really do hope that this also turns into a good analysis and
teaching tool for human player. That would be a fantastic benefit from
this advancement in computer Go.
On Wed, Jan 27, 2016 at 9:08 PM, Aja Huang wrote:
> 2016-01-27 18:46 GMT+00:00 Aja Huang :
>>
>> Hi all,
>>
>> We are very excited
Congratulations Aja, well done :)
On Wed, Jan 27, 2016 at 6:46 PM, Aja Huang wrote:
> Hi all,
>
> We are very excited to announce that our Go program, AlphaGo, has beaten a
> professional player for the first time. AlphaGo beat the European champion
> Fan Hui by 5 games to 0. We hope you enjoy o
82 matches
Mail list logo
