Re: [Computer-go] Zero is weaker than Master!?

[Hideki Kato]

The 40 block version (2nd instance) first appeared in 
Section 4 in the paper.  Section 2 and 3 are all for the 1st 
instance.

Hideki

Xavier Combelle: <39a79a0e-7c7d-2a01-a2ae-573cda8b1...@gmail.com>:
>Unless I mistake figure 3 shows the plot of supervised learning to

>reinforcement learning, not 20 bloc/40 block

>

>For searching mention of the 20 blocks I search for 20 in the whole

>paper and did not found any other mention

>

>than of the kifu thing.

>

>

>Le 26/10/2017 à 15:10, Gian-Carlo Pascutto a écrit :

>> On 26-10-17 10:55, Xavier Combelle wrote:

>>> It is just wild guesses  based on reasonable arguments but without

>>> evidence.

>> David Silver said they used 40 layers for AlphaGo Master. That's more

>> evidence than there is for the opposite argument that you are trying to

>> make. The paper certainly doesn't talk about a "small" and a "big" 
>Master.

>>

>> You seem to be arguing from a bunch of misreadings and

>> misunderstandings. For example, Figure 3 in the paper shows the Elo plot

>> for the 20 block/40 layer version, and it compares to Alpha Go Lee, not

>> Alpha Go Master. The Alpha Go Master line would be above the flattening

>> part of the 20 block/40 layer AlphaGo Zero. I guess you missed this when

>> you say that they "only mention it to compare on kifu prediction"?

>>

>

>___

>Computer-go mailing list

>Computer-go@computer-go.org

>http://computer-go.org/mailman/listinfo/computer-go
-- 
Hideki Kato 
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] November KGS bot tournament

[Hiroshi Yamashita]

Hi Nick,


this will be the last of the series of KGS bot tournaments.


Thank you for holding KGS tournament since 2005.
On CGOS, there are always some new comers.
I hope they also enter KGS bot tournament.

Thanks,
Hiroshi Yamashita


- Original Message - 
From: "Nick Wedd" 

To: 
Sent: Thursday, October 26, 2017 4:43 PM
Subject: [Computer-go] November KGS bot tournament


The November KGS bot tournament will be on Sunday, November 5th, starting
at 16:00 UTC and ending by 22:00 UTC.  It will use 19x19 boards, with
time limits
of 14 minutes each and very fast Canadian overtime, and komi of 7½.  It
will be a Swiss tournament.  See http://www.gokgs.com/tournInfo.jsp?id=112
7

Please register by emailing me at mapr...@gmail.com, with the words "KGS
Tournament Registration" in the email title.
With the falling interest in these events since the advent of AlphaGo, it
is likely that this will be the last of the series of KGS bot tournaments.

Nick
--
Nick Wedd  mapr...@gmail.com

___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?

[Jim O'Flaherty]

It's related to this line of thinking by Douglas Hoffstadter:
https://en.wikipedia.org/wiki/Copycat_(software)


Namaste,

Jim O'Flaherty
Founder/CEO
Precision Location Intelligence, Inc.
 • Irving, TX, USA
469-358-0633 <4693580633> • jim.oflaherty...@gmail.com •
www.linkedin.com/in/jimoflahertyjr

CONFIDENTIALITY / PROPRIETARY NOTICE:
The information contained in this e-mail, including any attachment(s), is
confidential information for Precision Location Intelligence, Inc.. As
such, it may be privileged and exempt from disclosure under applicable law.
If the reader of this message is not the intended recipient, or if you
received this message in error, then any direct or indirect disclosure,
distribution or copying of this message is strictly prohibited. If you have
received this message in error, please notify Precision Location
Intelligence, Inc. by calling (214) 489-7779 <2144897779> immediately and
by sending a return e-mail; delete this message; and destroy all copies,
including attachments.

On Thu, Oct 26, 2017 at 11:43 AM, Xavier Combelle  wrote:

> what are semantic genetic algorithm ?
>
> to my knowledge genetic algorithm lead to poor result except as a
> metaheuristic in optimisation problem
>
> Le 26/10/2017 à 14:40, Jim O'Flaherty a écrit :
>
> When I get time to spend dozens of hours on computer go again, I plan to
> play in Robert's area with semantic genetic algorithms. I am an Architect
> Software Engineer. Robert's work will allow me better than starting
> entirely from random in much the same way AlphaGo bootstrapped from the
> 100K of professional games. AG0 then leveraged AlphaGo in knowing an
> architecture that was close enough. My intuition is my approach will be
> something similar in it's evolution.
>
> This is the way we're going to "automate" creating provided proofing of
> human cognition styled computer go players to assist humans in a gradient
> ascent learning cycle.
>
> So, Robert, I admire and am encouraged by your research for my own
> computer go projects in this area. Keep kicking butt in your unique way. We
> are in an interesting transition in this community. Stick it out. It will
> be worth it long term.
>
> On Oct 26, 2017 4:38 AM, "Petri Pitkanen" 
> wrote:
>
>> Unfortunately there is no proof that you principles work better than
>> those form eighties. Nor there is any agreement that your pronciples form
>> any improvement over the old ones. Yes you are a  far better player than me
>> and shows that you are
>> - way better at reading
>> - have hugely better go understanding, principles if you like
>>
>> What is missing that I doubt that you can verbalise your go understanding
>> to degree that by applying those principles  I could become substantially
>> better player. again bulleting
>> - My reading skills would not get any better hence making much of value
>> any learning moot. Obviously issue on me not on your principles
>> - your principles are more complex than you understand. Much of you know
>> is automated to degree that it is subconsciousness information.
>> Transferring that information if hard. Usually done by re-playing master
>> games looking at problems i.e. training the darn neural net in the head
>>
>> If you can build Go bot about  KGS 3/4dan strength I am more than willing
>> to admit you are right and would even consider buying your  books.
>>
>> Petri
>>
>> 2017-10-26 6:21 GMT+03:00 Robert Jasiek :
>>
>>> On 25.10.2017 18:17, Xavier Combelle wrote:
>>>
 exact go theory is full of hole.

>>>
>>> WRT describing the whole game, yes, this is the current state. Solving
>>> go in a mathematical sense is a project for centuries.
>>>
>>> Actually, to my knowledge human can't apply only the exact go theory and
 play a decent game.

>>>
>>> Only for certain positions of a) late endgame, b) semeais, c) ko.
>>>
>>> If human can't do that, how it will teach a computer to do it magically ?

>>>
>>> IIRC, Martin Müller implemented CGT endgames a la Mathematical Go
>>> Endgames.
>>>
>>> The reason why (b) had became unpopular is because there is no go theory
 precise enough to implement it as an algorithm

>>>
>>> There is quite some theory of the 95% principle kind which might be
>>> implemented as approximation. E.g. "Usually, defend your weak important
>>> group." can be approximated by approximating "group", "important" (its loss
>>> is too large in a quick positional judgement), "weak" (can be killed in two
>>> successive moves), "defend" (after the move, cannot be killed in two
>>> successive moves), "usually" (always, unless there are several such groups
>>> and some must be chosen, say, randomly; the approximation being that the
>>> alternative strategy of large scale exchange is discarded).
>>>
>>> Besides, one must prioritise principles to solve conflicting principles
>>> by a higher order principle.

Re: [Computer-go] Source code (Was: Reducing network size? (Was: AlphaGo Zero))

[Shawn Ligocki]

On Thu, Oct 26, 2017 at 2:02 PM, Gian-Carlo Pascutto  wrote:

> On 26-10-17 15:55, Roel van Engelen wrote:
> > @Gian-Carlo Pascutto
> >
> > Since training uses a ridiculous amount of computing power i wonder
> > if it would be useful to make certain changes for future research,
> > like training the value head with multiple komi values
> > 
>
> Given that the game data will be available, it will be trivial for
> anyone to train a different network architecture on the result and see
> if they get better results, or a program that handles multiple komi
> values, etc.
>
> The problem is getting the *data*, not the training.
>

But the data should be different for different komi values, right?
Iteratively producing self-play games and training with the goal of
optimizing for komi 7 should converge to a different optimal player than
optimizing for komi 5. But maybe having high quality data for komi 7 will
still save a lot of the work for training a komi 5 (or komi agnostic)
network?
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Zero is weaker than Master!?

[Gian-Carlo Pascutto]

Figure 6 has the same graph as Figure 3 but for 40 blocks. You can compare
the Elo.

On Thu, Oct 26, 2017, 23:35 Xavier Combelle 
wrote:

> Unless I mistake figure 3 shows the plot of supervised learning to
> reinforcement learning, not 20 bloc/40 block
>
> For searching mention of the 20 blocks I search for 20 in the whole
> paper and did not found any other mention
>
> than of the kifu thing.
>
>
> Le 26/10/2017 à 15:10, Gian-Carlo Pascutto a écrit :
> > On 26-10-17 10:55, Xavier Combelle wrote:
> >> It is just wild guesses  based on reasonable arguments but without
> >> evidence.
> > David Silver said they used 40 layers for AlphaGo Master. That's more
> > evidence than there is for the opposite argument that you are trying to
> > make. The paper certainly doesn't talk about a "small" and a "big"
> Master.
> >
> > You seem to be arguing from a bunch of misreadings and
> > misunderstandings. For example, Figure 3 in the paper shows the Elo plot
> > for the 20 block/40 layer version, and it compares to Alpha Go Lee, not
> > Alpha Go Master. The Alpha Go Master line would be above the flattening
> > part of the 20 block/40 layer AlphaGo Zero. I guess you missed this when
> > you say that they "only mention it to compare on kifu prediction"?
> >
>
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go

-- 

GCP
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

[Computer-go] Amazon EC2 P3 instances with 8xV100!

[Rémi Coulom]

https://aws.amazon.com/about-aws/whats-new/2017/10/introducing-amazon-ec2-p3-instances/

1xGPU: p3.2xlarge: 8 vCPU, 61 GB RAM, $3.06/h
4xGPU: p3.8xlarge: 32 vCPU, 244 GB, $12.24/h
8xGPU: p3.16xlarge: 64v CPU, 488 GB., $24.48/h

Nice solution for a tournament. Probably more powerful than the 4xTPU of 
AlphaGo.

Rémi
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

[Computer-go] AGZ bootstrapping - Benefit of multi-labelled value net

[patrick.bardou via Computer-go]

@Gian-Carlo,
Indeed, multi-labelled value net/ head sounds a good way to inject more signal 
into the network, accorging to that paper, thus to inject more reinforcemenf 
learning signal for learning from scratch.
I was wondering if it could also be beneficial for bootstrapping the policy 
net/ head as well. Since score of randomly played games is likely to have high 
variance, I suppose that moves in most positions close to a final position will 
have very similar action values, as estimated  by the MCTS search, hence weak 
reinforcement signal (all/most of the moves leading to the same outcome, either 
win or loss). Using a multi-labelled value head could produce multi-labelled 
action values which might have more spread than with a fixed 6.5 komi and would 
allow better ranking of moves (by playing on the komi / averaging over the 
various possible komi ?)
Otherwise said, learning to increase the final score might be a good starting 
point for the policy net / head. At least in bootstrapping phase.
Combined with prioritized sampling biased towards low reverse move count from 
endgame, as I mentionned in an earlier post and as you propose for first round 
of learning win/loss of final position.
Patrick

 Message d'origine 
De : computer-go-requ...@computer-go.org 
Date : 26/10/2017  16:17  (GMT+01:00) 
À : computer-go@computer-go.org 
Objet : Computer-go Digest, Vol 93, Issue 34 


--

Message: 2
Date: Thu, 26 Oct 2017 15:17:43 +0200
From: Gian-Carlo Pascutto 
To: computer-go@computer-go.org
Subject: Re: [Computer-go] AlphaGo Zero
Message-ID: <8c872e71-4864-0a19-d3df-9fe1c48d2...@sjeng.org>
Content-Type: text/plain; charset=utf-8

On 25-10-17 16:00, Petr Baudis wrote:
> That makes sense.  I still hope that with a much more aggressive 
> training schedule we could train a reasonable Go player, perhaps at
> the expense of worse scaling at very high elos...  (At least I feel 
> optimistic after discovering a stupid bug in my code.)

By the way, a trivial observation: the initial network is random, so
there's no point in using it for playing the first batch of games. It
won't do anything useful until it has run a learning pass on a bunch of
"win/loss" scored games and it can at least tell who is the likely
winner in the final position (even if it mostly won't be able to make
territory at first).

This suggests that bootstrapping probably wants 500k starting games with
just random moves.

FWIW, it does not seem easy to get the value part of the network to
converge in the dual-res architecture, even when taking the appropriate
steps (1% weighting on error, strong regularizer).

-- 
GCP


--

Message: 3
Date: Thu, 26 Oct 2017 15:55:23 +0200
From: Roel van Engelen 
To: computer-go@computer-go.org
Subject: Re: [Computer-go] Source code (Was: Reducing network size?
(Was: AlphaGo Zero))
Message-ID:

Content-Type: text/plain; charset="utf-8"

@Gian-Carlo Pascutto

Since training uses a ridiculous amount of computing power i wonder if it
would
be useful to make certain changes for future research, like training the
value head
with multiple komi values 


___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Zero is weaker than Master!?

[Brian Sheppard via Computer-go]

I would add that "wild guesses based on not enough info" is an indispensable 
skill.

-Original Message-
From: Computer-go [mailto:computer-go-boun...@computer-go.org] On Behalf Of 
Hideki Kato
Sent: Thursday, October 26, 2017 10:17 AM
To: computer-go@computer-go.org
Subject: Re: [Computer-go] Zero is weaker than Master!?

Xavier Combelle: <62b977d7-d227-a74d-04b7-0d46db6a7...@gmail.com>:
>It is just wild guesses  based on reasonable arguments but without 
>evidence.


Yes, of course. Due to not enough info provided by Google.

Hideki


>Le 26/10/2017
à 07:51, Hideki Kato a écrit :
>> You can believe
>>> 
Of what I understand same network architecture imply the same number of
>>> block
>> 
but David Silver told AlphaGo Master used 40 layers in 
>> 
May. 
>> http://www.bestchinanews.com/Science-Technology/1
0371.html
>> # The paper was submitted in April.

>>
>> Usually, network "architecture" does not imply the num
ber of 
>> layers whereas "configulation" may do.

>>
>> Clearly they made 40 layers version first because it's
 
>> called "1st instance" where the 80 layers one is called
 "2nd 
>> instance."  The 1st was trained 3 days and overtoo
k AlphaGo 
>> Lee.  Then they changed to the 2nd.  Awaring t
his fact, and 
>> watching the growing curve of the 1st, I g
uess 40 layers was 
>> not enough to reach AlphaGo Master le
vel and so they 
>> doubled the layers.

>>
>> Hideki

>>
>> Xavier Combelle: <1550c907-8b96-e4ea-1f5e-2344f394b967
@gmail.com>:
>>> As I understand the paper they directly cre
ated alphago zero with a 40 
>>> block

>>> setup.
>>> They just made a reduced 20 block setup to co
mpare on kifu prediction
>>> (as far as I searched in the pa
per, it is the only
>>> place where they mention the 20 bloc
k setup)
>>> They specifically mention comparing several ver
sion of their software.
>>> with various parameter

>>> If the number of block was an important parameter I hope they would

>>> mention it.

>>> Of course they are a lot of things that they try and failed and we 
>>> will

>>> not know about

>>> But I have hard time to believe that alphago zero with a 20 block is 
>>> one

>>> of them

>>> About the paper, there is no mention of the number of block of master:

>>> "AlphaGo Master is the program that defeated top human players by 
>>> 600

>>> in January, 2017 34 .

>>> It was previously unpublished but uses the same neural network

>>> architecture, reinforcement

>>> learning algorithm, and MCTS algorithm as described in this paper.

>>> However, it uses the

>>> same handcrafted features and rollouts as AlphaGo Lee

>>> and training was initialised by

>>> supervised learning from human data."

>>> Of what I understand same network architecture imply the same number 
>>> of

>>> block

>>> Le 25/10/2017 à 17:58, Xavier Combelle a écrit :

 I understand better

 Le 25/10/2017 à 04:28, Hideki Kato a écrit :

> Are you thinking the 1st instance could reach Master level

> if giving more training days?

> I don't think so.  The performance would be stopping

> improving at 3 days.  If not, why they built the 2nd

> instance?

> Best,

> Hideki

> Xavier Combelle: <05c04de1-59c4-8fcd-2dd1-094faabf3...@gmail.com>:

>> How is it a fair comparison if there is only 3 days of training 
>> for

>>> Zero ?

>> Master had longer training no ? Moreover, Zero has bootstrap 
>> problem

>> because at the opposite of Master it don't learn from expert 
>> games

>> which means that it is likely to be weaker with little training.

>> Le 24/10/2017 à 20:20, Hideki Kato a écrit :

>>> David Silver told Master used 40 layers network in May. 

>>> According to new paper, Master used the same architecture

>>> as Zero.  So, Master used 20 blocks ResNet.  

>>> The first instance of Zero, 20 blocks ResNet version, is

>>> weaker than Master (after 3 days training).  So, with the

>>> same layers (a fair comparison) Zero is weaker than

>>> Master.

>>> Hideki

>> ___

>> Computer-go mailing list

>> Computer-go@computer-go.org

>> http://computer-go.org/mailman/listinfo/computer-go

>>> ___

>>> Computer-go mailing list

>>> Computer-go@computer-go.org

>>> http://computer-go.org/mailman/listinfo/computer-go

>

>___

>Computer-go mailing list

>Computer-go@computer-go.org

>http://computer-go.org/mailman/listinfo/computer-go
--
Hideki Kato  
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?

[Brian Sheppard via Computer-go]

Well,... good luck with that! :-)

Seriously: it is important to account for p-space completeness. That is, a set 
of rules that covers Go without conflict must be exponential in space usage.

Search has a triple role in system design. It accounts (at least 
asymptotically) for missing knowledge and also papers over disagreements 
between rules. It also evaluates the global situation, which allows rules to be 
expressed in terms of purely local effects.

From my perspective, that is too good a deal to pass by. But I don't want to be 
only a bearer of bad news. If you accept a limitation on your rule sets, then 
there is a higher level conflict resolution method that will lead to good 
results.

Your rules could express their effect as a local point gain, in the sense of 
"temperature". That is, temperature == the difference between moving first and 
letting the opponent move first. Then CGT provides a higher-order theory for 
rationalizing multiple priorities.

This suggestion only addresses one of the three roles of search, though perhaps 
the most important one.

Best,
Brian


-Original Message-
From: Computer-go [mailto:computer-go-boun...@computer-go.org] On Behalf Of 
Robert Jasiek
Sent: Thursday, October 26, 2017 10:17 AM
To: computer-go@computer-go.org
Subject: Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?

On 26.10.2017 13:52, Brian Sheppard via Computer-go wrote:
> MCTS is the glue that binds incompatible rules.

This is, however, not what I mean. Conflicting principles (call them rules if 
you like) must be dissolved by higher order principles. Only when all conflicts 
are dissolved should MCTS be applied.

What you describe has been used with success and better success than I expect 
what my knowledge-pure approach can currently achieve. But MCTS as glue for 
conflicting principles has also run into a boundary. I want to see that 
boundary surpassed by my pure approach.

--
robert jasiek
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Source code (Was: Reducing network size? (Was: AlphaGo Zero))

[Detlef Schmicker]

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 wonder if it
> would
> be useful to make certain changes for future research, like training the
> value head
> with multiple komi values 
> 
> On 26 October 2017 at 03:02, Brian Sheppard via Computer-go <
> computer-go@computer-go.org> wrote:
> 
>> I think it uses the champion network. That is, the training periodically
>> generates a candidate, and there is a playoff against the current champion.
>> If the candidate wins by more than 55% then a new champion is declared.
>>
>>
>>
>> Keeping a champion is an important mechanism, I believe. That creates the
>> competitive coevolution dynamic, where the network is evolving to learn how
>> to beat the best, and not just most recent. Without that dynamic, the
>> training process can go up and down.
>>
>>
>>
>> *From:* Computer-go [mailto:computer-go-boun...@computer-go.org] *On
>> Behalf Of *uurtamo .
>> *Sent:* Wednesday, October 25, 2017 6:07 PM
>> *To:* computer-go 
>> *Subject:* Re: [Computer-go] Source code (Was: Reducing network size?
>> (Was: AlphaGo Zero))
>>
>>
>>
>> Does the self-play step use the most recent network for each move?
>>
>>
>>
>> On Oct 25, 2017 2:23 PM, "Gian-Carlo Pascutto"  wrote:
>>
>> On 25-10-17 17:57, Xavier Combelle wrote:
>>> Is there some way to distribute learning of a neural network ?
>>
>> Learning as in training the DCNN, not really unless there are high
>> bandwidth links between the machines (AFAIK - unless the state of the
>> art changed?).
>>
>> Learning as in generating self-play games: yes. Especially if you update
>> the network only every 25 000 games.
>>
>> My understanding is that this task is much more bottlenecked on game
>> generation than on DCNN training, until you get quite a bit of machines
>> that generate games.
>>
>> --
>> GCP
>> ___
>> Computer-go mailing list
>> Computer-go@computer-go.org
>> http://computer-go.org/mailman/listinfo/computer-go
>>
>>
>> ___
>> Computer-go mailing list
>> Computer-go@computer-go.org
>> http://computer-go.org/mailman/listinfo/computer-go
>>
> 
> 
> 
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
> 
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Source code (Was: Reducing network size? (Was: AlphaGo Zero))

[Erik van der Werf]

Good point, Roel. Perhaps in the final layers one could make it predict a
model of the expected score distribution (before combining with the komi
and other rules specific adjustments for handicap stones, pass stones,
last-play parity, etc.). Should be easy enough to back-propagate win/loss
information (and perhaps even more) through such a model.


On Thu, Oct 26, 2017 at 3:55 PM, Roel van Engelen 
wrote:

> @Gian-Carlo Pascutto
>
> Since training uses a ridiculous amount of computing power i wonder if it
> would
> be useful to make certain changes for future research, like training the
> value head
> with multiple komi values 
>
> On 26 October 2017 at 03:02, Brian Sheppard via Computer-go <
> computer-go@computer-go.org> wrote:
>
>> I think it uses the champion network. That is, the training periodically
>> generates a candidate, and there is a playoff against the current champion.
>> If the candidate wins by more than 55% then a new champion is declared.
>>
>>
>>
>> Keeping a champion is an important mechanism, I believe. That creates the
>> competitive coevolution dynamic, where the network is evolving to learn how
>> to beat the best, and not just most recent. Without that dynamic, the
>> training process can go up and down.
>>
>>
>>
>> *From:* Computer-go [mailto:computer-go-boun...@computer-go.org] *On
>> Behalf Of *uurtamo .
>> *Sent:* Wednesday, October 25, 2017 6:07 PM
>> *To:* computer-go 
>> *Subject:* Re: [Computer-go] Source code (Was: Reducing network size?
>> (Was: AlphaGo Zero))
>>
>>
>>
>> Does the self-play step use the most recent network for each move?
>>
>>
>>
>> On Oct 25, 2017 2:23 PM, "Gian-Carlo Pascutto"  wrote:
>>
>> On 25-10-17 17:57, Xavier Combelle wrote:
>> > Is there some way to distribute learning of a neural network ?
>>
>> Learning as in training the DCNN, not really unless there are high
>> bandwidth links between the machines (AFAIK - unless the state of the
>> art changed?).
>>
>> Learning as in generating self-play games: yes. Especially if you update
>> the network only every 25 000 games.
>>
>> My understanding is that this task is much more bottlenecked on game
>> generation than on DCNN training, until you get quite a bit of machines
>> that generate games.
>>
>> --
>> GCP
>> ___
>> Computer-go mailing list
>> Computer-go@computer-go.org
>> http://computer-go.org/mailman/listinfo/computer-go
>>
>>
>> ___
>> Computer-go mailing list
>> Computer-go@computer-go.org
>> http://computer-go.org/mailman/listinfo/computer-go
>>
>
>
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?

[Xavier Combelle]

what are semantic genetic algorithm ?

to my knowledge genetic algorithm lead to poor result except as a
metaheuristic in optimisation problem


Le 26/10/2017 à 14:40, Jim O'Flaherty a écrit :
> When I get time to spend dozens of hours on computer go again, I plan
> to play in Robert's area with semantic genetic algorithms. I am an
> Architect Software Engineer. Robert's work will allow me better than
> starting entirely from random in much the same way AlphaGo
> bootstrapped from the 100K of professional games. AG0 then leveraged
> AlphaGo in knowing an architecture that was close enough. My intuition
> is my approach will be something similar in it's evolution.
>
> This is the way we're going to "automate" creating provided proofing
> of human cognition styled computer go players to assist humans in a
> gradient ascent learning cycle.
>
> So, Robert, I admire and am encouraged by your research for my own
> computer go projects in this area. Keep kicking butt in your unique
> way. We are in an interesting transition in this community. Stick it
> out. It will be worth it long term.
>
> On Oct 26, 2017 4:38 AM, "Petri Pitkanen"  > wrote:
>
> Unfortunately there is no proof that you principles work better
> than those form eighties. Nor there is any agreement that your
> pronciples form any improvement over the old ones. Yes you are a 
> far better player than me and shows that you are 
> - way better at reading 
> - have hugely better go understanding, principles if you like
>
> What is missing that I doubt that you can verbalise your go
> understanding to degree that by applying those principles  I could
> become substantially better player. again bulleting
> - My reading skills would not get any better hence making much of
> value any learning moot. Obviously issue on me not on your principles
> - your principles are more complex than you understand. Much of
> you know is automated to degree that it is subconsciousness
> information. Transferring that information if hard. Usually done
> by re-playing master games looking at problems i.e. training the
> darn neural net in the head
>
> If you can build Go bot about  KGS 3/4dan strength I am more than
> willing to admit you are right and would even consider buying
> your  books.
>
> Petri
>
> 2017-10-26 6:21 GMT+03:00 Robert Jasiek  >:
>
> On 25.10.2017 18:17, Xavier Combelle wrote:
>
> exact go theory is full of hole.
>
>
> WRT describing the whole game, yes, this is the current state.
> Solving go in a mathematical sense is a project for centuries.
>
> Actually, to my knowledge human can't apply only the exact
> go theory and
> play a decent game.
>
>
> Only for certain positions of a) late endgame, b) semeais, c) ko.
>
> If human can't do that, how it will teach a computer to do
> it magically ?
>
>
> IIRC, Martin Müller implemented CGT endgames a la Mathematical
> Go Endgames.
>
> The reason why (b) had became unpopular is because there
> is no go theory
> precise enough to implement it as an algorithm
>
>
> There is quite some theory of the 95% principle kind which
> might be implemented as approximation. E.g. "Usually, defend
> your weak important group." can be approximated by
> approximating "group", "important" (its loss is too large in a
> quick positional judgement), "weak" (can be killed in two
> successive moves), "defend" (after the move, cannot be killed
> in two successive moves), "usually" (always, unless there are
> several such groups and some must be chosen, say, randomly;
> the approximation being that the alternative strategy of large
> scale exchange is discarded).
>
> Besides, one must prioritise principles to solve conflicting
> principles by a higher order principle.
>
> IMO, such an expert system combined with tree reading and
> maybe MCTS to emulate reading used when a principle depends on
> reading can, with an effort of a few manyears of
> implementation, already achieve amateur mid dan. Not high dan
> yet because high dans can choose advanced strategies, such as
> global exchange, and there are no good enough principles for
> that yet, which would also consider necessary side conditions
> related to influence, aji etc. I need to work out such
> principles during the following years. Currently, the state is
> that weaker principles have identified the major topics
> (influence, aji etc.) to be considered in fights but they must
> be refined to create 95%+ principles.
>
> ***
>
> 

Re: [Computer-go] Zero is weaker than Master!?

[Xavier Combelle]

Unless I mistake figure 3 shows the plot of supervised learning to
reinforcement learning, not 20 bloc/40 block

For searching mention of the 20 blocks I search for 20 in the whole
paper and did not found any other mention

than of the kifu thing.


Le 26/10/2017 à 15:10, Gian-Carlo Pascutto a écrit :
> On 26-10-17 10:55, Xavier Combelle wrote:
>> It is just wild guesses  based on reasonable arguments but without
>> evidence.
> David Silver said they used 40 layers for AlphaGo Master. That's more
> evidence than there is for the opposite argument that you are trying to
> make. The paper certainly doesn't talk about a "small" and a "big" Master.
>
> You seem to be arguing from a bunch of misreadings and
> misunderstandings. For example, Figure 3 in the paper shows the Elo plot
> for the 20 block/40 layer version, and it compares to Alpha Go Lee, not
> Alpha Go Master. The Alpha Go Master line would be above the flattening
> part of the 20 block/40 layer AlphaGo Zero. I guess you missed this when
> you say that they "only mention it to compare on kifu prediction"?
>

___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] AlphaGo Zero

[Andy]

I agree with your main point that the first batch of games will be totally
random moves. I just wanted to make a small point that even for totally
random play, the network should be able to learn something about mid-game
positions as well. At move 100, a position with 50 white stones and 40
black stones is likely to be a win for white, even with completely random
play from there, since white has captured 10 black stones.


2017-10-26 8:17 GMT-05:00 Gian-Carlo Pascutto :

> On 25-10-17 16:00, Petr Baudis wrote:
> > That makes sense.  I still hope that with a much more aggressive
> > training schedule we could train a reasonable Go player, perhaps at
> > the expense of worse scaling at very high elos...  (At least I feel
> > optimistic after discovering a stupid bug in my code.)
>
> By the way, a trivial observation: the initial network is random, so
> there's no point in using it for playing the first batch of games. It
> won't do anything useful until it has run a learning pass on a bunch of
> "win/loss" scored games and it can at least tell who is the likely
> winner in the final position (even if it mostly won't be able to make
> territory at first).
>
> This suggests that bootstrapping probably wants 500k starting games with
> just random moves.
>
> FWIW, it does not seem easy to get the value part of the network to
> converge in the dual-res architecture, even when taking the appropriate
> steps (1% weighting on error, strong regularizer).
>
> --
> GCP
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?

[Robert Jasiek]

On 26.10.2017 08:52, Petri Pitkanen wrote:

Unfortunately there is no proof that you principles work better than those
form eighties.


No computer-go proof.

There is evidence in the form of my playing strength: with the 
principles "from the eighties", I got to circa 1 kyu. L+D reading 
practice etc. made me 3 dan. Afterwards, almost the only thing that made 
me stronger to 5 dan and then further improved my understanding was the 
invention of my own principles.


My principles etc. also work for (an unknown fraction of) readers of my 
books and for a high percentage of my pupils but I cannot compare what 
the effect on them would have been if instead they would only have 
learnt the principles "from the eighties". I do, however, know that my 
principles provide me with very much more efficient means of teaching 
contents compared to using the principles "from the eighties".


The principles "from the eighties" and my principles can be compared 
with each other. IMO, such a comparison is shocking: the principles 
"from the eighties" are very much weaker on average and altogether 
convey very much less contents.



Nor there is any agreement that your pronciples form any
improvement over the old ones.


Only time constraints prevent me from doing an extensive comparison and 
so better support formation of an agreement.



What is missing that I doubt that you can verbalise your go understanding
to degree that by applying those principles  I could become substantially
better player.


Different players are different. So different that some players claim to 
only learn from examples. Therefore, I cannot know whether you are a 
player who could learn well from principles etc.



- My reading skills would not get any better


Do you say so after having learnt and invested effort in applying the 
contents of Tactical Reading?


Regardless of the possible impact of that book, a great part of reading 
skill must be obtained by reading practice in games and problem solving. 
If your reading is much weaker than your knowledge of go theory, then it 
may be the case that almost only reading practise (plus possibly reading 
theory about improving one's reading practice) can significantly improve 
your strength at the moment.



- your principles are more complex than you understand.


I do not think so:)


Much of you know is
automated to degree that it is subconsciousness information.


From ca. 10 kyu to now, especially from 3 dan to now, I have reduced 
the impact of my subconscious thinking on my go decision-making and 
replaced it by knowledge, reading and positional judgement based on 
knowledge and reading. The still remaining subconscious thinking is 
small. Most of my remaining mistakes are related to psychology or 
subconscious thinking, when necessary because of explicit knowledge gaps 
or thinking time constraints.



Transferring that information if hard.


Transferring it from principles etc. to code - yes.


If you can build Go bot about  KGS 3/4dan strength


Using my approach, I expect several manyears, which I do not have for 
that purpose.


--
robert jasiek
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] November KGS bot tournament

[Hideki Kato]

The link should be 
.

Hideki

Nick Wedd: :
>The November KGS bot tournament will be on Sunday, November 5th, starting
>at 16:00 UTC and ending by 22:00 UTC.  It will use 19x19 boards, with
>time limits
>of 14 minutes each and very fast Canadian overtime, and komi of 7½.  It
>will be a Swiss tournament.  See http://www.gokgs.com/tournInfo.jsp?id=112
>7
>
>Please register by emailing me at mapr...@gmail.com, with the words "KGS
>Tournament Registration" in the email title.
>With the falling interest in these events since the advent of AlphaGo, it
>is likely that this will be the last of the series of KGS bot tournaments.
>
>Nick
-- 
Hideki Kato 
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] I present my apologizes to Robert Jasiek

[Robert Jasiek]

Accepted, thank you!

--
robert jasiek
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?

[Robert Jasiek]

On 26.10.2017 13:52, Brian Sheppard via Computer-go wrote:

MCTS is the glue that binds incompatible rules.


This is, however, not what I mean. Conflicting principles (call them 
rules if you like) must be dissolved by higher order principles. Only 
when all conflicts are dissolved should MCTS be applied.


What you describe has been used with success and better success than I 
expect what my knowledge-pure approach can currently achieve. But MCTS 
as glue for conflicting principles has also run into a boundary. I want 
to see that boundary surpassed by my pure approach.


--
robert jasiek
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Zero is weaker than Master!?

[Hideki Kato]

Xavier Combelle: <62b977d7-d227-a74d-04b7-0d46db6a7...@gmail.com>:
>It is just wild guesses  based on reasonable arguments but without 
>evidence.


Yes, of course. Due to not enough info provided by Google.

Hideki


>Le 26/10/2017 
à 07:51, Hideki Kato a écrit :
>> You can believe
>>> 
Of what I understand same network architecture imply the 
same number of
>>> block
>> 
but David Silver told AlphaGo Master used 40 layers in 
>> 
May. 
>> http://www.bestchinanews.com/Science-Technology/1
0371.html
>> # The paper was submitted in April.

>>
>> Usually, network "architecture" does not imply the num
ber of 
>> layers whereas "configulation" may do.

>>
>> Clearly they made 40 layers version first because it's
 
>> called "1st instance" where the 80 layers one is called
 "2nd 
>> instance."  The 1st was trained 3 days and overtoo
k AlphaGo 
>> Lee.  Then they changed to the 2nd.  Awaring t
his fact, and 
>> watching the growing curve of the 1st, I g
uess 40 layers was 
>> not enough to reach AlphaGo Master le
vel and so they 
>> doubled the layers.

>>
>> Hideki

>>
>> Xavier Combelle: <1550c907-8b96-e4ea-1f5e-2344f394b967
@gmail.com>:
>>> As I understand the paper they directly cre
ated alphago zero with a 40 
>>> block

>>> setup.
>>> They just made a reduced 20 block setup to co
mpare on kifu prediction
>>> (as far as I searched in the pa
per, it is the only
>>> place where they mention the 20 bloc
k setup)
>>> They specifically mention comparing several ver
sion of their software.
>>> with various parameter

>>> If the number of block was an important parameter I hope they would

>>> mention it.

>>> Of course they are a lot of things that they try and failed and we will

>>> not know about

>>> But I have hard time to believe that alphago zero with a 20 block is one

>>> of them

>>> About the paper, there is no mention of the number of block of master:

>>> "AlphaGo Master is the program that defeated top human players by 600

>>> in January, 2017 34 .

>>> It was previously unpublished but uses the same neural network

>>> architecture, reinforcement

>>> learning algorithm, and MCTS algorithm as described in this paper.

>>> However, it uses the

>>> same handcrafted features and rollouts as AlphaGo Lee

>>> and training was initialised by

>>> supervised learning from human data."

>>> Of what I understand same network architecture imply the same number of

>>> block

>>> Le 25/10/2017 à 17:58, Xavier Combelle a écrit :

 I understand better

 Le 25/10/2017 à 04:28, Hideki Kato a écrit :

> Are you thinking the 1st instance could reach Master level 

> if giving more training days?

> I don't think so.  The performance would be stopping 

> improving at 3 days.  If not, why they built the 2nd 

> instance?

> Best,

> Hideki

> Xavier Combelle: <05c04de1-59c4-8fcd-2dd1-094faabf3...@gmail.com>:

>> How is it a fair comparison if there is only 3 days of training for 

>>> Zero ?

>> Master had longer training no ? Moreover, Zero has bootstrap problem

>> because at the opposite of Master it don't learn from expert games

>> which means that it is likely to be weaker with little training.

>> Le 24/10/2017 à 20:20, Hideki Kato a écrit :

>>> David Silver told Master used 40 layers network in May. 

>>> According to new paper, Master used the same architecture 

>>> as Zero.  So, Master used 20 blocks ResNet.  

>>> The first instance of Zero, 20 blocks ResNet version, is 

>>> weaker than Master (after 3 days training).  So, with the 

>>> same layers (a fair comparison) Zero is weaker than 

>>> Master.

>>> Hideki

>> ___

>> Computer-go mailing list

>> Computer-go@computer-go.org

>> http://computer-go.org/mailman/listinfo/computer-go

>>> ___

>>> Computer-go mailing list

>>> Computer-go@computer-go.org

>>> http://computer-go.org/mailman/listinfo/computer-go

>

>___

>Computer-go mailing list

>Computer-go@computer-go.org

>http://computer-go.org/mailman/listinfo/computer-go
-- 
Hideki Kato 
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Source code (Was: Reducing network size? (Was: AlphaGo Zero))

[Roel van Engelen]

@Gian-Carlo Pascutto

Since training uses a ridiculous amount of computing power i wonder if it
would
be useful to make certain changes for future research, like training the
value head
with multiple komi values 

On 26 October 2017 at 03:02, Brian Sheppard via Computer-go <
computer-go@computer-go.org> wrote:

> I think it uses the champion network. That is, the training periodically
> generates a candidate, and there is a playoff against the current champion.
> If the candidate wins by more than 55% then a new champion is declared.
>
>
>
> Keeping a champion is an important mechanism, I believe. That creates the
> competitive coevolution dynamic, where the network is evolving to learn how
> to beat the best, and not just most recent. Without that dynamic, the
> training process can go up and down.
>
>
>
> *From:* Computer-go [mailto:computer-go-boun...@computer-go.org] *On
> Behalf Of *uurtamo .
> *Sent:* Wednesday, October 25, 2017 6:07 PM
> *To:* computer-go 
> *Subject:* Re: [Computer-go] Source code (Was: Reducing network size?
> (Was: AlphaGo Zero))
>
>
>
> Does the self-play step use the most recent network for each move?
>
>
>
> On Oct 25, 2017 2:23 PM, "Gian-Carlo Pascutto"  wrote:
>
> On 25-10-17 17:57, Xavier Combelle wrote:
> > Is there some way to distribute learning of a neural network ?
>
> Learning as in training the DCNN, not really unless there are high
> bandwidth links between the machines (AFAIK - unless the state of the
> art changed?).
>
> Learning as in generating self-play games: yes. Especially if you update
> the network only every 25 000 games.
>
> My understanding is that this task is much more bottlenecked on game
> generation than on DCNN training, until you get quite a bit of machines
> that generate games.
>
> --
> GCP
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
>
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] AlphaGo Zero

[Gian-Carlo Pascutto]

On 25-10-17 16:00, Petr Baudis wrote:
> That makes sense.  I still hope that with a much more aggressive 
> training schedule we could train a reasonable Go player, perhaps at
> the expense of worse scaling at very high elos...  (At least I feel 
> optimistic after discovering a stupid bug in my code.)

By the way, a trivial observation: the initial network is random, so
there's no point in using it for playing the first batch of games. It
won't do anything useful until it has run a learning pass on a bunch of
"win/loss" scored games and it can at least tell who is the likely
winner in the final position (even if it mostly won't be able to make
territory at first).

This suggests that bootstrapping probably wants 500k starting games with
just random moves.

FWIW, it does not seem easy to get the value part of the network to
converge in the dual-res architecture, even when taking the appropriate
steps (1% weighting on error, strong regularizer).

-- 
GCP
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Zero is weaker than Master!?

[Gian-Carlo Pascutto]

On 26-10-17 10:55, Xavier Combelle wrote:
> It is just wild guesses  based on reasonable arguments but without
> evidence.

David Silver said they used 40 layers for AlphaGo Master. That's more
evidence than there is for the opposite argument that you are trying to
make. The paper certainly doesn't talk about a "small" and a "big" Master.

You seem to be arguing from a bunch of misreadings and
misunderstandings. For example, Figure 3 in the paper shows the Elo plot
for the 20 block/40 layer version, and it compares to Alpha Go Lee, not
Alpha Go Master. The Alpha Go Master line would be above the flattening
part of the 20 block/40 layer AlphaGo Zero. I guess you missed this when
you say that they "only mention it to compare on kifu prediction"?

-- 
GCP
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?

[Jim O'Flaherty]

When I get time to spend dozens of hours on computer go again, I plan to
play in Robert's area with semantic genetic algorithms. I am an Architect
Software Engineer. Robert's work will allow me better than starting
entirely from random in much the same way AlphaGo bootstrapped from the
100K of professional games. AG0 then leveraged AlphaGo in knowing an
architecture that was close enough. My intuition is my approach will be
something similar in it's evolution.

This is the way we're going to "automate" creating provided proofing of
human cognition styled computer go players to assist humans in a gradient
ascent learning cycle.

So, Robert, I admire and am encouraged by your research for my own computer
go projects in this area. Keep kicking butt in your unique way. We are in
an interesting transition in this community. Stick it out. It will be worth
it long term.

On Oct 26, 2017 4:38 AM, "Petri Pitkanen" 
wrote:

> Unfortunately there is no proof that you principles work better than those
> form eighties. Nor there is any agreement that your pronciples form any
> improvement over the old ones. Yes you are a  far better player than me and
> shows that you are
> - way better at reading
> - have hugely better go understanding, principles if you like
>
> What is missing that I doubt that you can verbalise your go understanding
> to degree that by applying those principles  I could become substantially
> better player. again bulleting
> - My reading skills would not get any better hence making much of value
> any learning moot. Obviously issue on me not on your principles
> - your principles are more complex than you understand. Much of you know
> is automated to degree that it is subconsciousness information.
> Transferring that information if hard. Usually done by re-playing master
> games looking at problems i.e. training the darn neural net in the head
>
> If you can build Go bot about  KGS 3/4dan strength I am more than willing
> to admit you are right and would even consider buying your  books.
>
> Petri
>
> 2017-10-26 6:21 GMT+03:00 Robert Jasiek :
>
>> On 25.10.2017 18:17, Xavier Combelle wrote:
>>
>>> exact go theory is full of hole.
>>>
>>
>> WRT describing the whole game, yes, this is the current state. Solving go
>> in a mathematical sense is a project for centuries.
>>
>> Actually, to my knowledge human can't apply only the exact go theory and
>>> play a decent game.
>>>
>>
>> Only for certain positions of a) late endgame, b) semeais, c) ko.
>>
>> If human can't do that, how it will teach a computer to do it magically ?
>>>
>>
>> IIRC, Martin Müller implemented CGT endgames a la Mathematical Go
>> Endgames.
>>
>> The reason why (b) had became unpopular is because there is no go theory
>>> precise enough to implement it as an algorithm
>>>
>>
>> There is quite some theory of the 95% principle kind which might be
>> implemented as approximation. E.g. "Usually, defend your weak important
>> group." can be approximated by approximating "group", "important" (its loss
>> is too large in a quick positional judgement), "weak" (can be killed in two
>> successive moves), "defend" (after the move, cannot be killed in two
>> successive moves), "usually" (always, unless there are several such groups
>> and some must be chosen, say, randomly; the approximation being that the
>> alternative strategy of large scale exchange is discarded).
>>
>> Besides, one must prioritise principles to solve conflicting principles
>> by a higher order principle.
>>
>> IMO, such an expert system combined with tree reading and maybe MCTS to
>> emulate reading used when a principle depends on reading can, with an
>> effort of a few manyears of implementation, already achieve amateur mid
>> dan. Not high dan yet because high dans can choose advanced strategies,
>> such as global exchange, and there are no good enough principles for that
>> yet, which would also consider necessary side conditions related to
>> influence, aji etc. I need to work out such principles during the following
>> years. Currently, the state is that weaker principles have identified the
>> major topics (influence, aji etc.) to be considered in fights but they must
>> be refined to create 95%+ principles.
>>
>> ***
>>
>> In the 80s and 90s, expert systems failed to do better than ca. 5 kyu
>> because principles were only marginally better than 50%. Today, (my)
>> average principles discard the weaker, 50% principles and are ca. 75%.
>> Tomorrow, the 75% principles can be discarded for an average of 95%
>> principles. Expert systems get their chance again! Their major disadvantage
>> remains: great manpower is required for implementation. The advantage is
>> semantical understanding.
>>
>> --
>> robert jasiek
>>
>> ___
>> Computer-go mailing list
>> Computer-go@computer-go.org
>> http://computer-go.org/mailman/listinfo/computer-go
>>
>
>
> 

Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?

[Brian Sheppard via Computer-go]

Robert is right, but Robert seems to think this hasn't been done. Actually 
every prominent non-neural MCTS program since Mogo has been based on the exact 
design that Robert describes. The best of them achieve somewhat greater 
strength than Robert expects.

MCTS is the glue that binds incompatible rules. It rationalizes different 
heuristics into a coherent whole by testing the ideas in a competition against 
one another using a meaningful evaluation (win/loss).

Best,
Brian

-Original Message-
From: Computer-go [mailto:computer-go-boun...@computer-go.org] On Behalf Of 
Xavier Combelle
Sent: Thursday, October 26, 2017 1:50 AM
To: computer-go@computer-go.org
Subject: Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?



>> The reason why (b) had became unpopular is because there is no go 
>> theory precise enough to implement it as an algorithm
>
> There is quite some theory of the 95% principle kind which might be 
> implemented as approximation. E.g. "Usually, defend your weak 
> important group." can be approximated by approximating "group", 
> "important" (its loss is too large in a quick positional judgement), 
> "weak" (can be killed in two successive moves), "defend" (after the 
> move, cannot be killed in two successive moves), "usually" (always, 
> unless there are several such groups and some must be chosen, say, 
> randomly; the approximation being that the alternative strategy of 
> large scale exchange is discarded).
>
> Besides, one must prioritise principles to solve conflicting 
> principles by a higher order principle.
>
> IMO, such an expert system combined with tree reading and maybe MCTS 
> to emulate reading used when a principle depends on reading can, with 
> an effort of a few manyears of implementation, already achieve amateur 
> mid dan. Not high dan yet because high dans can choose advanced 
> strategies, such as global exchange, and there are no good enough 
> principles for that yet, which would also consider necessary side 
> conditions related to influence, aji etc. I need to work out such 
> principles during the following years. Currently, the state is that 
> weaker principles have identified the major topics (influence, aji
> etc.) to be considered in fights but they must be refined to create 
> 95%+ principles.
>
> ***
>
> In the 80s and 90s, expert systems failed to do better than ca. 5 kyu 
> because principles were only marginally better than 50%. Today, (my) 
> average principles discard the weaker, 50% principles and are ca. 75%.
> Tomorrow, the 75% principles can be discarded for an average of 95% 
> principles. Expert systems get their chance again! Their major 
> disadvantage remains: great manpower is required for implementation.
> The advantage is semantical understanding.
>
From a software developer point of view enlighten by my knowledge of history of 
ai and history of go development,
 such approximate definition is close to useless to build a software at the 
current state of art.
One of the reason is as you state the considerable work it would require to 
implement a huge number of imprecise rules.
As you are not a software developer, I want you to look on this comics which 
state the difference between apparent difficulty and real difficulty of 
developping software. https://xkcd.com/1425/ As far as I understand your task 
to implement such an expert system would require the many years of 
implementations would be thousands of years.
As far as my experience speak the expected reward would be a win of one or two 
rank and so definitely not a mid dan amateur level.

Xavier Combelle

___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

[Computer-go] I present my apologizes to Robert Jasiek

[Xavier Combelle]

I present my apologizes to Robert jasiek.
To my knowledge all his behavior on this list was always correct
and my initial and my subsequent mail was inappropriate

Xavier Combelle


___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Zero is weaker than Master!?

[Xavier Combelle]

It is just wild guesses  based on reasonable arguments but without evidence.


Le 26/10/2017 à 07:51, Hideki Kato a écrit :
> You can believe
>> Of what I understand same network architecture imply the same number of
>> block
> but David Silver told AlphaGo Master used 40 layers in 
> May. 
> http://www.bestchinanews.com/Science-Technology/10371.html
> # The paper was submitted in April.
>
> Usually, network "architecture" does not imply the number of 
> layers whereas "configulation" may do.
>
> Clearly they made 40 layers version first because it's 
> called "1st instance" where the 80 layers one is called "2nd 
> instance."  The 1st was trained 3 days and overtook AlphaGo 
> Lee.  Then they changed to the 2nd.  Awaring this fact, and 
> watching the growing curve of the 1st, I guess 40 layers was 
> not enough to reach AlphaGo Master level and so they 
> doubled the layers.
>
> Hideki
>
> Xavier Combelle: <1550c907-8b96-e4ea-1f5e-2344f394b...@gmail.com>:
>> As I understand the paper they directly created alphago zero with a 40 
>> block
>> setup.
>> They just made a reduced 20 block setup to compare on kifu prediction
>> (as far as I searched in the paper, it is the only
>> place where they mention the 20 block setup)
>> They specifically mention comparing several version of their software.
>> with various parameter
>> If the number of block was an important parameter I hope they would
>> mention it.
>> Of course they are a lot of things that they try and failed and we will
>> not know about
>> But I have hard time to believe that alphago zero with a 20 block is one
>> of them
>> About the paper, there is no mention of the number of block of master:
>> "AlphaGo Master is the program that defeated top human players by 60–0
>> in January, 2017 34 .
>> It was previously unpublished but uses the same neural network
>> architecture, reinforcement
>> learning algorithm, and MCTS algorithm as described in this paper.
>> However, it uses the
>> same handcrafted features and rollouts as AlphaGo Lee
>> and training was initialised by
>> supervised learning from human data."
>> Of what I understand same network architecture imply the same number of
>> block
>> Le 25/10/2017 à 17:58, Xavier Combelle a écrit :
>>> I understand better
>>> Le 25/10/2017 à 04:28, Hideki Kato a écrit :
 Are you thinking the 1st instance could reach Master level 
 if giving more training days?
 I don't think so.  The performance would be stopping 
 improving at 3 days.  If not, why they built the 2nd 
 instance?
 Best,
 Hideki
 Xavier Combelle: <05c04de1-59c4-8fcd-2dd1-094faabf3...@gmail.com>:
> How is it a fair comparison if there is only 3 days of training for 
>> Zero ?
> Master had longer training no ? Moreover, Zero has bootstrap problem
> because at the opposite of Master it don't learn from expert games
> which means that it is likely to be weaker with little training.
> Le 24/10/2017 à 20:20, Hideki Kato a écrit :
>> David Silver told Master used 40 layers network in May. 
>> According to new paper, Master used the same architecture 
>> as Zero.  So, Master used 20 blocks ResNet.  
>> The first instance of Zero, 20 blocks ResNet version, is 
>> weaker than Master (after 3 days training).  So, with the 
>> same layers (a fair comparison) Zero is weaker than 
>> Master.
>> Hideki
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>> ___
>> Computer-go mailing list
>> Computer-go@computer-go.org
>> http://computer-go.org/mailman/listinfo/computer-go

___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

[Computer-go] November KGS bot tournament

[Nick Wedd]

The November KGS bot tournament will be on Sunday, November 5th, starting
at 16:00 UTC and ending by 22:00 UTC.  It will use 19x19 boards, with
time limits
of 14 minutes each and very fast Canadian overtime, and komi of 7½.  It
will be a Swiss tournament.  See http://www.gokgs.com/tournInfo.jsp?id=112
7

Please register by emailing me at mapr...@gmail.com, with the words "KGS
Tournament Registration" in the email title.
With the falling interest in these events since the advent of AlphaGo, it
is likely that this will be the last of the series of KGS bot tournaments.

Nick
-- 
Nick Wedd  mapr...@gmail.com
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?

[Petri Pitkanen]

Unfortunately there is no proof that you principles work better than those
form eighties. Nor there is any agreement that your pronciples form any
improvement over the old ones. Yes you are a  far better player than me and
shows that you are
- way better at reading
- have hugely better go understanding, principles if you like

What is missing that I doubt that you can verbalise your go understanding
to degree that by applying those principles  I could become substantially
better player. again bulleting
- My reading skills would not get any better hence making much of value any
learning moot. Obviously issue on me not on your principles
- your principles are more complex than you understand. Much of you know is
automated to degree that it is subconsciousness information. Transferring
that information if hard. Usually done by re-playing master games looking
at problems i.e. training the darn neural net in the head

If you can build Go bot about  KGS 3/4dan strength I am more than willing
to admit you are right and would even consider buying your  books.

Petri

2017-10-26 6:21 GMT+03:00 Robert Jasiek :

> On 25.10.2017 18:17, Xavier Combelle wrote:
>
>> exact go theory is full of hole.
>>
>
> WRT describing the whole game, yes, this is the current state. Solving go
> in a mathematical sense is a project for centuries.
>
> Actually, to my knowledge human can't apply only the exact go theory and
>> play a decent game.
>>
>
> Only for certain positions of a) late endgame, b) semeais, c) ko.
>
> If human can't do that, how it will teach a computer to do it magically ?
>>
>
> IIRC, Martin Müller implemented CGT endgames a la Mathematical Go Endgames.
>
> The reason why (b) had became unpopular is because there is no go theory
>> precise enough to implement it as an algorithm
>>
>
> There is quite some theory of the 95% principle kind which might be
> implemented as approximation. E.g. "Usually, defend your weak important
> group." can be approximated by approximating "group", "important" (its loss
> is too large in a quick positional judgement), "weak" (can be killed in two
> successive moves), "defend" (after the move, cannot be killed in two
> successive moves), "usually" (always, unless there are several such groups
> and some must be chosen, say, randomly; the approximation being that the
> alternative strategy of large scale exchange is discarded).
>
> Besides, one must prioritise principles to solve conflicting principles by
> a higher order principle.
>
> IMO, such an expert system combined with tree reading and maybe MCTS to
> emulate reading used when a principle depends on reading can, with an
> effort of a few manyears of implementation, already achieve amateur mid
> dan. Not high dan yet because high dans can choose advanced strategies,
> such as global exchange, and there are no good enough principles for that
> yet, which would also consider necessary side conditions related to
> influence, aji etc. I need to work out such principles during the following
> years. Currently, the state is that weaker principles have identified the
> major topics (influence, aji etc.) to be considered in fights but they must
> be refined to create 95%+ principles.
>
> ***
>
> In the 80s and 90s, expert systems failed to do better than ca. 5 kyu
> because principles were only marginally better than 50%. Today, (my)
> average principles discard the weaker, 50% principles and are ca. 75%.
> Tomorrow, the 75% principles can be discarded for an average of 95%
> principles. Expert systems get their chance again! Their major disadvantage
> remains: great manpower is required for implementation. The advantage is
> semantical understanding.
>
> --
> robert jasiek
>
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Zero is weaker than Master!?

[Hideki Kato]

You can believe
>Of what I understand same network architecture imply the same number of
>block
but David Silver told AlphaGo Master used 40 layers in 
May. 
http://www.bestchinanews.com/Science-Technology/10371.html
# The paper was submitted in April.

Usually, network "architecture" does not imply the number of 
layers whereas "configulation" may do.

Clearly they made 40 layers version first because it's 
called "1st instance" where the 80 layers one is called "2nd 
instance."  The 1st was trained 3 days and overtook AlphaGo 
Lee.  Then they changed to the 2nd.  Awaring this fact, and 
watching the growing curve of the 1st, I guess 40 layers was 
not enough to reach AlphaGo Master level and so they 
doubled the layers.

Hideki

Xavier Combelle: <1550c907-8b96-e4ea-1f5e-2344f394b...@gmail.com>:
>As I understand the paper they directly created alphago zero with a 40 
>block

>setup.

>

>They just made a reduced 20 block setup to compare on kifu prediction

>(as far as I searched in the paper, it is the only

>place where they mention the 20 block setup)

>

>They specifically mention comparing several version of their software.

>with various parameter

>

>If the number of block was an important parameter I hope they would

>mention it.

>

>Of course they are a lot of things that they try and failed and we will

>not know about

>

>But I have hard time to believe that alphago zero with a 20 block is one

>of them

>

>About the paper, there is no mention of the number of block of master:

>

>"AlphaGo Master is the program that defeated top human players by 60–0

>in January, 2017 34 .

>It was previously unpublished but uses the same neural network

>architecture, reinforcement

>learning algorithm, and MCTS algorithm as described in this paper.

>However, it uses the

>same handcrafted features and rollouts as AlphaGo Lee

>and training was initialised by

>supervised learning from human data."

>

>Of what I understand same network architecture imply the same number of

>block

>

>Le 25/10/2017 à 17:58, Xavier Combelle a écrit :

>> I understand better

>>

>>

>> Le 25/10/2017 à 04:28, Hideki Kato a écrit :

>>> Are you thinking the 1st instance could reach Master level 

>>> if giving more training days?

>>>

>>> I don't think so.  The performance would be stopping 

>>> improving at 3 days.  If not, why they built the 2nd 

>>> instance?

>>>

>>> Best,

>>> Hideki

>>>

>>> Xavier Combelle: <05c04de1-59c4-8fcd-2dd1-094faabf3...@gmail.com>:

 How is it a fair comparison if there is only 3 days of training for 
>Zero ?

 Master had longer training no ? Moreover, Zero has bootstrap problem

 because at the opposite of Master it don't learn from expert games

 which means that it is likely to be weaker with little training.

 Le 24/10/2017 à 20:20, Hideki Kato a écrit :

> David Silver told Master used 40 layers network in May. 

> According to new paper, Master used the same architecture 

> as Zero.  So, Master used 20 blocks ResNet.  

> The first instance of Zero, 20 blocks ResNet version, is 

> weaker than Master (after 3 days training).  So, with the 

> same layers (a fair comparison) Zero is weaker than 

> Master.

> Hideki

 ___

 Computer-go mailing list

 Computer-go@computer-go.org

 http://computer-go.org/mailman/listinfo/computer-go

>

>

>___

>Computer-go mailing list

>Computer-go@computer-go.org

>http://computer-go.org/mailman/listinfo/computer-go
-- 
Hideki Kato 
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?

[Xavier Combelle]

>> The reason why (b) had became unpopular is because there is no go theory
>> precise enough to implement it as an algorithm
>
> There is quite some theory of the 95% principle kind which might be
> implemented as approximation. E.g. "Usually, defend your weak
> important group." can be approximated by approximating "group",
> "important" (its loss is too large in a quick positional judgement),
> "weak" (can be killed in two successive moves), "defend" (after the
> move, cannot be killed in two successive moves), "usually" (always,
> unless there are several such groups and some must be chosen, say,
> randomly; the approximation being that the alternative strategy of
> large scale exchange is discarded).
>
> Besides, one must prioritise principles to solve conflicting
> principles by a higher order principle.
>
> IMO, such an expert system combined with tree reading and maybe MCTS
> to emulate reading used when a principle depends on reading can, with
> an effort of a few manyears of implementation, already achieve amateur
> mid dan. Not high dan yet because high dans can choose advanced
> strategies, such as global exchange, and there are no good enough
> principles for that yet, which would also consider necessary side
> conditions related to influence, aji etc. I need to work out such
> principles during the following years. Currently, the state is that
> weaker principles have identified the major topics (influence, aji
> etc.) to be considered in fights but they must be refined to create
> 95%+ principles.
>
> ***
>
> In the 80s and 90s, expert systems failed to do better than ca. 5 kyu
> because principles were only marginally better than 50%. Today, (my)
> average principles discard the weaker, 50% principles and are ca. 75%.
> Tomorrow, the 75% principles can be discarded for an average of 95%
> principles. Expert systems get their chance again! Their major
> disadvantage remains: great manpower is required for implementation.
> The advantage is semantical understanding.
>
From a software developer point of view enlighten by my knowledge of
history of ai and history of go development,
 such approximate definition is close to useless to build a software at
the current state of art.
One of the reason is as you state the considerable work it would require
to implement a huge number of imprecise rules.
As you are not a software developer, I want you to look on this comics
which state the difference between apparent difficulty and real difficulty
of developping software. https://xkcd.com/1425/
As far as I understand your task to implement such an expert system
would require the many years of implementations would be thousands of years.
As far as my experience speak the expected reward would be a win of one
or two rank and so definitely not a mid dan amateur level.

Xavier Combelle

___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go