[Computer-go] it's alphago

[Ray Tayek]

https://games.slashdot.org/story/17/01/04/2022236/googles-alphago-ai-secretively-won-more-than-50-straight-games-against-worlds-top-go-players

--
Honesty is a very expensive gift. So, don't expect it from cheap people - Warren Buffett
http://tayek.com/

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

[Yamato]

Yes, it is AlphaGo. I am relieved that DeepMind clarified this.

Honestly I got a little frustrated that many people didn't think that
was AlphaGo. It was almost clear to me because I know the difficulty of
developing AlphaGo-like bots.

I hope Aja can comment here, also about GodMoves :)

Yamato

[Stefan Kaitschick]

Honestly I got a little frustrated that many people didn't think that
was AlphaGo. It was almost clear to me because I know the difficulty of
developing AlphaGo-like bots.

 I feel with you. People seem to think that the Nature paper gave away the full recipe.

[terry mcintyre]

It's one thing to know the recipe; it's another to have an industrial-size kitchen. Google was able to throw truly gargantuan amounts of computing resources at this problem.

A few years back, a researcher - was it  Remi Coulon? - was able to scrounge a few thousand cores for a tournament. Google designed an ASIC specifically for the task of accelerating their neural networks, and made thousands of them available for massive tests and training.

 

Terry McIntyre <terrym...@yahoo.com> Unix/Linux Systems Administration Taking time to do it right saves having to do it twice.tr

[Xavier Combelle]

Le 05/01/2017 à 02:16, Yamato a écrit :
> Yes, it is AlphaGo. I am relieved that DeepMind clarified this.
>
> Honestly I got a little frustrated that many people didn't think that
> was AlphaGo. It was almost clear to me because I know the difficulty of
> developing AlphaGo-like bots.

thanks for this insight, if I understand well developing a bot
competitive with alphago
is nearly an impossible task?

[Adrian Petrescu]

As an individual? Probably, yes.

[Lukas van de Wiel]

It helps a lot if you have to do it as a job, as a paid researcher. I once tried it as a volunteer job for a company I worked for at the time, but we only got the basic infrastructure going, after half a year of work, with two people.

We were trying a neural network approach, while everybody said NN was done for in AI go, and we had to do MCTS. We were stubborn, though. As we did not get paid for it, we essentially could do whatever we liked.

We also both had social lives that we did not want to neglect totally, so we might have been able to do more, had we been more dedicated. ;-)

[Xavier Combelle]

I mean as a company too, until this point none has succeed

[anoniem]

The sheer amount of processing power required is kinda frustrating..

Not able to use my computer for a whole month in order to train knowing 

it is only 1/100th the training time AlphaGo was trained with..

Yet it is extremely satisfying to see it grow stronger and surpass oneself..

I hope at some point DeepMind, an research instute, another big company
or somebody else with acces to a big cluster will release pre-trained models

for everybody to experiment with=)

a poor poor student;-)

[Yamato]

Hello,

On 2017/01/06 3:34, Xavier Combelle wrote:
>> Honestly I got a little frustrated that many people didn't think that
>> was AlphaGo. It was almost clear to me because I know the difficulty of
>> developing AlphaGo-like bots.
> thanks for this insight, if I understand well developing a bot
> competitive with alphago is nearly an impossible task?

It depends on which version of AlphaGo.

v13: possible
v18: very hard but not impossible
v19 or later: nearly impossible

in one year, I think.

One reason of this is that DeepMind has not published improvements
since the Nature version.

Yamato

[fot...@smart-games.com]

Competitive with Alpha-go, one developer, not possible. I do think it is possible to make a pro level program with one person or a small team. Look at Deep Zen and Aya for example. I expect I’ll get there (pro level) with Many Faces as well.

 

David

[Hiroshi Yamashita]

If value net is the most important part for over pro level, the problem is making strong selfplay games.

1. make 30 million selfplay games.
2. make value net.
3. use this value net for selfplay program.
4. go to (1)

I don't know when the progress will stop by this loop.
But if once strong enough selfplay games are published, everyone can make pro level program.
30 million is big number. It needs many computers.
Computer Go community may be able to share this work.
I can offer Aya, it is not open-source though. Maybe Ray(strongest open source so far) is better choice.

Thanks,
Hiroshi Yamashita

----- Original Message -----
From: <fot...@smart-games.com>
To: <compu...@computer-go.org>
Sent: Friday, January 06, 2017 4:50 PM
Subject: Re: [Computer-go] it's alphago

Competitive with Alpha-go, one developer, not possible. I do think it is possible to make a pro level program with one
person or a small team. Look at Deep Zen and Aya for example. I expect I’ll get there (pro level) with Many Faces as
well.

David

_______________________________________________

[Detlef Schmicker]

Hi,

this sounds interesting! AlphaGo paper plays only with RL network, if I
understood correctly. If we start this huge approach we should try to
carefully discuss the way (and hopefully get some hints from people
tried with much computational power :)

If I understood correctly you would try to use a program using value net
with (let's say 2000 playouts) in selfplay? Using only one result, or
doing some games per position? Or are you thinking of using only the win
percentage such a program gives from his own mixing of SL network,
search and value net?

By the way to make some promotion :) oakfoam is not far away from Ray
for this kind of approach, where you will probably try to reduce cpu/gpu
usage per game (at least if Rn3.3-4c is Ray on CGOS with 4 cores, NG04b
is oakfoam on CGOS with 10k and saving GPU usage by using only 50% of GX970)

Detlef

[Sebastian Scheib]

Maybe now we need AlphaGo vs. Tartrate to see who is the definitive Sai XD

--

Dracux

http://www.dracux.com

[Andy]

What is Ray? Strongest open source bot? Anyone have a link to it?

[Lukas van de Wiel]

A project similar to the Great Mersenne Prime search might be a
possibility to distribute the work of training the network among many
enthousiasts, and to keep improving it by self play.

[daniel rich]

A closer example than the mersenne prime search is fishtest from the chess engine world. My understanding is that it is a key part of why stockfish is such a strong chessengine.

https://github.com/glinscott/fishtest

A large group of volunteers that essentially donate compute power to test changes and improve the bot. That would be a fairly cool way compute time to be made available to the community. The plus is that eventually big corporate players may lose interest to devote the same level of spending and compute that we have seen so far.

[Marc Landgraf]

And why would it be desirable that 'the big corporate players lose interest to devote computer power'?

And who are those big corporate players? Deepmind? Who are not even selling their bot? Or are you talking about CS/Zen who are having indeed financial interests here?

What would be the benefit of any of those parties in losing interest?

[daniel rich]

Oh sorry I mispoke, corporate players losing interest is a bad thing in my mind but also more or less inevitable(to some degree anyway). I was simply saying that as delighted as I am that google and other players are putting so much money and research into go I suspect eventually the resources will be re-allocated to other things, and having a large community network would help push things forward even after/if resources dwindle.

Just as you point out deepmind is not selling their bot and I think aren't motivated as much by the game of go so much as the AI breakthroughs it represents. Even if many others players stay fairly active I think the resources currently invested in go are larger than the market for go bots. 

For example in the chess world there are some valuable chess engines but the engines are now limited mostly to companies that directly sell the engines or community/non-profit efforts. I suspect at some point go will be the same.

[Hiroshi Yamashita]

Ray was Japanese student program that went on 7th, UEC cup 2016.

Ray
http://computer-go-ray.com/eng/index.html
Thare is a stronger version of Ray, with policy net and value net.
https://github.com/zakki/Ray/tree/nn
CGOS BayesElo is 3463 (Rn.3.3-4c).
http://www.yss-aya.com/cgos/19x19/bayes.html

Hiroshi Yamashita

----- Original Message -----
From: "Andy" <andy.o...@gmail.com>
To: "computer-go" <compu...@computer-go.org>
Sent: Friday, January 06, 2017 11:48 PM
Subject: Re: [Computer-go] it's alphago

What is Ray? Strongest open source bot? Anyone have a link to it?

[Hiroshi Yamashita]

Hi,

> If I understood correctly you would try to use a program using value net
> with (let's say 2000 playouts) in selfplay? Using only one result, or

Yes. 2000 playouts/move MCTS with policy net and value net.

> doing some games per position? Or are you thinking of using only the win

I thought one game per position, but some games per position looks nice option.

> doing some games per position? Or are you thinking of using only the win
> percentage such a program gives from his own mixing of SL network,

In my experience, game result is better than win percentage.

> usage per game (at least if Rn3.3-4c is Ray on CGOS with 4 cores, NG04b

Oh NG04b is oakfoam. AlphaGo RL is about 2800 (CGOS BayesElo).
So around this rating program seems nice. And many computers don't have GPU.
To calculate DCNN on CPU, maybe we can not use big network(filter 192), but
smaller one(filter 64 or 32).

[Xavier Combelle]

To my knowledge, fishtest is also a major part of stockfish engine. It is essential because there is lot of possible improvement and most of them win only 2 or 3 elo points, but added, it lead to 60-70 elo points between each release (every one year or something like that)

[Darren Cook]

> https://games.slashdot.org/story/17/01/04/2022236/googles-alphago-ai-secretively-won-more-than-50-straight-games-against-worlds-top-go-players

The five Lee Sedol games last year never felt like they were probing
Alpha Go's potential weaknesses. E.g. things like whole board semeai,
complex whole board ko fights, obscure under-the-stones tesuji, etc.

I wondered if anyone here had studied those 50 games and found anything
interesting or impressive, along those lines? I.e. if I was going to
look at just one game, which one should it be?

Thanks,
Darren

--
Darren Cook, Software Researcher/Developer
My New Book: Practical Machine Learning with H2O:
http://shop.oreilly.com/product/0636920053170.do