We are doomed - AlphaGo Zero, learning only from basic rules

Discussion of anything and everything relating to chess playing software and machines.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Daniel Shawul
Posts: 3974
Joined: Tue Mar 14, 2006 10:34 am
Location: Ethiopia
Contact:

Re: We are doomed - AlphaGo Zero, learning only from basic r

Post by Daniel Shawul » Thu Oct 19, 2017 1:48 pm

Vinvin wrote:I hope for such an experience for chess : starting a very deep learning with only basic rules, piece coordinate and piece interaction.
Even no "piece value" concept hardcoded.
Giraffe already did that with the NN evaluation. It used hand-selected features as inputs (presence & location of pieces) etc, and was able to compete with Stockfish's eval. It is most likely possible to have a Giraffe-zero atleast for the evaluation only --i.e. it will learn everything the chess world knows about good static evaluation (not search) from self-play games only in a couple of hours.

User avatar
reflectionofpower
Posts: 1461
Joined: Fri Mar 01, 2013 4:28 pm
Location: USA

Re: We are doomed - AlphaGo Zero, learning only from basic r

Post by reflectionofpower » Thu Oct 19, 2017 2:04 pm

Vinvin wrote:
reflectionofpower wrote:It would be interesting for chess to be done this way BUT how will it benefit the average chess user if we do not have the "juice" to run the program?
When the neural network is trained, it can be exported in a file to be use on every computer.

Like a BrainFish concept? Highly doubtful ... If that was the case we could use that file and run the chess program on a 286. The only way the AlphaGO program was so successful was the neural network ALONG with the juice (their optimal setup was 1202 CPUs and 176 GPU). I did not get the specs for the AlphaGo Zero configuration.
reflectionofpower wrote:The programs of today on an average machine can beat the best of humanity.

Yes, but sometimes top engines do very poorly on some positions. The idea is to get an engine strong in EVERY chess positions !
I agree for the most part but humans do very poorly compared to (C) on tactics on a large majority of the times so does it really matter?
"Without change, something sleeps inside us, and seldom awakens. The sleeper must awaken." (Dune - 1984)

Lonnie

Leo
Posts: 927
Joined: Fri Sep 16, 2016 4:55 pm
Location: USA/Minnesota
Full name: Leo Anger

Re: We are doomed - AlphaGo Zero, learning only from basic r

Post by Leo » Thu Oct 19, 2017 2:38 pm

Cardoso wrote:For that one I think they will take milleniums, or simply never.
They can't cure migranes or diabetes, much less stop aging.
And I think some people expect too much from science.
My mother had a severe skin disease on her feet called hyperkeratosis, with profound cracks in the sckin wich hurted badly, she was treated with the best doctor in the field in the country (portugal), with very agressive medications, and none of the several treatements worked. Desperate my mother tryed some plant called "malvas" in portuguese, after 2 weeks she was much better, after 6 weeks she was completed cured and the problem went completely away.
I think the human body is too complex for science.
Even a single cell is tremendously complex.
I very much agree. I am tired of the bombast from Google on what great things they are going to do for humanity. AI is overrated.
Advanced Micro Devices fan.

User avatar
reflectionofpower
Posts: 1461
Joined: Fri Mar 01, 2013 4:28 pm
Location: USA

Re: We are doomed - AlphaGo Zero, learning only from basic r

Post by reflectionofpower » Thu Oct 19, 2017 3:05 pm

Leo wrote:
Cardoso wrote:For that one I think they will take milleniums, or simply never.
They can't cure migranes or diabetes, much less stop aging.
And I think some people expect too much from science.
My mother had a severe skin disease on her feet called hyperkeratosis, with profound cracks in the sckin wich hurted badly, she was treated with the best doctor in the field in the country (portugal), with very agressive medications, and none of the several treatements worked. Desperate my mother tryed some plant called "malvas" in portuguese, after 2 weeks she was much better, after 6 weeks she was completed cured and the problem went completely away.
I think the human body is too complex for science.
Even a single cell is tremendously complex.
I very much agree. I am tired of the bombast from Google on what great things they are going to do for humanity. AI is overrated.
Yep! NONE of these companies are here to save humanity. They are in it to make $$ regardless if human bodies or land are in the way. Used to be decades ago health insurance was included in your job AND loyalty was rewarded and most of the time you retired from a company with a pension. Now you pay for your insurance and if there is a way to save $$ they will fire people left & right to accomplish it.
"Without change, something sleeps inside us, and seldom awakens. The sleeper must awaken." (Dune - 1984)

Lonnie

Uri Blass
Posts: 8771
Joined: Wed Mar 08, 2006 11:37 pm
Location: Tel-Aviv Israel

Re: We are doomed - AlphaGo Zero, learning only from basic r

Post by Uri Blass » Thu Oct 19, 2017 3:20 pm

Daniel Shawul wrote:
Vinvin wrote:I hope for such an experience for chess : starting a very deep learning with only basic rules, piece coordinate and piece interaction.
Even no "piece value" concept hardcoded.
Giraffe already did that with the NN evaluation. It used hand-selected features as inputs (presence & location of pieces) etc, and was able to compete with Stockfish's eval. It is most likely possible to have a Giraffe-zero atleast for the evaluation only --i.e. it will learn everything the chess world knows about good static evaluation (not search) from self-play games only in a couple of hours.
Is there an evidence that Giraffe was able to compete with stockfish's evaluation?

I do not claim that it is not the case but I do not know about comparison between quality of evaluation of Giraffe and Stockfish.

PK
Posts: 861
Joined: Mon Jan 15, 2007 10:23 am
Location: Warsza
Contact:

Re: We are doomed - AlphaGo Zero, learning only from basic r

Post by PK » Thu Oct 19, 2017 3:49 pm

The only direct comparison of Giraffe eval with a decently strong engine that I am aware of:

http://talkchess.com/forum/viewtopic.ph ... 39&t=64096

Daniel Shawul
Posts: 3974
Joined: Tue Mar 14, 2006 10:34 am
Location: Ethiopia
Contact:

Re: We are doomed - AlphaGo Zero, learning only from basic r

Post by Daniel Shawul » Thu Oct 19, 2017 4:15 pm

PK wrote:The only direct comparison of Giraffe eval with a decently strong engine that I am aware of:

http://talkchess.com/forum/viewtopic.ph ... 39&t=64096
I wan't actually aware of that test but the one given in Giraffe's paper on Page 25: https://arxiv.org/pdf/1509.01549v1.pdf

Peter's test also confirm that Giraffee's eval is close to Stockfish's, but it is not equally efficient due to 10x slowdown incurred by the NN evaluation. So it seems Giraffe has
already learned (probably not tabula rasa ? ) all the human chess knowledge of the past century...

Code: Select all

Giraffe (1s) 2400 258570 9641
Giraffe (0.5s) 2400 119843 9211
Giraffe (0.1s) 2400 24134 8526
Stockfish 5 3387 108540 10505
Senpai 1.0 3096 86711 9414
Texel 1.04 2995 119455 8494
Arasan 17.5 2847 79442 7961
Scorpio 2.7.6 2821 139143 8795
Crafty 24.0 2801 296918 8541
GNU Chess 6 / Fruit 2.1 2685 58552 8307
Sungorus 1.4 2309 145069 7729

jdart
Posts: 3951
Joined: Fri Mar 10, 2006 4:23 am
Location: http://www.arasanchess.org

Re: We are doomed - AlphaGo Zero, learning only from basic r

Post by jdart » Thu Oct 19, 2017 4:24 pm

The problem is search speed.

You can have a Stockfish-equivalent eval implemented in a NN, but the eval speed will not be the same (likely worse).

--Jon

jhellis3
Posts: 460
Joined: Fri Aug 16, 2013 10:36 pm

Re: We are doomed - AlphaGo Zero, learning only from basic r

Post by jhellis3 » Thu Oct 19, 2017 4:32 pm

I would say it depends on how the NNs scale. If they have been designed to run on Googles TPUs or Nvidia's Volta+ chips, the raw performance available will be monstrous. If that is the case, I would expect they could be ported to Chess, trained up, and be 200+ Elo above the top chess engines today within a few months. Both amazing and kind of sad at the same time...

gladius
Posts: 538
Joined: Tue Dec 12, 2006 9:10 am

Re: We are doomed - AlphaGo Zero, learning only from basic r

Post by gladius » Fri Oct 20, 2017 3:34 am

jhellis3 wrote:I would say it depends on how the NNs scale. If they have been designed to run on Googles TPUs or Nvidia's Volta+ chips, the raw performance available will be monstrous. If that is the case, I would expect they could be ported to Chess, trained up, and be 200+ Elo above the top chess engines today within a few months. Both amazing and kind of sad at the same time...
That would be incredible to me, especially if it was all self trained. I was thinking how to structure the network in a similar fashion for Chess. It's not trivial, as there is not a well defined move output like in Go (fixed number of places you can place a stone, so the network just always outputs a probability for each of those).

You could do a deterministic legal move generator, and use the index of the move in that generator as the set of probabilities the network predicts, but that's a bit ugly. You could also do the more general thing, and have source to destination square output. That would 64x63 outputs (actually a little less, as possible moves from each square are limited). I'm not sure if that's too many for the network to learn though.

A pretty revolutionary advance!

Post Reply