Page 10 of 11
Re: AlphaZero
Posted: Thu Dec 07, 2017 1:28 pm
by Roger de Coverly
We've been discussing this under a thread titled Computer Go.
The computer used is something home brewed by Google. It's massively powerful.
viewtopic.php?f=36&t=8097&start=120
Re: Computer Go
Posted: Thu Dec 07, 2017 1:32 pm
by Roger de Coverly
Peter Shaw wrote: ↑Thu Dec 07, 2017 12:30 pm
Of course I don't really understand this, but surely it would take billions of games before it would even start to develop any simple strategy such as 'it's a generally a good idea to take the opponent's pieces'.
I'd think it's a similar program to one they got to play arcade games. I'd imagine it starts with some rudimentary strategy like taking opponent's pieces and retaining its own. They might even have given it a nudge by asking it to solve some simple mates.
Re: AlphaZero
Posted: Thu Dec 07, 2017 1:33 pm
by Clive Blackburn
Re: Computer Go
Posted: Thu Dec 07, 2017 2:10 pm
by Matt Fletcher
Peter, I think it did start playing random moves, but I'm not convinced that this is drawish. Agree it would be interesting to see some early games if they're stored!
I also think they do mention the number of games per step - the report says 'mini-batches of 4,096'. So if my maths is right, it self-played almost 3 billion games..!
Roger, I think the whole point was that they just told it the rules and set it running - no training on positions, just self-play and optimisation.
Re: Computer Go
Posted: Thu Dec 07, 2017 2:21 pm
by MartinCarpenter
There are really good fundamental reasons to set it off from scratch too - once you program anything into it beyond the rules you have to fix on some sort of abstractions to definite it all with. Get those wrong and you can really limit its ultimate strength.
A bit like how using existing Go games to train AlphaGo left it with some ultimately unhelpful human derived hang ups....
Deep mind are showing fairly strongly it works better to let the computer pick these sorts of things out itself
Nothing left for us to do but worshipping our silicon overlords
The other thing about the training is that I presume that those 3 billion games won't have all been purely random by any means - the first X% will have been, later on it'll have been a mixture of 'intelligent' and random moves and near the end it'll have been playing nearly all intelligent moves.
Re: Computer Go
Posted: Mon Dec 11, 2017 8:31 am
by Paul McKeown
I have just gone through the 10 published games from the match. Alpha Zero is amazing. It does not play like any computer program I have ever seen before. It seems to me more a superhuman initiative seeker, with a real sense of just going for it. Tal on steroids. The depth at which its material sacrifices are seen to be justified is staggering. Beautiful stuff. Of course, this is preliminary, as there are the other hundred odd unpublished games, which may ultimately give a different impression. I'm not sure who it was selected the games to be published, but do you get the feeling they liked seeing the Queen's Indian getting tin canned? Chuckle.
Re: AlphaZero
Posted: Wed Dec 13, 2017 5:23 pm
by Christopher Kreuzer
Chessbase article:
https://en.chessbase.com/post/alpha-zer ... and-apples
(There are other articles out there, if anyone wants to add other links.)
Re: AlphaZero
Posted: Wed Dec 13, 2017 7:29 pm
by Paul McKeown
It might be useful it the other Alpha Go thread were moved into the main forum, as it now concerns chess.
Re: AlphaZero
Posted: Wed Dec 13, 2017 8:41 pm
by Carl Hibbard
Paul McKeown wrote: ↑Wed Dec 13, 2017 7:29 pm
It might be useful it the other Alpha Go thread were moved into the main forum, as it now concerns chess.
The topics where combined.
This chap is a little loud but it was interesting:
https://www.youtube.com/watch?v=9CoNk3EYOpc&t=1032s
Re: Computer Go
Posted: Wed Dec 13, 2017 9:18 pm
by Paul McKeown
Thanks!
Re: Computer Go
Posted: Thu Dec 14, 2017 10:25 pm
by Christopher Kreuzer
We do really still need an AlphaZero thread that is separate from the AlphaGo thread, surely?
Just played through all the 10 sample wins by AlphaZero. Difficult to understand some of it, but some scarily impressive play and mix of tactics, positional play, and (to me) incomprehensible moves.
Re: AlphaZero
Posted: Mon Dec 18, 2017 10:38 am
by Mick Norris
Re: Computer Go
Posted: Mon Jan 01, 2018 1:07 pm
by Carl Hibbard
For those of you with Netflix a documentary on AlphaGo appeared today and is worth a watch.
Re: Computer Go
Posted: Mon Jan 01, 2018 1:09 pm
by Carl Hibbard
Carl Hibbard wrote: ↑Mon Jan 01, 2018 1:07 pm
For those of you with Netflix a documentary on AlphaGo appeared today and is worth a watch I am told.
I presume just on the Go side.
Re: Computer Go
Posted: Mon Jan 01, 2018 3:29 pm
by Carl Hibbard
Carl Hibbard wrote: ↑Mon Jan 01, 2018 1:07 pm
For those of you with Netflix a documentary on AlphaGo appeared today and is worth a watch.
Watched and enjoyed although I would have liked a little more tech details like number of nodes etc otherwise well worth the time.