Alpha Zero (redux)

Discuss anything you like about chess related matters in this forum.
User avatar
Christopher Kreuzer
Posts: 7277
Joined: Fri Aug 06, 2010 2:34 am
Location: London

Alpha Zero (redux)

Post by Christopher Kreuzer » Thu Feb 01, 2018 1:20 pm

The previous thread on Alpha Zero got merged with the earlier thread on Computer Go, which is silly as no-one is discussing computer go any more. Here are some more links actually on Alpha Zero:

http://chessimprover.com/its-all-over-for-humanity/
http://chessimprover.com/news-of-the-century/

There is a You Tube video by someone called agadmator analysing AlphaZero's "Immortal Zugzwang Game".



Does anyone know of a text-based or website-based analysis of that game, rather than a video?

And why hasn't there been more discussion of AlphaZero here? Is it the end of chess? What does it all mean for the different flavours of human chess and the numerous communities of chess players?

MJMcCready
Posts: 1234
Joined: Mon Jun 24, 2013 2:30 pm

Re: Alpha Zero (redux)

Post by MJMcCready » Thu Feb 01, 2018 2:25 pm

The whole thing has been over-hyped. People will continue to enjoy chess for hundreds of years. The fact that a new computer has suddenly got better than the existing computers is of no consequence for most people.

MartinCarpenter
Posts: 2434
Joined: Tue May 24, 2011 10:58 am

Re: Alpha Zero (redux)

Post by MartinCarpenter » Thu Feb 01, 2018 2:37 pm

It means a tiny bit more than that - it means that the entire sum of human knowledge gathered about chess in all the time we've been playing it is worth less than nothing. Identically for Go.

If that doesn't feel like a bit of a slap in the face then I'm not sure why :)

Meanwhile of course, we'll keep happily playing the game.

User avatar
JustinHorton
Posts: 6340
Joined: Mon Aug 04, 2008 10:06 am
Location: Somewhere you're not

Re: Alpha Zero (redux)

Post by JustinHorton » Thu Feb 01, 2018 2:40 pm

MartinCarpenter wrote:
Thu Feb 01, 2018 2:37 pm
It means a tiny bit more than that - it means that the entire sum of human knowledge gathered about chess in all the time we've been playing it is worth less than nothing.
You sure about that?
"Do you play chess?"
"Yes, but I prefer a game with a better chance of cheating."

lostontime.blogspot.com

MJMcCready
Posts: 1234
Joined: Mon Jun 24, 2013 2:30 pm

Re: Alpha Zero (redux)

Post by MJMcCready » Thu Feb 01, 2018 3:02 pm

No he isn't, he's just regurgitating some woeful journalism without thinking about what he's said.

MartinCarpenter
Posts: 2434
Joined: Tue May 24, 2011 10:58 am

Re: Alpha Zero (redux)

Post by MartinCarpenter » Thu Feb 01, 2018 9:21 pm

No, its what I mean :) Ok, the information is worth a lot *to us* and every other chess engine.

To the problem of how best to play chess? It turned out to be worth less than nothing.

They notably improved at Go when they stopped trying to use the existing knowledge. Essentially there are enough incorrect simplifications/assumptions built into the extant human knowledge base that it was better to have the computer reproduce it 'right'.
(Or arguably simply in a more suitable form for how that computer engine thinks.).

MJMcCready
Posts: 1234
Joined: Mon Jun 24, 2013 2:30 pm

Re: Alpha Zero (redux)

Post by MJMcCready » Thu Feb 01, 2018 10:15 pm

Understood as a cultural phenomenon, chess has lots to offer on many levels. What computers do and don't do is of little consequence to all below the upper echelon of our game.

Roger de Coverly
Posts: 18056
Joined: Tue Apr 15, 2008 2:51 pm

Re: Alpha Zero (redux)

Post by Roger de Coverly » Thu Feb 01, 2018 10:42 pm

MartinCarpenter wrote:
Thu Feb 01, 2018 9:21 pm
They notably improved at Go when they stopped trying to use the existing knowledge.
On the limited basis of what the Alpha Zero team published, they confirmed or perhaps reinforced existing knowledge on the d5 sacrifice in the Queens Indian and the lines with e5 and a slow treatment in the Caro (usually termed the Short system).

It would be fascinating if they let Alpha Zero or a clone running in a similar manner loose on the position after 1. e4 e5 2. f4 . Is it any good for White? The subsidiary question being what are the most promising lines for both players?

The opinions of the existing engines are biased by the weights placed by their programmers on various factors.

MJMcCready
Posts: 1234
Joined: Mon Jun 24, 2013 2:30 pm

Re: Alpha Zero (redux)

Post by MJMcCready » Fri Feb 02, 2018 3:17 pm

If your rating is below ELO 2700 it will make little or no difference. As soon as you deviate from what it uncovers, mistake after mistake after mistake will inevitably follow. All that will come of it is that you will enjoy your games less because you are tying to follow something rather than adhere to your own preferences.

Matthew Turner
Posts: 2701
Joined: Fri May 16, 2008 11:54 am

Re: Alpha Zero (redux)

Post by Matthew Turner » Fri Feb 02, 2018 10:56 pm

Alan Turing created a ‘computer’, Turochamp; The fact that it could play chess at a reasonable level was both surprising and revolutionary. That has basically formed the basis of all computing ever since. Alpha Go and Alpha Zero have a new approach which is undoubtedly much more successful at winning games. It is easy then to draw the inference that the ideas behind AI will dominate computing in the same way as Turing’s algorithms did. A computer that is much better at Chess, or Go, must also be better equipped for dealing with the logistics of the NHS. That involves a bigger leap in logic than many people appreciate, so the jury is still out on the eventual impact of AI. However, if we consider what has stemmed from Turing’s innovation the possibilities are difficult to comprehend.

User avatar
Joey Stewart
Posts: 1082
Joined: Wed Apr 11, 2007 2:35 pm
Location: All Of Them
Contact:

Re: Alpha Zero (redux)

Post by Joey Stewart » Sat Feb 03, 2018 4:56 pm

Right now the only people who will care about these super computers "solving" chess are correspondence players, unless the future brings some sort of technology which allows us to interface our brains with machines then there is no worry about anybody human being able to use this to their advantage in competitive play.
Lose one queen and it is a disaster, Lose 1000 queens and it is just a statistic.

NickFaulks
Posts: 5043
Joined: Sat Jan 02, 2010 1:28 pm

Re: Alpha Zero (redux)

Post by NickFaulks » Sun Feb 04, 2018 11:38 am

Matthew Turner wrote:
Fri Feb 02, 2018 10:56 pm
A computer that is much better at Chess, or Go, must also be better equipped for dealing with the logistics of the NHS. That involves a bigger leap in logic than many people appreciate, so the jury is still out on the eventual impact of AI.
The crucial leap in logic is clear. Alpha Zero is based entirely on learning from experience, at which it is astonishingly good. It is a basic tenet of the political class that under no circumstances should any lessons be learned from past experience - the only way forward is to repeat failed policies with redoubled enthusiasm.

Jonathan Rogers
Posts: 3852
Joined: Tue Nov 18, 2008 9:26 pm

Re: Alpha Zero (redux)

Post by Jonathan Rogers » Sun Feb 04, 2018 2:42 pm

NickFaulks wrote:
Sun Feb 04, 2018 11:38 am
Matthew Turner wrote:
Fri Feb 02, 2018 10:56 pm
A computer that is much better at Chess, or Go, must also be better equipped for dealing with the logistics of the NHS. That involves a bigger leap in logic than many people appreciate, so the jury is still out on the eventual impact of AI.
The crucial leap in logic is clear. Alpha Zero is based entirely on learning from experience, at which it is astonishingly good. It is a basic tenet of the political class that under no circumstances should any lessons be learned from past experience - the only way forward is to repeat failed policies with redoubled enthusiasm.
I might agree with that but wonder whether Nick and I have the same examples in mind ...

I wonder how helpful it is to start reimagining the merits of major openings on the basis of the games so far. If Alphazero plays a bit more, maybe it will independently start to share the preferences forged by human experience.

David Robertson
Posts: 2143
Joined: Tue Apr 03, 2007 6:24 pm
Contact:

Re: Alpha Zero (redux)

Post by David Robertson » Sun Feb 04, 2018 3:40 pm

Jonathan Rogers wrote:
Sun Feb 04, 2018 2:42 pm
maybe it will independently start to share the preferences forged by human experience
Now that would be a breakthrough - computer says humans are right!

NickFaulks
Posts: 5043
Joined: Sat Jan 02, 2010 1:28 pm

Re: Alpha Zero (redux)

Post by NickFaulks » Sun Feb 04, 2018 4:13 pm

Jonathan Rogers wrote:
Sun Feb 04, 2018 2:42 pm
If Alphazero plays a bit more, maybe it will independently start to share the preferences forged by human experience.
My understanding is that it reached that point in its first run, and carried on.

Post Reply