Page 3 of 3

Re: Alpha Zero (redux)

Posted: Fri Jul 12, 2019 4:36 pm
by Paul McKeown
NickFaulks wrote:
Sun Feb 04, 2018 11:38 am
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.
A perfect description of the Tory's current leadership election: all objective knowledge about Brexit must be ignored in favour of the selectorate's private prejudices.

Re: Alpha Zero (redux)

Posted: Fri Jul 12, 2019 4:41 pm
by Christopher Kreuzer
Noticed a reference (from January 2019) to DeepMind turning to computer gaming:

https://towardsdatascience.com/the-scie ... 4bd7824d4b

Plenty of chess comparisons in the article.
I must confess to never having heard of StarCraft (or StarCraft II) before. Anyone?

Re: Alpha Zero (redux)

Posted: Fri Jul 12, 2019 4:50 pm
by IM Jack Rudd
I've heard of it. I think John Sargent plays it a lot.

Re: Alpha Zero (redux)

Posted: Fri Jul 12, 2019 4:55 pm
by Paul McKeown
I remember some nonsense on this forum and elsewhere in the past, about how little hardware was used, some of it based on protests by Deepmind that the Google Tensorflow was just GPU class hardware, not supercomputers. Yet here is Google, itself, talking about the new third generation of TPUs:
The latest-generation Cloud TPU v3 Pods are liquid-cooled for maximum performance. Each one delivers more than 100 petaFLOPs of computing power. In terms of raw mathematical operations per second, a Cloud TPU v3 Pod is comparable with a Top 5 supercomputer worldwide, Google said -- though it operates at lower numerical precision.
https://www.zdnet.com/article/google-ma ... e-in-beta/

The hardware that Deepmind was trained on (a cluster of 5,000 TPU 1st Gens to generate games and 64 TPU 2nd Gens for the neural network) was absolutely of supercomputer class. It would be interesting to see the result from training it for 24 hours on a TPU3 network, rather than a TPU2 network. It is possible that no improvement might be reached due to a very deep, pseudo-stable local minimum having been reached. On the other hand...

Re: Alpha Zero (redux)

Posted: Wed Aug 14, 2019 12:05 pm
by Christopher Kreuzer
May be of interest:

Fat Fritz – What on Earth is that?
It's a semi-secret development, an AlphaZero clone, engineered over the past nine months for ChessBase. Fat Fritz was tested by some of the best players in the world, who expressed unmitigated delight over the ideas and improvements it came up with. Now the program is publicly available on the ChessBase Engine Cloud. And it is running on awesome hardware.

Re: Alpha Zero (redux)

Posted: Thu Aug 15, 2019 10:59 am
by Geoff Chandler
Hi Chris,

This bit:

" Fat Fritz was tested by some of the best players in the world, who expressed
unmitigated delight over the ideas and improvements it came up with."

Was followed by:

"A number of top players will probably not allow us to reveal what they had found — the positions and moves they shared with us.
"I cannot let you use them in your articles," one of them typically wrote. "They haven't yet been used in practice, and for obvious
reasons I want to reserve them for future games, to score points."

A number of them did not even want us to tell people that they had been working with the machine."

Who ever wrote this advert knows their market well. Chess players and paranoia go hand in hand.

Every future opponent may have been one of the chosen few. What have the found? what TN bomb is ticking?
Who has been posted missing for 2-3 weeks. I bet they are one of the 'Nameless Ones'. Why was I not picked?
My entire repertoire has been examined, exposed and I'm a sitting duck whilst they have 'unmitigated delight.'

Within days of reading the article buying a Fat Fritz will be the overriding priority of every sleepless chess player.

Think I'll write an opening book with the blurb:

The variations inside have been shown to some of the best players in the world who have requested to remain
nameless so they can use the revolutionary moves inside to score points. (...and points mean prizes.)

'Don't tell people you have this book which can only be bought in cash leaving no paper trail to preserve your anonymity
if you use a move from this book say you found it OTB because the fewer that know of this book the better it is for you.

Re: Alpha Zero (redux)

Posted: Thu Aug 15, 2019 1:39 pm
by John McKenna
I think 'Fat Fritz' is a truly terrible product name.

It puts me in mind of Landburgher Gessler in this old TV series -

https://en.m.wikipedia.org/wiki/The_Adv ... lliam_Tell

Why not rename it - William Tell!?

Edit - The answer is in the theme song -

... Follow on, follow on, at the leader's heel
With a thrust of a pike and a clash of steel
Follow on with the fight till the tyrants reel
For Tell and Switzerland.

Give ‘em one for the day they burned the grain,
Give ‘em two for the night that Fritz was slain,
Give ‘em three, give ‘em four and hooray for more,
For Tell and Switzerland....

Re: Alpha Zero (redux)

Posted: Thu Aug 15, 2019 4:52 pm
by Geoff Chandler
Hi John,

I have an old ChessBase C.D. knocking about somewhere called FatBase.
Maybe choosing the name is a hang up from that.

Image

Why not call it Robin Hood.

Robin Hood, Robbing you Robbing all your friends,
Robin Hood, Robbing you it's become a chessbase trend.

Feared by the good, welcomed by the cheats.

Robin Hood....Robbing You...Robbing you...

Re: Alpha Zero (redux)

Posted: Thu Aug 15, 2019 6:11 pm
by John McKenna
Continuing that theme, but to another tune -

Robbin' you, robbin' me, (Ah-haa)

Where's the fat lady who can hit high C?

A more suitable name for the new product could be -

Deep Fryer Tuck!

Re: Alpha Zero (redux)

Posted: Mon Nov 04, 2019 5:58 pm
by Christopher Kreuzer
More on Fat Fritz here:

What sets Fat Fritz apart (part 1)

Re: Alpha Zero (redux)

Posted: Tue Dec 01, 2020 12:36 am
by Christopher Kreuzer
Christopher Kreuzer wrote:
Fri Jul 12, 2019 4:41 pm
Noticed a reference (from January 2019) to DeepMind turning to computer gaming:

https://towardsdatascience.com/the-scie ... 4bd7824d4b
Coming back to this topic:

https://www.bbc.co.uk/news/science-environment-55133972

"One of biology's biggest mysteries 'largely solved' by AI"

It's about protein folding.
London-based AI lab, DeepMind, has largely cracked the problem, say the organisers of a scientific challenge.
The article also has a link to "DeepMind's Demis Hassabis on The Life Scientific (BBC Radio 4)":

https://www.bbc.co.uk/programmes/m0009zbj

[Apologies if that has been mentioned before.]

Re: Alpha Zero (redux)

Posted: Tue Dec 01, 2020 1:10 am
by David Sedgwick

Re: Alpha Zero (redux)

Posted: Tue Dec 01, 2020 10:16 am
by MartinCarpenter
That really is a bit astonishing, and it'll be fascinating to see how they did it when it comes out.

Its a conceptually much less obvious problem for it than chess/Go etc.

Re: Alpha Zero (redux)

Posted: Sun Dec 27, 2020 7:54 pm
by Christopher Kreuzer
In case anyone missed it (and as far as I can see it wasn't mentioned in other threads):

DeepMind's AI agent MuZero could turbocharge YouTube (BBC article)

MuZero (Wikipedia article)

Only a couple of mentions of chess, but from the BBC article:
The Nature paper reports that MuZero proved to be slightly better than AlphaZero at playing Go, despite doing less tree-search computation per move.
...other potential uses included next-generation virtual assistants, personalised medicine and search-and-rescue technologies.