News from AlphaZero

[shrapnel]

Nice to see the comparison with Sultan Khan, an Indian, who was a true chess Genius.

[shrapnel]

Well, it’s an article/blog written off the backof a PR release from a publisher who doesn’t yet have the book, which is not yet actually written by some chess guys who apparently get to see some hundred of games that nobody else has seen, selectively provided them by Google DeepMind, using not known criteria. And some quotes about something more in that classic forward moving window “in a couple of weeks”.

Excuse me for being an old cynic.

Sneer all you like, but the fact remains that AlphaZero's very deep Positional understanding cannot be matched by Stockfish or any other chess engine no matter how much they progress, as their very basis is a fallacy. Most chess programmers made the basic mistake of trying to reduce Chess to just a set of mathematical/scientific principles, which it is NOT.
As is said, Chess is not just a Science, it is also an Art. Programmers ignore the Art aspect at their own peril or perhaps they have no choice as they cannot quantify Art.
Whatever the reason, as long as they see Chess in terms of Piece/Material value, they can never hope to master Chess completely.
This aspect is confirmed by the Article itself, where its says that even the Programmers of AlphaZero are THEMSELVES unsure how AlphaZero goes about its business, or words to that effect.
You can quibble about 'unknown criteria' all you like, but you are still missing the Woods for the Trees.
The detractors are so obsessed by the Conditions under which the Match is being played that they fail to see ( or choose not to see) the MANNER in which AlphaZero decimates Stockfish. Stockfish appears to be so completely clueless as to the long-term intentions of AlphaZero, that it hardly seems to matter what Hardware or Time Control is being used.
Another interesting Term I've seen used w.r.t. AlphaZero is the "Boa-Constrictor approach", that is, it seems to be more interested in anticipating and stifling the Play of the opponent rather than playing brilliant moves itself.
I've seen this approach during my Club playing days against human opponents ; of course AlphaZero has the advantage of not making any Tactical errors as humans would make, which makes it more deadly.
This is explained better in this Article....
https://www.nextbigfuture.com/2018/01/c ... style.html.
To quote..."Modern chess engines are focused on activity, and have special safeguards to avoid blocked positions as they have no understanding of them and often find themselves in a dead end before they realize it. AlphaZero has no such prejudices or issues, and seems to thrive on snuffing out the opponent’s play. It is singularly impressive, and what is astonishing is how it is able to also find tactics that the engines seem blind to."

[chrisw]

Well, it’s an article/blog written off the backof a PR release from a publisher who doesn’t yet have the book, which is not yet actually written by some chess guys who apparently get to see some hundred of games that nobody else has seen, selectively provided them by Google DeepMind, using not known criteria. And some quotes about something more in that classic forward moving window “in a couple of weeks”.

Excuse me for being an old cynic.

Sneer all you like, but the fact remains that AlphaZero's very deep Positional understanding cannot be matched by Stockfish or any other chess engine no matter how much they progress, as their very basis is a fallacy. Most chess programmers made the basic mistake of trying to reduce Chess to just a set of mathematical/scientific principles, which it is NOT.
As is said, Chess is not just a Science, it is also an Art. Programmers ignore the Art aspect at their own peril or perhaps they have no choice as they cannot quantify Art.
Whatever the reason, as long as they see Chess in terms of Piece/Material value, they can never hope to master Chess completely.
This aspect is confirmed by the Article itself, where its says that even the Programmers of AlphaZero are THEMSELVES unsure how AlphaZero goes about its business, or words to that effect.
You can quibble about 'unknown criteria' all you like, but you are still missing the Woods for the Trees.
The detractors are so obsessed by the Conditions under which the Match is being played that they fail to see ( or choose not to see) the MANNER in which AlphaZero decimates Stockfish. Stockfish appears to be so completely clueless as to the long-term intentions of AlphaZero, that it hardly seems to matter what Hardware or Time Control is being used.
Another interesting Term I've seen used w.r.t. AlphaZero is the "Boa-Constrictor approach", that is, it seems to be more interested in anticipating and stifling the Play of the opponent rather than playing brilliant moves itself.
I've seen this approach during my Club playing days against human opponents ; of course AlphaZero has the advantage of not making any Tactical errors as humans would make, which makes it more deadly.
This is explained better in this Article....
https://www.nextbigfuture.com/2018/01/c ... style.html.
To quote..."Modern chess engines are focused on activity, and have special safeguards to avoid blocked positions as they have no understanding of them and often find themselves in a dead end before they realize it. AlphaZero has no such prejudices or issues, and seems to thrive on snuffing out the opponent’s play. It is singularly impressive, and what is astonishing is how it is able to also find tactics that the engines seem blind to."

Thanks for the lecture. When I need to revisit the basic principles of 2+2=4, I’ll know who to ask. Thanks.

The problem of Lc0 and by extension Alpha Zero is the factual mismatch of alleged and actual performance vs Stockfish, for LC0 has surely had enough resources thrown at it by now to have equated to AZ. But where are the 100 game matches consisting solely of wins and draws with zero losses? Where’s the raw data?

[Nay Lin Tun]

The statistics of zero loss in 100 games between 100 elo difference is very odd!
Alpha Zero and Stockfish 8 would have 200-300 elo gap with their hardware setup and time control, however Alpha Zero likely spoiled a lot of winning endgames as draw, and statistically end as +100 elo!
( It can be seen in Leela end games too)

[dkappe]

Having trained Ender nets and tested the various Leela nets extensively, I can say that stockfish 9 is at a different level from 16p on down. If A0 has managed to solve this problem, I would be fascinated to find out how.

[Leo]

Well, it’s an article/blog written off the backof a PR release from a publisher who doesn’t yet have the book, which is not yet actually written by some chess guys who apparently get to see some hundred of games that nobody else has seen, selectively provided them by Google DeepMind, using not known criteria. And some quotes about something more in that classic forward moving window “in a couple of weeks”.

Excuse me for being an old cynic.

Sneer all you like, but the fact remains that AlphaZero's very deep Positional understanding cannot be matched by Stockfish or any other chess engine no matter how much they progress, as their very basis is a fallacy. Most chess programmers made the basic mistake of trying to reduce Chess to just a set of mathematical/scientific principles, which it is NOT.
As is said, Chess is not just a Science, it is also an Art. Programmers ignore the Art aspect at their own peril or perhaps they have no choice as they cannot quantify Art.
Whatever the reason, as long as they see Chess in terms of Piece/Material value, they can never hope to master Chess completely.
This aspect is confirmed by the Article itself, where its says that even the Programmers of AlphaZero are THEMSELVES unsure how AlphaZero goes about its business, or words to that effect.
You can quibble about 'unknown criteria' all you like, but you are still missing the Woods for the Trees.
The detractors are so obsessed by the Conditions under which the Match is being played that they fail to see ( or choose not to see) the MANNER in which AlphaZero decimates Stockfish. Stockfish appears to be so completely clueless as to the long-term intentions of AlphaZero, that it hardly seems to matter what Hardware or Time Control is being used.
Another interesting Term I've seen used w.r.t. AlphaZero is the "Boa-Constrictor approach", that is, it seems to be more interested in anticipating and stifling the Play of the opponent rather than playing brilliant moves itself.
I've seen this approach during my Club playing days against human opponents ; of course AlphaZero has the advantage of not making any Tactical errors as humans would make, which makes it more deadly.
This is explained better in this Article....
https://www.nextbigfuture.com/2018/01/c ... style.html.
To quote..."Modern chess engines are focused on activity, and have special safeguards to avoid blocked positions as they have no understanding of them and often find themselves in a dead end before they realize it. AlphaZero has no such prejudices or issues, and seems to thrive on snuffing out the opponent’s play. It is singularly impressive, and what is astonishing is how it is able to also find tactics that the engines seem blind to."

When its all been said and done, more will be said than done.

[jp]

This aspect is confirmed by the Article itself, where its says that even the Programmers of AlphaZero are THEMSELVES unsure how AlphaZero goes about its business, or words to that effect.

That is a completely inconsequential statement. It's just saying NNs are black boxes that you cannot look inside.
That's true for good NNs and bad NNs.
It does not tell anything about how good or bad an NN is for a task.

[jp]

Having trained Ender nets and tested the various Leela nets extensively, I can say that stockfish 9 is at a different level from 16p on down.

What is the difference in endgame performance between Ender & Lc0?
How much stronger in endgames is SF9?