AlphaZero Chess is not that strong ...

[Ozymandias]

1) Only 1 GB hashtable (64GB to 128GB would have been more way more suitable)
2) 1 minute by move fixed time control
3) No opening book for Stockfish (AZ saved his opening knowledge from its previous games)
4) Only version 8 (the current development version is already about 40 Elo above)

1) I don't think that this makes for a big difference.
2) Time management is something that leaves to be desired in any engine, so I wouldn't expect a change in TC, to actually benefit SF.
3) If anything, you're selling this point short. The book makes ALL the difference.
4) Hundreds of tests around the web, say something along those lines.

With this 4 points upgraded to a regular level, the current Stockfish version would been already 100 Elo stronger.

I wouldn't talk in terms of Elo points, but rather in the feasibility of a stronger program to be able to demonstrate its strength. With a good book, engines currently at the top become unbeatable, so it really does't matter how much stronger an engine is, compared to another, it won't win a single game in a 100 games match.

One more thing, when you look at the graph : https://www.dropbox.com/s/kaq37zf0z66d6 ... g.png?dl=0
you see there's no progress from 300,000 to 700,000 steps. This AZ is already at his best !

This graph illustrates that A0 was at the best it was going to get... training against SF under the same conditions. I don't know if the program could've improved further changing one of those two things.

[syzygy]

This graph illustrates that A0 was at the best it was going to get... training against SF under the same conditions.

Except that it did NOT train against SF.

Maybe you really should read those papers?

[Ozymandias]

This graph illustrates that A0 was at the best it was going to get... training against SF under the same conditions.

Except that it did NOT train against SF.

Maybe you really should read those papers?

I can't change "SF" for ""itself" now. But thanks anyway. I will continue NOT reading the papers, though. It's unnecessary work, as long as I can keep relying on external fact checking.

[Uri Blass]

1) Only 1 GB hashtable (64GB to 128GB would have been more way more suitable)
2) 1 minute by move fixed time control
3) No opening book for Stockfish (AZ saved his opening knowledge from its previous games)
4) Only version 8 (the current development version is already about 40 Elo above)

1) I don't think that this makes for a big difference.
2) Time management is something that leaves to be desired in any engine, so I wouldn't expect a change in TC, to actually benefit SF.
3) If anything, you're selling this point short. The book makes ALL the difference.
4) Hundreds of tests around the web, say something along those lines.

With this 4 points upgraded to a regular level, the current Stockfish version would been already 100 Elo stronger.

I wouldn't talk in terms of Elo points, but rather in the feasibility of a stronger program to be able to demonstrate its strength. With a good book, engines currently at the top become unbeatable, so it really does't matter how much stronger an engine is, compared to another, it won't win a single game in a 100 games match.

One more thing, when you look at the graph : https://www.dropbox.com/s/kaq37zf0z66d6 ... g.png?dl=0
you see there's no progress from 300,000 to 700,000 steps. This AZ is already at his best !

This graph illustrates that A0 was at the best it was going to get... training against SF under the same conditions. I don't know if the program could've improved further changing one of those two things.

I am not sure that with a good book engines currently at the top become unbeatable.

I would like somebody to make a match of Stokcfish against itself when one side get something equivalent to 1 minutes per move on 64 cores'(maybe 120 minutes/40 moves on 4 cores) and another side get 10 times more than it to see if the result of all the games is a draw.

I believe that most games are going to be a draw but not all of them.

Of course we need to agree about specifc book that we consider to be a good book before the match.

[Tobber]

This graph illustrates that A0 was at the best it was going to get... training against SF under the same conditions.

Except that it did NOT train against SF.

Maybe you really should read those papers?

Yes, please read the papers. The figure 2 will also show that AZ scales better than Stockfish, the graph Molina refers to show Elo difference with 1 second per move. Figure 2 shows AZ gets better compared to SF with increased move time.
The Elo difference from the 1200 games with "Human openings" is +78. The calculation for the 100 games, 1 minute/move and with no opening book, is +100. It make sense that the 1200 games were played with shorter move time and the Elo difference according to figure 2 is then less than +100.

The openings played in the 1200 games were of course not the choice of AZ but forced and can give some indication of how AZ could handle an engine with an opening book.
I think this is the most impressive result from the AZ training, it found out by itself some reasonable opening theory.

/John

[hgm]

So a good book makes it impossible to see the difference between a 3500 and a 6000-Elo engine: all games will be draw. To test engines it is thus essential to play them without a good book. Tests with a book are meaningless; they say nothing about the engine, but only something about the book.

[syzygy]

This graph illustrates that A0 was at the best it was going to get... training against SF under the same conditions.

Except that it did NOT train against SF.

Maybe you really should read those papers?

I can't change "SF" for ""itself" now. But thanks anyway. I will continue NOT reading the papers, though. It's unnecessary work, as long as I can keep relying on external fact checking.

OK, so we all should just ignore whatever you are asserting here, since whatever you assert is not based on fact. I'll add you to my personal list of talkchess users that talk with no authority and just add more uninformed noise (and are proud of that to boot).

[Ozymandias]

I am not sure that with a good book engines currently at the top become unbeatable.

Neither Zor nor Akhtar lost a single game, many months ago. Both used Cryptic and strong HW, so you could argue that their invincibility was based on that, rather than the book, but I only used a 5820K, and my machine didn't lose a single game either, and it played against ALL the top centaurs. In the past, centaurs proved in many tournaments, to be 100-150 Elo points ahead of standalone engines, but that's no longer the case. The obvious conclusion is that books have only left lines to play, that are possible to draw by top engines. It doesn't matter if you can play better, there's no room for a win. You can obtain some advantage, but it will even out before the game is over.

[Ozymandias]

To test engines it is thus essential to play them without a good book.

You could also test them against weaker engines. The stronger engine would get better results and a higher Elo.

[Ozymandias]

This graph illustrates that A0 was at the best it was going to get... training against SF under the same conditions.

Except that it did NOT train against SF.

Maybe you really should read those papers?

I can't change "SF" for ""itself" now. But thanks anyway. I will continue NOT reading the papers, though. It's unnecessary work, as long as I can keep relying on external fact checking.

OK, so we all should just ignore whatever you are asserting here, since whatever you assert is not based on fact. I'll add you to my personal list of talk chess users that talk with no authority and just add more uninformed noise (and are proud of that to boot).

You should quote the full text:

This graph illustrates that A0 was at the best it was going to get... training against SF under the same conditions. I don't know if the program could've improved further changing one of those two things.

I was expressing a doubt, which you simply left out. Changing the bolded word, for the one in the papers (I'm trusting you on this), doesn't solve my doubt. It leaves it intact.

In the other thread, I made two posts that you revised:

The training phase... didn't it consist of 44 million games or something like that? If that's the case, I don't see how they could be played in just four hours.

I didn't think they could output so many games per second, people argued that the HW difference between the PC used by SF and the machine in which the NN worked, wasn't that high. Unless they replaced the machine for the match, it clearly was they case.

Where you see assertions, I was actually asking:

- How where they able to play so many games in such a short period of time?
- Did they change the machine used in the training, for the match?

I got an answer to both of them, even if they weren't formulated in a way you'd recognise.

As to whether people "with no authority" (in this case, someone who doesn't read some papers) should or shouldn't talk, you're free to make whatever list you want. You can also ask for other users to ignore such posts, but I'm sure you'll also grant them the same freedom, to act accordingly... or not.