Which opening is most often used by AlphaZero?

I heard that it prefers the "English opening" as white, but I only got that information by hearsay.

Does anybody know?

Found the source of the hearsay (Time = 0:53)

EDIT: The video is already 3 years old (2018). AlphaZero also played against the Stockfish 8 engine. Does anybody know something about AlphaZero's openings in 2021?

I think it did prefear the English, although the only deference between this and other openings is that it delays the draw for much longer I reckon.

Moreover, the various graphs are based upon human openings ; nowadays, the engine can come up with slightly more concise moves that may affect the results of hundred-games matchs.

Well I guess that depends on the programming of the AI. If it wants to achieve wins or draws, but not lose, then delaying a draw would not be preferable. In the contrary, it would make more sense for the AI to make the game shorter to be able to play more games and find more winning positions (if it can predict the draw many moves ahead). Yet, I don't think that this is a criterion either, since the goal is to find the best overall strategies, not the fastest ones.

I think the information to extract in this case is the WIN/DRAW/LOSS-ratio described above under each opening. The English opening seems to have the best ratio of all openings for now (in games with AlphaZero vs. AlphaZero).

It still might not be the best opening for humans though, since I find AlphaZero's play rather unintuitive as a human (especially the pawn movements on the h- and a-file).

I am so excited to see what AlphaZero is coming up with next time.

I think there were 6 games which it lost against Stockfish 13 so there is still room for improvement I guess.

Maybe the data is faulty though. We can not see any plateaus (stabilization of values) so it could be save to assume that it hasn't played enough yet to be able to properly display a final comparison with the best overall opening strategy.

Interesting point is that Stockfish seems to have problems not only with the English opening (8), but also the Reti (7) and Ruy Lopez (6).

Perhaps I'm being nit-picky but still important to note, AlphaZero wasn't programmed with anything.  It was given the rules of the game and then it played itself millions of times.

"Which opening is most often used by AlphaZero?"

How would we know? We use Stockfish.

_______

It actually depends on what you play first, but if you were playing as black then it might be the English?

alpha is crazy, it plays weird things even the evans gambit at one point and atonio got happy so idk at this point cause alpha really just does whatever and i enjoy its attacking power 💕

It has a huge advantage of being able to play WAY more games than any human possibly could, and remember everything it learned.

It's learned that it can sacrifice lots of material for position, using what it has to lock down the opponent's ability to maneuver, essentially equalizing the effective material and AlphaZero has the superior maneuverability. 

In some ways, it's amazing and humbling to see how long humans have been playing chess and learning strategies only to have a machine learn the game and in 4 hours show us strategies adn ways of thinking we just never considered, or considered at large.

It was all there right in front of us... :-)

alpha is the coolest engine overall so awesome 💕

Wish there would be more news. Alpha Zero has practically gone dark for the last few months I guess. We only got to see VERY few games.

There is a newer article
https://arxiv.org/pdf/2111.09259.pdf

AlphaZero now prefers 1 e4 as white.

Actually, according to the paper you linked, that's only after 100k training steps (Figure 7). If you look at Figure 5, which looks at the dominant openings from 0 to 1 million training steps, AlphaZero prefers d4 more once it reaches 1 million steps. But e4 and c4 tend to follow closely behind.

Nevertheless, d4 and e4 are both great first moves by white that are viable at all levels of competitive play.

There's also a supplementary document/web app published by the authors of the aforementioned paper that looks at AlphaZero's policy evaluation of several opening positions. You can check that out here:

https://storage.googleapis.com/uncertainty-over-space/alphachess/index.html

I'm doing an AI research master's, so I can read research papers well, but I'm still quite new to chess, so I'm having some difficulty analyzing the openings that are depicted in the supplementary document.

It would be nice to give that document some extra visibility!

Interesting. I had thought that Alphazero mainly played 1.d4 as I can only recall crushing victories against Stockfish in the QID and the French. One thing is for sure: if Stockfish plays 1...e6 in response, Alphazero will always go for the transposition into the French.