7

In his process of learning Chess and getting stronger and stronger each second AlphaZero learned openings. Some were played often at the beginning, but lost in AlphaZero's popularity over the time (e.g. French Defence). You can find four examples here: The future is here – AlphaZero learns chess

I read somewhere that AlphaZero played (and so: evaluated) 12 openings. I would like to see a list of them, perhaps with the graphs (not only the 4 mentioned above).

Thanks in advance.

10

You can find the complete table in their paper. See table 2 in the arXiv version linked below:

How to read them:

  • The plots show the proportion of times alphazero played a given opening during its self-training games as a function of training time. So e.g., you can see that its interest in employing the French defense peaked after 2 hours, but past that, it abruptly drops to near 0, indicating that after 2 hours+ training it realised the emergent lines from that opening are non-optimal compared to other choices, such as the Caro-Kann defense (which had a promising plateau after 2 hours but eventually dropped too, plot shown below).
  • Then below the diagrams, you see how it fared against Stockfish in 100 game matches for each line. Finally, alphazero's principal variation for each opening is also indicated below the plots. Please see the table's caption in the paper for any other details.

Overall, the English opening stands out: it kept employing it consistently throughout its training. Ultimately, the pattern suggests a tendency towards more versatile openings.

Caro-Kann diagram from table 2: [Ref]

enter image description here


[Ref]: Silver, David, et al. "Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm." arXiv preprint arXiv:1712.01815 (2017)

  • Thanks! Interesting, that the Queens Gambit seems to be very drawish and AlphaZero in the King's Gambit does not even win one game with Black! – BNetz Feb 20 '18 at 8:32
  • It's quite some time I was reading the paper, but I have feeling that he almost stopped playing e4 later and started to employ 1.Nf3 if I recall well. That would be strong reason for no Caro-Kan, French, Sicilian... – hoacin Feb 21 '18 at 12:05
  • 1
    @hoacin As far as I see, nothing is mentioned in the paper about the proportion of times 1. Nf3 was chosen. On the other hand, in the table 2 of the paper they only analyse the common human openings, among which 1. Nf3 does not take part. Moreover, as it says in the paper: "Each of these openings is independently discovered and played frequently by AlphaZero during self-play training. " Which means that the discussed statistics are gathered independently of whether 1. Nf3 was more frequently adopted or not. By the way, I believe you meant "it" and not "he." :) – Phonon Feb 21 '18 at 13:34
  • We must have into account, that Alpha Zero learned but playing against an engine that have an opening book. Therefore, it is more apropiate to think the Alpha Zero algorith as a kind of probe that decodified the oponent engine openings book. – djnavas Feb 22 '18 at 9:16
  • @djnavas Stockfish did not use an opening book in that match, see the link. Moreover, in the paper, they explicitly say the openings were self-discovered independently, meaning simply by considering optimal moves (in the games against itself) without any input knowledge of existing opening theory. – Phonon Feb 22 '18 at 9:49
2

Check the paper itself, in particular page six. There are indeed 12 openings, some of which AlphaZero kept playing, others which it played for a while and then discarded.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.