2

Leela Chess Zero is a neural-network based chess engine trained entirely from self play (that's why it has the word "zero" in its name). Leelenstein uses the same binary,* but a different method to train its neural network, utilizing supervised learning.

The latest results indicate that Leela is stronger than Leelenstein. Why?

I'm especially interested in an explanation of why the zero approach is better than the supervised learning approach, if it is indeed superior (presumably it is, or I don't see why Leela would win in the first place).

*As I understand it, NN-engines need two things to operate: the binary file, and the neural network itself. Leela and Leelenstein share the same binary.

  • The link has no useful information unless you sign up and start exploring the site. Do you have a link that actually proves your claim? It would appear that the zero approach has been trained for a long time already and the new one is just starting to do training. Clearly they have different networks but can not have the same binary file if they are learning differently. – edwina oliver Jan 31 at 2:29
  • 2
    @edwinaoliver what link are you talking about? I see everything about how Leelenstein is trained on that Patreon page. Also, Leela regularly resets to start afresh (the latest test is at least the 4th time it's done this), so "has been around longer" does not seem accurate. – Allure Jan 31 at 2:35
  • neither link showed me anything useful . I stopped when josh site said to sign up. – edwina oliver Jan 31 at 3:58
  • @edwinaoliver Scroll down. This is a direct quote from the Patreon website. "Leelenstein started out mostly CCRL games, with some 40b and T30 games. Currently uses CCRL games + more T30 games + SGDR + GGT + SE, but started with the same weights as the previous to retain some information from those other games." And no, I did not sign up either. – Allure Jan 31 at 4:09
2

As far as I can see Leelenstein is mostly derivative of Leela. It uses a fraction of the games produced by Leela's self-play to get to a similar playing strength.

Given that most of these games have also been used to train Leela, there is little reason to expect Leelenstein to be better.

Basically Leela produces customized trainingsdata via self-play, Leelenstein uses games produced by other engines, which may or may not be consistent with each other.

After self-play Leela uses the data in a supervised fashion. Leelenstein is not a different approach. It's the same approach, only skipping the step of producing its own trainingsdata.

| improve this answer | |
  • Hmm, wouldn't it in principle be possible to use Leela's losses as a training set for Leelenstein, thereby training Leelenstein to exploit Leela's weaknesses? – Allure Feb 5 at 1:48
  • Well, most of the games produced by Leela are games against itself. But there is a project that is aimed at training a NN-engine to beat Stockfish. – BlindKungFuMaster Feb 5 at 8:53

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.