I [Kasparov] can look at AlphaZero’s games and understand the potential weaknesses. And I believe it has made some inaccurate evaluations, which is natural.

The reason he says that is AlphaZero uses statistics of ALL games to determine piece value.

Which pieces did he specifically mention?

What other shortcomings does alpha zero have that humans could exploit?

  • Needs details - where did Kasparov say that? What's the source? – Allure Feb 21 '20 at 21:51
  • @Allure I linked the comments in my answer. – PhishMaster Feb 21 '20 at 21:54
  • Kasparov was quoted in an article by Wired magazine online. – edwina oliver Feb 21 '20 at 23:01

First, as great as Kasparov is, compared to AlphaZero, even he is a patzer; so for him to say that he sees weaknesses, he may be correct, but it could also be, simply, that AlphaZero is playing at a level beyond even his great understanding, and with positional/tactical minutia we, humans, have yet to realize are important, or cannot see at all.

In this interview, he only mentioned bishops versus knights, saying:

I can look at AlphaZero’s games and understand the potential weaknesses. And I believe it has made some inaccurate evaluations, which is natural. For example, it values bishop over knight. It sees over 60 million games that statistically, you know, the bishop was dominant in many more games. So I think it added too much advantage to bishop in terms of numbers. So what you should do, you should try to get your engine to a position where AlphaZero will make inevitable mistakes [based on this inaccuracy].

Note: I believe the 60 million games he mentions are all games played against itself as it was teaching itself how to play, rather than a database of human games. After all, Mega 2020, has only 8 million games in it.

I also found this article that Kasparov penned, in which he said:

But in my observation, AlphaZero prioritizes piece activity over material, preferring positions that to my eye looked risky and aggressive.

Considering the recent question, "Is Stockfish 11 (level 20) beatable?", and that AlphaZero destroyed a recent, although previous, version of Stockfish, I do not believe that you will find any weaknesses that any human will be able to exploit. That will be the domain of the next, stronger, computer.

  • 1
    AlphaZero beat Stockfish 8, which is about 150 elo weaker than Stockfish 11 on a single core. Also, NN engines have clear weaknesses, e.g. when depth of calculation is most important (endgames) they can be significantly inferior to AB engines. Like Stockfish they're not unbeatable either. It's true you can't beat NN engines as an unassisted human either, but you can as an assisted human (in correspondence chess). – Allure Feb 21 '20 at 22:09
  • 1
    @Allure. Indeed, one the projects that some of the Leela people are working on is "teaching" Leela and some version of Stockfish to work together, mostly trying to evaluate positions and see which engine would handle it better. – Don Thousand Feb 22 '20 at 0:16
  • I think the algorithm is going to be, when a probabilistic engine sees multiple best moves as having an equal probability of winning, the algorithm will switch over to a minimax engine to find the quickest win. – Dave Babbitt Jul 30 '20 at 13:13

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.