code.talks 2019 - The Building Blocks of Superhuman Poker AI

Поделиться
HTML-код
  • Опубликовано: 9 ноя 2024

Комментарии • 7

  • @acdemiralp2998
    @acdemiralp2998 3 года назад +4

    Starts at 13:20

  • @dmitry6609
    @dmitry6609 4 года назад +1

    how come it costed only 150$? You told it took 8 days with 64 cores and 512 GB of memory. I thought on AWS it would cost around 800$. How is it possible to spend only 150$ for this?

  • @dmitriys4279
    @dmitriys4279 5 лет назад

    so deep learning (actor-critic) doesnt work for texas holdem poker , does it ?

    • @is0morph
      @is0morph 4 года назад

      systems like deep stack use deep learning and there's certainly potential to it. also note that the same authors have a deep learning variant for the abstraction part of the algorithm. so it's just that nobody managed to build a deep learning version that beats top pros, but that might just be a matter of time. still, it's impressive that pluribus doesn't use any DL at all in the first place

    • @dmitriys4279
      @dmitriys4279 4 года назад

      @@is0morph ohh now I understand, thanks for reply. It's a great achievement. Let me clarify, all players have fixed stack size at the start of every hand?

    • @is0morph
      @is0morph 4 года назад

      @@dmitriys4279 yeah, that seems to stabilize training. the authors note this in a half-sentence in the addendum of the paper. in fact, there are a lot of these little details that are hard to decrypt. it's not clear to me why this doesn't tamper with the ability to learn to deal with all-in situations etc., but it seems to work for them.

  • @pokeraibot328
    @pokeraibot328 4 года назад +1

    Hello)