Googles AI beats European Go Champion 5-0

Discussion in 'Unrelated Discussion' started by cola_colin, January 27, 2016.

  1. cola_colin

    cola_colin Moderator Alumni

    Messages:
    12,074
    Likes Received:
    16,221
    Absolutely.
    Same concept applies when somebody asks you to formally describe how you can tell there is a dog in the picture.
    You can't.
    stuart98, zihuatanejo and elkanfirst like this.
  2. zihuatanejo

    zihuatanejo Well-Known Member

    Messages:
    798
    Likes Received:
    577
    [​IMG]
    stuart98 and Devak like this.
  3. cola_colin

    cola_colin Moderator Alumni

    Messages:
    12,074
    Likes Received:
    16,221
    arseface, stuart98 and Quitch like this.
  4. cola_colin

    cola_colin Moderator Alumni

    Messages:
    12,074
    Likes Received:
    16,221




    3-0 for AlphaGo, which apparently has improved quite a lot since last year. They'll still play the 4th and the 5th match I think, but just for show. I guess now it really can be said for sure: Go no longer is a domain in which humans outrank AIs. Albeit I wonder how long it will be until an AI like this can do it's magic without a huge cluster of CPUs and GPUs.
    arseface likes this.
  5. websterx01

    websterx01 Post Master General

    Messages:
    1,682
    Likes Received:
    1,063
    Quite a while, I imagine. Without knowing anything about the AI, I'm assuming that it computes all possible next moves and determines the best next move. And that is a lot of combinations to do in a short period of time. It also wouldn't surprise me if it does go in with an optimal strategy to get started.
  6. cola_colin

    cola_colin Moderator Alumni

    Messages:
    12,074
    Likes Received:
    16,221
    There are more possible legal Go positions than atoms in the universe. The whole point of this is that Go is simply impossible to brute force. You have games with 100+ turns and 100+ (or 200?) possible moves each turn for each player. On top of that it is really hard to judge how well a position is for a player.
    To be precise there are 208168199381979984699478633344862770286522453884530548425639456820927419612738015378525648451698519643907259916015628128546089888314427129715319317557736620397247064840935 legal positions you'd have to consider to completely solve Go by brute force.

    AlphaGo certainly considers tens of thousands of ways it can be play out from any given location, but that is a tiny, tiny fraction of the possibilities that Go offers. The amazing thing about AlphaGo is that using two neural nets, one for selecting a move and the other for evaluating how good a board is, it actually has managed to basically get an "intuition" about how to play. It's not a victory of raw computation power, it is mainly a victory of machine learning.
    It plays sort of like a human plays ("consider which moves seem to be good, think about what happens if you do those moves only, ignore moves that look bad from the start") and it has learned sort of like a human as well by studying games of the best players and than experimenting against itself.
    Last edited: March 12, 2016
    Quitch and websterx01 like this.
  7. arseface

    arseface Post Master General

    Messages:
    1,804
    Likes Received:
    502
    I need to watch the second and third games. The first was pretty brutal.
  8. squishypon3

    squishypon3 Post Master General

    Messages:
    7,971
    Likes Received:
    4,356
    After seeing GO here first I've given it a go on mobile. Since I liked it's idea.

    Fun, but I'm getting rekt by the computers, even on easy. XD

    Its good fun to waste time and improve on my downtime. XP
  9. cptconundrum

    cptconundrum Post Master General

    Messages:
    4,186
    Likes Received:
    4,900
    Wasn't the score 4-1, not 5-0?
  10. DeathByDenim

    DeathByDenim Post Master General

    Messages:
    4,327
    Likes Received:
    2,125
    The 5-0 was against the European top player, Fan Hui, in January, which started this thread. The 4-1 is against the world top player, Lee Sedol, that was played just last week.
  11. cola_colin

    cola_colin Moderator Alumni

    Messages:
    12,074
    Likes Received:
    16,221
    Though the AlphaGo that Fan Hui faced in October of last year (it was only made public in late January this year) apparently was vastly inferior to the one that defeated Lee Sedol.
    Fan Hui actually played 10 games overall. He won a single game if I remember correctly, but it was an unrated practice game.
  12. cptconundrum

    cptconundrum Post Master General

    Messages:
    4,186
    Likes Received:
    4,900
    Reading timestamps is hard.
    cola_colin and proeleert like this.
  13. BulletMagnet

    BulletMagnet Post Master General

    Messages:
    3,263
    Likes Received:
    591
    I believe I read that AlphaGo was running on a single desktop for this.

    Training the computer is a much, much more time consuming process than using what has been learned.
  14. squishypon3

    squishypon3 Post Master General

    Messages:
    7,971
    Likes Received:
    4,356
    And I'm sure it didn't take any GPUs, right? Since there's no graphic processing going on?
  15. Quitch

    Quitch Post Master General

    Messages:
    5,850
    Likes Received:
    6,045
    No, it was using a distributed network of around 1200 CPUs and 180 GPUs.

    GPUs are much better than CPUs for operations you want to run in parallel so are used all the time for non-graphics processing.
    websterx01 likes this.
  16. websterx01

    websterx01 Post Master General

    Messages:
    1,682
    Likes Received:
    1,063
    Think of a single high end GPU as a 3000 core CPU. It might only run at 1GHz, but that's far better for many things than 4 cores at 4GHz.
  17. cola_colin

    cola_colin Moderator Alumni

    Messages:
    12,074
    Likes Received:
    16,221
    More like a very stupid 1ghz cpu. They can't do all that much, but they're really really fast if you have simple number crunching to do that can be parallized very well. Like the 1 million+ pixels you'd work on for an imagine. Or the many small calculations you do when you evaluate an artifical neural network.
    websterx01 likes this.
  18. cola_colin

    cola_colin Moderator Alumni

    Messages:
    12,074
    Likes Received:
    16,221
    tunsel11 and ljfed like this.
  19. tatsujb

    tatsujb Post Master General

    Messages:
    12,902
    Likes Received:
    5,385
    awesome. though really FA would be a healthier environement to grow an ai in. there really is only one best strategy in starcraft that you modulate according to the situation.

    a case-by-case bot build from the counsel of the starcraft top 50 will systematically beat an neural net AI figuring it out by itself.
  20. cola_colin

    cola_colin Moderator Alumni

    Messages:
    12,074
    Likes Received:
    16,221
    I don't think there is much of a difference, you're not giving sc2 enough credit :p

    Well yeah, the current state of sc AIs clearly is more in the region of hard coded stuff. But that is because figuring out how to learn to play without hardcoding it is incredibly hard and nobody really knows how to do it.
    Ergo a perfect next step towards better AI.
    tatsujb likes this.

Share This Page