back to article What does it take for an OpenAI bot to best Dota 2 heroes? 128,000 CPU cores, 256 Nvidia GPUs

OpenAI's video-game-playing bots are getting much better at mastering sci-fi strategy war game Dota 2, seeing off semi pro players with ease in team matchups. However, they can’t quite master the whole game to beat top professional teams – yet. Last August, machine-learning software built by the OpenAI lab headquartered in …

  1. Aaiieeee
    Thumb Up

    I would imagine the bots could possibly win a team fight at a pro level due to faster reaction times and precision calcuations (e.g. staying out of range by 1 pixel for the entire time then stepping 1 pixel closer to attack then back again before the human can react).

    But I'm not convinced it would win at broader strategy and map management. I imagine the humans would need to ensure unfair fights by way of superior game insight that the bots can't know or communicate.

    1. Dr Stephen Jones

      Last year OpenAI gave them a reaction time of zero. It’s hard to lose when the odds are so loaded in your favour. Nevertheless they were being outwitted in a few hours. OpenAI went very quiet.

      Can’t wait to see what “improvements” they’ve made this year.

    2. Anonymous Coward
      Anonymous Coward

      Finally a machine that should be able to run Linux and compete with a mid range Windows 10 PC for gaming!

    3. streaky Silver badge

      But I'm not convinced it would win at broader strategy and map management.

      It's free to innovate. Evidence suggests it does. You can see from the 1v1 matchups some of the ways it does this, it's been known to do very weird things that work and pros are saying, y'know, I didn't even consider that.

      Sure people managed to cheese it but eventually it'll learn to deal with cheese and then you're losing again.

  2. Lee D Silver badge

    Summary:

    The AI cheats, and still loses.

    On a level playing field (where it has to issue commands via a mouse / keyboard protocol with appropriate simulated human latency, and interpret the screen as per what's visible to a human rate-limited to the same extent as a human), it would get its backside kicked.

    "AI" is still not AI.

    IBM just admitted they can't even work out how to sell Watson or what to do with it that's useful. Machine learning etc. plateaus REALLY quickly and then it's almost impossible to untrain/retrain it.

    But even on a single, simple game, it has to cheat, and still loses to humans.

    1. Anonymous Coward
      Anonymous Coward

      Fracking Toasters

      Come on, Starbuck. Tells us what you really feel.

    2. JDX Gold badge

      It's not a simple game you lame troll. Chess is a simple game. Go is a simple(ish) game. ML trounces human players and is on a par (at least) with traditional non-AI software representing man-centuries of work.

      In a complex game requiring substantial strategic thinking and planning, mouse-click reactions and CPM really shouldn't be the dominating factor. I don't know if this IS the case in Dota 2 but it sounds very different to Starcraft.

  3. 89724102172714582892524I7751670349743096734346773478647892349863592355648544996312855148583659264921
  4. Anonymous Coward
    Anonymous Coward

    How many orders of magnitude

    of speedup does Reinforcement Learning need before it's even a little bit useful to anybody?

    Give me that much compute and I will give you a string that SHA-256's to a C program that plays Dota.

    1. GrumpyOldBloke

      Re: How many orders of magnitude

      Of course this is not about playing Dota. This is about playing WW3 Syria, Iran, Russia and China when the other half of the team get the robots up and running. The skills learnt here are already useful, not useful enough against teams of humans but that is useful to know as well.

      1. cray74

        Re: How many orders of magnitude

        This is about playing WW3

        "A strange game. The only winning move is not to play. How about a nice game of chess?"

        1. quartzie
          Mushroom

          Re: How many orders of magnitude

          yet strangely, some people seem determined to play it out...

          1. Destroy All Monsters Silver badge
            Holmes

            Re: How many orders of magnitude

            So the question quickly reduces to "Is a Dota-2 playing superpowered bot good for Israel"?

  5. razorfishsl

    you can bet that if the machine was not given access to see the whole map, then the results would be so abysmal it would be the end of their funding.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2019