18 votes

DeepMind's StarCraft II AI "AlphaStar" will play some games anonymously on the Europe competitive ladder against opt-in opponents

5 comments

  1. SUD0 Link
    I love following this stuff! I used to be a really big into SC2 and played it a decent amount. I watched the live stream of Alphastar playing a pro player back when they first revealed it. I...

    I love following this stuff! I used to be a really big into SC2 and played it a decent amount. I watched the live stream of Alphastar playing a pro player back when they first revealed it. I thought the demonstration was a little unfair, however. It was apparent while watching from Alphastar's POV that it didn't beat its human opponents with some brilliant strategy. It won by making units that are impossible to kill if handled perfectly (like blink stalkers) and had superhuman micromanagement. It basically just out-muscled the human player. I am glad to see that the researchers consulted pro players to determine an APM that is a little more restricting. It should force Alphastar to try and out strategize its opponents instead of just forcing its way to a win.

    10 votes
  2. [3]
    Douglas Link
    I don't play any Starcraft whatsoever, but just following this thing's progress has been pretty fascinating. I would love to see this type of aggregated machine learning in a game like Left 4 Dead...

    I don't play any Starcraft whatsoever, but just following this thing's progress has been pretty fascinating.

    I would love to see this type of aggregated machine learning in a game like Left 4 Dead or Dying Light. Like the game's opponent would have a hive-mind that the entire userbase would collectively try to defeat, and the AI hive-mind learns from each tactic employed against it enough to find new ways to counter it.

    It feels like there's a lot of exploration in storytelling that games can harness with this concept.

    6 votes
    1. Deimos (edited ) Link Parent
      It's not really what you're talking about, but there are a couple of games in the "AI War" series (second one is currently in Early Access) that set up kind of an interesting scenario related to...

      It's not really what you're talking about, but there are a couple of games in the "AI War" series (second one is currently in Early Access) that set up kind of an interesting scenario related to needing to fight back against an overwhelmingly powerful AI that you can't possibly defeat in a straightforward fight.

      They wrote a number of good blog posts about the approach to the AI in the original game, this is the first one in the series: Designing Emergent AI, Part 1: An Introduction

      7 votes
    2. stu2b50 Link Parent
      They can't really do that. They learn from at least a few hundred thousand samples. If you set the learning rate to be so high the weights would adjust from a few games with players it wouldn't...

      They can't really do that. They learn from at least a few hundred thousand samples. If you set the learning rate to be so high the weights would adjust from a few games with players it wouldn't even get close to converge.

      Remember, in the end it's just gradient descent, not magic.

      6 votes
  3. mrbig (edited ) Link
    When computers became better than World Champions at chess my interest in the sport was greatly reduced. I know it’s silly: cars didn’t make running competitions obsolete. But when it comes to...

    When computers became better than World Champions at chess my interest in the sport was greatly reduced. I know it’s silly: cars didn’t make running competitions obsolete. But when it comes to reasoning activities I can’t help but feel cheated. It’s the one thing no animal can do better than us. IDK if that makes sense.

    1 vote