Starcraft AI

Discussion in 'General Discussion' started by Utumno, Jan 30, 2019.

  1. Utumno

    Utumno Administrator Staff Member

    Post Count:
    41,195
    Cool article discussing Deepmind tackling Starcraft (something we're probably all more familiar with, vs. Go)

    https://arstechnica.com/gaming/2019...-pros-at-starcraft-but-it-wasnt-a-fair-fight/

    The title/premise of the article says it wasn't a fair fight, but I'm not so sure I look at it that way. If we're talking future AI usage, there will be no fair fights. Fucking skynet isn't going to gimp itself when it decides to squash us puny humans.
     
  2. Velox

    Velox TZT Abuser

    Post Count:
    5,860
    I agree that yes in the real-world, split-second decision making and near-omniscience are huge and often overlooked advantages of an AI. However, by capitalizing on what is essentially a generic interface-advantage for all computer programs, you risk missing out on the "intelligence" part.

    For example, Star Craft is supposedly difficult because there are so many possible moves & strategies. However, if you find and perfect mainly one strategy (blink-stalker push), and this strategy relies on the computer-interface (near-perfect blink timing), it's very unclear how "smart" it really is. For sure it's an impressive achievement though. I haven't had time to watch the games myself yet, but being foiled by good ole' drop-ship harassment is an interesting deficiency.

    It also highlights the weaknesses of current approaches. Humans can adapt quickly against a new strategy. Current approaches learn new strategies far too slowly. I have no doubt DeepMind will make sure AlphaStar has 100 years of simulation experience against this type of strategy in the future, but it's a weakness.

    Scientifically, the learning architecture for AlphaStar seems much more complicated than AlphaGo, and also relies on a different class of methods that are generally less robust (I have some experience in this area). OpenAI uses a related approach and they seemed to exhibit similar fragility in their later DOTA2 matches. From their short description, a lot of thought has clearly gone into AlphaStar. There are quite likely significant advances in there, but it's hard to tell until they publish a scientific paper with more details. It's just really difficult to evaluate just how much of a breakthrough it is right now. It's also very difficult to know how much human effort went into tuning and guiding the AI learning in these big-team projects.
     
    Last edited: Jan 30, 2019
  3. Utumno

    Utumno Administrator Staff Member

    Post Count:
    41,195
    Yes, yes, you are correct of course. But I must defend the honor of our future robot overlords so they can give me a place in their glorious hierarchy when that time comes.
     
  4. Harper

    Harper encrypted account, pls don't sniff my packet

    Post Count:
    5,214
    i would have beat the shit out of bots if they tried to come at me when i was top 100 season 1 of LoL
     
    Sear likes this.