1. 2

    Impressive. Seems like the main technical contribution here is that bots were occasionally forked during training, played against each other in a league, and the final “bot” is actually a team of five complementary (“least exploitable”) bots. Each bot has a slightly different strategy. Didn’t catch how they select the bot before a game (random?) or if the AI can switch between strategies mid-game like a human player can.

    This is nitpicky, but I do wish they chose an RTS that doesn’t reward micromanaging units as heavily as Starcraft. Even though DeepMind capped action per minute to something reasonable like 300, the AI still leaned pretty heavily on its inhumanly precise and efficient micro for most of the games. Nonetheless, the AI still demonstrated long-term planning and would have been competitive even with ordinary human levels of micro, and that’s amazing.

    Don’t think anyone thought we’d solve SC2 so quickly - probably including parts of DeepMind.

    1. 3

      “Didn’t catch how they select the bot before a game (random?) or if the AI can switch between strategies mid-game like a human player can.”

      From what I remember, it was their opinion of the best-performing bots of the league. There might have been measurements. They often talked pretty subjectively, though.

      “Even though DeepMind capped action per minute to something reasonable like 300, the AI still leaned pretty heavily on its inhumanly precise and efficient micro for most of the games.”

      That might be addressed by throwing it off randomly. It did something else, though, which is more interesting.

      They kept mentioning it kept its whole army together a lot. One of commenters said there was a highly-disputed action that would select all units. So, it might have been doing that. Although I can’t say as not SC2 player, I think that’s a great explanation if that action exists and AI could use it. The APM’s were limited. It could move more units per time period by selecting all. So, the strategies to maximize unit movement might have evolved to use select all more often with many strategies built around that.

      “Nonetheless, the AI still demonstrated long-term planning and would have been competitive even with ordinary human levels of micro, and that’s amazing.”

      It did even better than that. It either knew or could fake a perception of what humans were doing or capable of at each moment. That was impossible for prior agents. They got tricked so easily. The very, last game was only one where one glitched on that. Same kind of exploit as older AI’s. Just like the articles I’ve seen talking about how humans and AI’s have different, blind spots with different, massive screwups. I’m most interested in whether they can address that. Its reaction still shows a lack of deep understanding and strategy that would lead a human to instantly address or ignore something for good reasons. All the other actions indicate an understanding of a lot of what can be done, feigning it, or working around it with good strategy nicely.

      “Don’t think anyone thought we’d solve SC2 so quickly - probably including parts of DeepMind.”

      Yeah, one dumbass in particular comes to mind. He was still happy to be right just that one time out of 11. Gotta cherish what victories the human race has left in constrained environments against the products of rich companies.

      Nah, really, I’d love for them to put this thing online against other pro players if they haven’t. There’s no telling what would happen with all that exposure. I think they should keep it out of chat, though. The results might be entertaining to some. I’m just worried that prolonged exposure to interacting with certain kinds of people could make every encounter with these things a bad, PR result that could affect the project’s funding. ;)

      1. 3

        I was about to call you out for insulting the author of that comment, then I realized it’s self-deprecation. :)

        Yeah, this is pretty cool. Not a full solution yet, but clearly getting there.

        1. 4

          “I was about to call you out for insulting the author of that comment, then I realized it’s self-deprecation. :)”

          I was almost always the one hit by personal attacks here, not delivering them. I counter ideas, not people, the most. It was setup to look like something worse, though. Glad you got a fun surprise out of that as intended. :)

          “Yeah, this is pretty cool. Not a full solution yet, but clearly getting there.”

          Oh come on. I want to say that, too. It dominated them, though. The stress in their eyes and reactions trying to understand what it was doing was a complete reversal of previous fights with Starcraft AI’s which made the AI look as dumb as it was at the slightest trick. That change hit me the hardest. That last victory against the AI spotting same weakness as old ones is my sliver of hope for the average laborer in anything as complex as Starcraft where the graphics hardware is cheaper than people. They’ll demonstrate their value and keep their jobs partly by pointing out to avoidable, colossal screw-ups the AI’s can make. :)