At the risk of underselling everything that preceded it, the real Oh Shit moment of DeepMind’s StarCraft II demonstration didn’t arrive until the ninth game.
I’m no expert here, but I’m having trouble seeing how stalker blink micro differs from say, an aimbot in an fps? There are some impressive things going on here (how good it is at judging engagements stands out to me), but I feel like it’s being propped up by something computers are just naturally good at. Am I missing something here?
No, and I think the casters and players sort of agreed? If you don’t somehow limit the ‘cameras’ and inputs the AI makes, then this is a lot less interesting.
Seeing it play DotA (I think it was the same AI?) made more sense, since there’s already some limitations to what can be done with more input, but when what you primarily end up highlighting is ‘computers can react faster than humans’ it’s sort of a ‘well, yeah, duh’. Which is a shame, because there is something inherently interesting about the simulation aspect itself, which drowns in the ‘real good at stalker micro’ talk.
Like the ‘is oversaturating workers better?’ question is really interesting for instance.
Even looking at pure micro, this is at least a lot more advanced than a simple aim bot. After all, aimbots with perfect accuracy have existed for decades, but this level of skill in an AI for more nuanced tactical movement is quite new. It’s also important to point out that this wasn’t something hard-coded into the AI itself, but rather something it learned to do through experience.
It’s true that this will never be “fair” until the AI has to inhabit a physical body that interfaces with the game through a mouse and keyboard, but I’d be surprised if AlphaStar didn’t become a source of innovation in Starcraft the way AlphaZero has in chess.
One additional point on the subject of the AI not being limited to a physical body, something that got pointed out during the stream was that the AI wasn’t necessarily exploiting its speed (as much as it could have). Its APM was actually one or two hundred lower than the human player.
So it’s not that it was constantly controlling every unit simultaneously in an inhuman way. It gave fewer orders overall than the human player, but it gave the right orders (more or less) every time, while the human player made mistakes.
(in case there’s any confusion, I mean this as a ‘yes, and’ support for what Noelle was saying, not as an ‘actually’ rebuttal)
The main thing that sets Deepmind apart from other existing SC AI’s is that it is more or less self-taught. The other AIs more or less follow a script “taught” by the human programmers, but Deepmind learns only by watching replays and playing against itself thousands of times. I think that gets lost in the perfect-stalker-micro debate. This is supposed to be a demonstration of how an AI can learn a complex system, not how an AI performs inhuman tasks.
Hi all. It’s Josh, the writer, and this is exactly the point.
There are a bunch of other minor details that I didn’t have room to fit in here, but seem to keep re-appearing as misunderstandings in conversations about AlphaStar that make this all the more impressive:
*Its APM is capped to approximate human levels. From Oriol Vinyals in their reddit AMA: “We set a maximum of 600 APMs over 5 second periods, 400 over 15 second periods, 320 over 30 second periods, and 300 over 60 second period.”
*Its “reaction time,” to intake what is happening on the screen, communicate it back to the system, process it and make a decision, send the response back is 350 ms, which is actually lower than human pro players.
*Yes, the version that won 10-0 uses the screen in a huge zoomed out way that a human can not. However, when they tested where AlphaStar’s “focus” was, it was refocusing on different areas at approximately the same rate that pros do. And, the version that MaNa beat in the live game, that has been trained to have to use the camera more like a human, was less experienced and still only had an estimated MMR rating a couple hundred points below the zoomed out full screen version.
Finally, people had previously built Brood War bots with ridiculous APM and unit control, who could do wild things with their armies, and they still got crushed by decent human pros because their decision making was so poor. Watching AlphaStar do things like decide when to engage or disengage from fights, micro skills aside, is light years beyond anything we’ve seen before.
Thanks to everyone who read the piece.
