|
Google's DeepMind AI just beat two Pros at Starcraft 2
|
Starcraft mastery is fundamentally different from civ though, from what I could glean it was mostly that the AI outmicros humans (not a surprise), but that is not a thing in civ.
I was part of a project that developed an AI that we "taught" to play Freeciv. (I placed sixth, well my AI did... it's how I obtained my copy of civ 6 as a matter of fact) That was far closer to a future with good AI for civ. The method was very similar it seems, learning AI is fascinating. Here's the project, if you're interested: https://arago.co/ai/freeciv/
I love AI! Fascinating link, Japper.
I Think I'm Gwangju Like It Here
A blog about my adventures in Korea, and whatever else I feel like writing about.
I found some interesting analysis linked in the comments on Twenty Sided.
Basically, the developers of the AI claim that it doesn't click faster than humans, because otherwise this wouldn't be interesting. That is true on average, but at peak, it's twice as fast as the fastest player in the world. On top of that, human players tend to "clickspam" which inflates their APM but does nothing in the game.  (January 28th, 2019, 15:56)RFS-81 Wrote: Basically, the developers of the AI claim that it doesn't click faster than humans, because otherwise this wouldn't be interesting. While I agree with the blog you posted that the developers tweaked the graphs to make AlphaStar look better, wouldn't it still be interesting? Let's assume the AI indeed has superhuman clicking capabilities in certain situations for a short amount of time. Isn't microing your units a huge part of RTS strategy games? Doesn't that mean that the AIs strategic approach to the game is at least good enough to beat very strong human players through superior micro skills? Maybe I am not very demanding, but I find that impressive. Some people seem to expect that the AI wins through a superior strategic understanding only, when even in the quoted blog post it is mentioned that the World's number 1 player Serral supposedly has an up to 50%(!) higher EPM number than other top players. That sounds to me like his #1 spot is quite dependent on his micro skills and not only on his strategic understanding of the game. I also find the whole idea of "limiting the AI micro skills" not as straightforward as it may seem at first glance: Where do you set your limits? At the strongest current player? Do you code in a chance for typical human errors, such as mouse misclicks? If we go down that route, should it start playing worse after a few games because a human player may become tired? Should it play worse in decisive games because a human may be nervous? For example, Google's chess AI AlphaZero calculates around 60.000 moves per second. Yet nobody does seriously complain that humans can only calculate a fraction of that number and demand that the developers should reduce its calculating speed to a human level to see how good the AIs decision making really is. blog.usejournal.com Wrote:Why is Deepmind interested in restricting the agent to play like a human? Why not just let it run wild with no limitations? The reason is that Starcraft 2 is a game that can be broken by mechanical perfection. I think this is the crux of the matter: People seem to equal AI to gameplan only, while I feel microing an army of different units perfectly is also part of AI progression and impressive as well. If you feel an AI should not have any advantage whatsoever in the "mechanical execution" part of a game, then probably AlphaStar is not an impressive feat. But then again, if you follow that thought, the AI in almost all genres outside turn based strategy will probably have an unfair advantage. (January 29th, 2019, 12:03)Gustaran Wrote:(January 28th, 2019, 15:56)RFS-81 Wrote: Basically, the developers of the AI claim that it doesn't click faster than humans, because otherwise this wouldn't be interesting. Counterpoint: I agree, but consider it to be a generous-to-the-AI compromise, as the AI does not have to use the same interface humans do, which in this game where mechanical skill is relevant is a big deal. The AI is excused from it's duty to do vision and audio processing, and motor control of mouse and keyboard.
"Not interesting" was too strong a statement. But the power of perfect micro has been demonstrated before: https://www.youtube.com/watch?v=IKVFZ28ybQs
January 30th, 2019, 00:41
(This post was last modified: January 30th, 2019, 00:47 by antisocialmunky.)
Peak was 1.5k APM during the 3 pronged micro battle in game 4, humans usually are around 300 APM while spamming. Likewise, the AI could see the entire map at once in all but the 1 game it lost. Its a great showing though, much better than I expected. Hopefully they can do some more convincing follow ups.
https://www.youtube.com/watch?v=H3MCb4W7-kM is probably the best source on this. I like how he brings up how nice it would be to study the 'AI confidence' level for victory because that's traditionally been a very hard thing to gauge by players. Usually players play conservatively and bank incremental advantage instead of going for an outright kill so seeing how an AI analyzes % chance to win at any given time would be a useful tool to have.
In Soviet Russia, Civilization Micros You!
"Right, as the world goes, is only in question between equals in power, while the strong do what they can and the weak suffer what they must." “I have never understood why it is "greed" to want to keep the money you have earned but not greed to want to take somebody else's money.” 
Turns out AlphaStar is susceptible to having its puppet strings pulled: https://www.escapistmagazine.com/v2/2019...precision/
(though that was a version that had played fewer practice games, no idea if the later ones are better about it) |
|
Are you, in fact, a pregnant lady who lives in the apartment next door to Superdeath's parents? - Commodore |
