AlphaStar AI SC2 News

[Jaeger]

In the last few days, there have been major news regarding AlphaStar, the DeepMind AI which was trying to learn SC2. Now, the AI view the game through a camera like a human would, and has had its APM reduced even further. By playing on the ladder, it has reached the rank of grandmaster, and it has been better than 99.8% of all the human players.

I just wanted to share this information, and I also had a question. How impressive is a rank of GM, and its ability to beat 99.8% of humans? How many players are between 99.8% and the top player?

@Goodspeed
@Diarouga

[Goodspeed]

GM is top 200 of the region. They probably played in all regions, I'm curious if they reached GM in only the US/EU or in Korea as well (the US ladder is relatively easy, Korea relatively hard). Also the skill difference between the worst GM and the top players is pretty big (nearing or at 100% win rate). So did it reach top 100, top 50? They didn't share that iirc.

Depending on the APM restriction, which they didn't release details about yet afaik, this is either news or something we already knew. If the restrictions are good enough, the AI won't be able to rely on mechanics to win games, but I have a feeling it still does. It seems unlikely to me that the AI would actually outsmart GM players without significant changes to their method/algorithms or a significant amount more practice time.

Keep in mind that limiting it to say 300 APM still means that it will win most games through mechanics, since its actions are much more efficient than any human's. They would have to limit it to 150-200 or so to account for that.

It's still very much a more impressive feat to reach GM with all races than the PvP AlphaStar they showed in the exhibition, which only knew one build order. It's something else entirely to beat humans in extended series though (with the same agent each game, that is), because they will always find ways to exploit weaknesses. For example, the AI doesn't scout enough to effectively adapt. @blackwidow actually sniped it once on ladder with some all in rush.

[RefluxSemantic]

Sc2 is a hard game. Supposedly if you're good at macro you can actually get to Masters (top 4%) by just building SCVs and Marines and nothing else but doing it well. So basically 96% of the players can be beaten by pure mechanical skill. Assuming that they didn't limit the AI too much, reaching top 99.8% is not that impressive, because something like 300 APM (effective, precise robot apm!) already gets you in at least Masters. That's what makes it very hard to tell how much it is accomplishing. It's clear that the AI can function reasonably well in this setting, but it is hard to tell how strategically advanced it is. Currently it seems like Alphastar has somewhat mediocre strategical understanding backed up by very good mechanics. With the limitations they have set up it is also still much worse than the actual top players.

[blackwidow]

Alphastar only played on EU, and reached top 50GM with its protoss and terran (protoss ~top30), and top100 with its zerg.

From what I've seen the protoss agents are actually doing interesting strategies, reacting properly to scouted tech and in general they look like humans much more than the terran and zerg agents.
I remember it scouting for a specific building by going through the opponents base, on the mini map you can see the size of a certain building. Alphastar recognized that the size of the dot on the mini map was likely a spire, and quickly started up the counter to spire-units. Then it checked (by actually moving the camera over there) to see if the dot on the mini map truly was a spire. After seeing that it wasn't (it was a static-defense building with the same size), it instantly cancelled the counter units Alphastar had just started producing. All of this happened over the course of ~3 seconds.
I thought it was pretty cool as this kind of reveals that Alphastar was really looking for stuff, and reacting properly to what it was seeing. I think it could've gotten top 10 at least if it continued playing (it had over 80% winrate)

The terran however, really only got carried by mechanics, I've see a bunch of its games and its pretty silly. The only games it wins are where the opponent does nothing special, and just sits there macro'ing. In these games alphastar wins by being the better mechanical player and just showing up with a bigger army. Any cheese/all in or special build order and it straight up dies looking very silly. Like gs mentioned I sniped the terran one and did an all-in (knowing I was playing alphastar after seeing its reaper control and chat), and it just reacted horribly. Also it refused to leave the game for a long time, lifting all of its buildings going everywhere. (Here is the game for those interested

The zerg knows one build and its really powerful, but it's not smart whatsoever. It just sits there executing a very optimized all in, it doesn't scout and it doesn't react. The build order is strong enough to reach top 100GM on its own. But it couldn't get higher than that. And Alphastar couldn't adapt by doing different things as it literally just knows to do that one thing (it doesn't even know how to counter air units).

So all in all, the protoss alphastar is looking sick (not just carried by mechanics), the terran and zerg are carried to their GM spot solely by their mechanical ability.

[RefluxSemantic]

So did the Terran and Zerg agents just have less training?

Or do we finally have proof that Protoss is the easiest race in sc2?

[Goodspeed]

It is interesting that it does better with P. Maybe it's because P can get more cost-effectiveness out of (near) perfect micro than the other races. Or maybe it's because it didn't find a good all in build with P, like it did with Z, and was forced to actually L2P. The Z agent seems to have tunnel visioned on this 1 build because it worked so well.

[blackwidow]

It is interesting that it does better with P. Maybe it's because P can get more cost-effectiveness out of (near) perfect micro than the other races. Or maybe it's because it didn't find a good all in build with P, like it did with Z, and was forced to actually L2P. The Z agent seems to have tunnel visioned on this 1 build because it worked so well.

I think the protoss agent had more development going into it as they started out with protoss only.

But you're right it could be that it didn't find any "execute this and win" builds like it did with terran and zerg. This would kind of make sense as with Z you can get really high with any midgame timing, because the macro is just so hard to perfect. Constantly making probes is much easier, so its hard to gain an advantage over its opponent by doing this perfectly. Same with terran, the macro aspect is much harder, so its gonna be doing that significantly better than humans.

So it would have to try different things to be better than human p's, leading to it exploring more and more builds.