The thing with programs like Deepmind is as soon as it understands the rules, it only takes a few iterations (human intervention iterations) before they become superhuman.

My bet is Terran. Mostly because perfect micro of tanks and marines seem to be the most superhuman thing with most potential to increase win equity. I mean... how do you defend against someone that can perfectly micro 3, 4, even 5 drops simultaneously?

MODESTY HWAITING

they actually have a bot vs bot stream/tourney every so often and they are terrible, but not made on the scale of this. someone got worker rushed pvp and lost everything without moving from mining.

hopefully ill remember to watch, but understanding is maybe untouchable, but how long will it take to get there? each game will have such a large amount of unknowns that i dont think it can ever fully "understand". will it ever be able to say, army is x size, worker count appears to be y; they are macroing vs army is x worker is z they are allining, or just behind

What you wrote describes AlphaZero, but I'm not sure if they are using that or a completely different tool/strategy to tackle StarCraft. Having imperfect information (not to mention an exponentially larger number of variables) probably has a big impact on how much of their previous work in chess/go applies to SC2.

Also worth noting - I think they (DeepMind) want their bots to have to perform under human-like constraints, i.e. a few hundred APM and not 20k. While superhuman multitasking (drops/drop defense) is still within those parameters, I think things like tank micromanagement or superhuman marine splits are not. Even if they were, it would kinda be cheating - the goal is to outsmart humans, not just take advantage of a faster clock speed.

I remarked to a coworker that it would be absolutely hilarious if, after dozens of iterations of training and playing against itself to distill the absolute perfect StarCraft bot, it comes out on stage and shows the world the most unexploitable strategy of all: cannons.

They also program in a 200ms reaction delay to be more human like.

I can't remember how the 5v5 Dota 2 was going but they became unbeatable in 1v1 damn quickly, more quickly than anyone predicted and they did some cool **** no one had thought of before that looked genius.

Hoping to see some of the same.

lol PvP

I'm reading through the TL thread, did it seriously 4 gate mass stalkers? That's what I used to do

There were 10 different "agents" that each did something a bit different. One of those was a weird 4 gate with 2 of them proxied, where it just straight up busted up a ramp vs Mana.

I thought it was two agents - one they trained to play vs TLO, one they trained to play vs Mana (and each of which learned by playing against tons of other agents).

Each set of 5 games involved 5 different agents. The way they were trained was a sort of league where the best agents in each stage continued and were allowed to evolve for the next league. There were 5 agents from the final stage of that training process that were selected each time to play TLO and then after another week of training to play Mana.

Human hope not entirely lost yet as Mana wins the show match against a new version that is more "human" in that it actually uses the camera as a human would instead of seeing the whole map - the initial versions saw everything at once but did still have fog of war, allowing for completely inhuman micro.

Seems to me like all that was proven is that getting to see the whole map at once is a huge advantage.

Almost certain the bot could not see the whole map.

in the intro they mentioned its camera view and how fast it switched, seemed like that could be a big edge and not a maphack type. kinda cool but the parts i watched were pretty boring. somehow they got casters to make it kindergarten casting which was impressive they could keep it up

I watched the whole thing and as I understood it the versions that played the 5 game series were able to see everything that wasn't under fog of war. The developers noticed that while it could see everything it still focused on specific areas so on their visualisation they showed where it was focusing.

In the version that was used for the show match they had a trained a new agent to play using a normal camera. That version was noticeably weaker, although it's not entirely clear how much of that was due to the camera restriction and how much was Mana being able to find a way to exploit a specific tendency.

I think it won't be long before there is a version that is capable of playing with "humanlike" restrictions and beating any human at PvP but there is a long way to go before it's capable of playing well in all matchups.

I don't see it that way. To me it seems more like they have a working setup and now they just need to throw more resources at it to get good at all matchups. And it seems like Deepmind has virtually unlimited computing budget.

wcs groups up and gsl qualifiers up. gsl notables; MC, scarlett (only non korean) and ALL of jinair

DeepMind will be unbeatable, even with human like restrictions, by humans in any three game series in a few weeks if it’s not already.

I actually don’t think ability to “see” is that big of a deal. From what I can tell, some of the top players already do that basically with rapid camera switching to manage multiple battle fields.

I didn't like how both the DeepMind rep as well as the announcers pointed to AlphaStar's lower average apm as an indication of AlphaStar winning on decision making while completely ignoring the dynamic APM indicator showing AlphaStar APMs in the 600s in microintensive big battles and in the 30s in the beginning of the game, thus bringing down the average.

Wasn't there a fight of AlphaStar's mass stalker against Mana's immortal-heavy army where AS came out even, or ahead, based on superior micro?

That's the sort of thing that seems like BS. If AlphaStar's micro takes SC2 to the level where immortals don't even counter stalkers that hard, it's not playing the same game that us humans are playing.

Its not unheard of for humans to spike APM as high or nearly as high as AlphaStar was. Off the top of my head, there was a famous-ish game between Sheth and a Korean protoss in WoL (Puzzle?) where the toss did a blink all-in and was hitting 700 APM on the engagements easily. Not to mention Mana's/TLO's APM were also spiking quite high in the games as well. And realistically, the only reason 30 looks odd early is because humans spam. I guess it's a little disingenuous of the developers if it was programmed/taught that to "dodge" the spirit of the APM cap, but it's really not that weird.

It's blink micro was impossibly perfect, but most of it also happened before Mana really got more than like 3 immortals. What was way more impressive IMO (and also way more problematic) was how it later manuevered and microed on like 3 different fronts when Mana moved out in that same game. It was brilliant positional play, but completely impossible for a human.

But of course, all of this pales in comparison to the real problem.

They should be able to fine tune the human realistic elements as they go. It's a lot harder to do in SC2 compared to Dota.

Heh, I see Afreeca s2 listed as a team and wondered what s2 was doing nowadays, not noticing he'd been retired for five years

Reminder, GSL starts tonight, in 7 hours! Hypeeeeeee

First impression is that Group E is the group of death, oughta be lots of fun