Watch Deepmind Play StarCraft 2 at Noon

AlphaAtlas

[H]ard|Gawd
Staff member
Joined
Mar 3, 2018
Messages
1,713
DeepMind, Alphabet's subsidiary focused on neural-network bases game AIs, has been steadily training its titular AI to play a variety of games. It's beat the world's best human Go player in 2016. and achieved human-level performance in Quake 3 last year, while the company announced loftier aspirations late last year. Now, following up on their previous progress with Starcraft 2, DeepMind is scheduled a stream a StarCraft 2 "demonstration" at noon central time today. KitGuru, who spotted the stream yesterday, notes that Deepmind was beating bots on insane difficulty with a 50 percent success rate last year, so we'll see how quickly this AI learns.

Check out the stream here.

Join Artosis, RottterdaM and a cast of special guests for a unique StarCraft II showcase live from DeepMind in London, in partnership with Blizzard.
 
  • Like
Reactions: dgz
like this
Now lets watch Deepmind play World of Warcraft Battle For Azeroth. It seems to be just grinding all day long and playing pet battles.
 
Now lets watch Deepmind play World of Warcraft Battle For Azeroth. It seems to be just grinding all day long and playing pet battles.

I fully expect to see something like Legion's gamer profile in the news someday:

Gamer Profile for Infiltrait0rN7

Galaxy of Fantasy:
Most Used Character: John Smith, Level 612 Ardat-Yakshi Necromancer
Group Affiliation: N/A
Most Recent Boss Defeated: K'l'rh, Rachni Blood Wizard
Awards:
- Best Supporter/Healer (Event: Scourge of the Thresher Dragon)
- Best Unit Efficiency (Event: Return of the Cyber-Protheans)
- Winner (Event: Crystal Genophage Elimination Platinum)
Infractions:
- Suspected use of VI play assistance (direct control of twenty-seven pets without use of behavior macros); challenged and overturned
- Suspected use of VI play assistance (reaction time better than possible for organics); challenged and overturned
- Suspected use of hacking for direct server access (tactics better than possible without knowledge of underlying code behavior); challenged and overturned
- Unsportsmanlike behavior (taunting during Crystal Genophage Elimination Platinum); accepted 3-day account suspension

N7 Code of Honor: Medal of Duty:
Player Score: 15,999,999,999 (max)
Most Preferred Class: Sniper
Least Preferred Class: Melee
Sniper Rifle Kills: 200,917 since last server reset
Shotgun Kills: 3

Grim Terminus Alliance:
Award: Abolitionist (Complete full playthrough without any slave kills, free all slaves encountered)
Award: Cure for What Ails You (Kill 100+ quarians)

Geth Attack: Eden Prime Fundraising Edition:
Donation Level: Ultra Platinum
Player Score: 0 (Purchased but not played)

Fleet and Flotilla: Interactive Cross-Species Relationship Simulator:
"Based on the Bestselling Vid!"
Playtime: 75 hours, 6 minutes
Player Score: 15 (Hopeless)
 
So is this another one of those demonstrations where the AI gets to see the entire battlefield or will it actually be under normal playing conditions?
 
I personally have playd one of the best ai's back in 2014 for Starcraft 2. (was master level at the time)
It was really easy and not worth it as training even.

4 years later and lets see what they can do.
 
So is this another one of those demonstrations where the AI gets to see the entire battlefield or will it actually be under normal playing conditions?


hopefully.. but..


this is one area i would love to see to have a "cloud" option. Gaming AI

no matter how awesome of a machine we may have at home, if gaming AI could be offloaded to a cloud service, it would be a whole new ballgame.

I would gladly pay for such a service
 
I personally have playd one of the best ai's back in 2014 for Starcraft 2. (was master level at the time)
It was really easy and not worth it as training even.

4 years later and lets see what they can do.

There's been a move on the AI front; the way it used to be done was to try and code behaviors into the AI and just rely on the AI being faster. Nowadays, there's been a move toward neural networks, or basically just having the AI play itself over and over until it optimizes its strategy. The only downside to neural networks is that "bad" behaviors that are learned in very early generations can persist unless you have some mechanism to occasionally purge behaviors learned in early generations.
 
There's been a move on the AI front; the way it used to be done was to try and code behaviors into the AI and just rely on the AI being faster. Nowadays, there's been a move toward neural networks, or basically just having the AI play itself over and over until it optimizes its strategy. The only downside to neural networks is that "bad" behaviors that are learned in very early generations can persist unless you have some mechanism to occasionally purge behaviors learned in early generations.

Most AI stuff is just plain stupid.
But so far, Color me impressed with this one so far.

TLO is a noob as protoss for sure, not really that good as his main (zerg) either :p
 
Thank you so much for this. Can't wait for Maru vs an AI. Artosis commenting lol

Edit: Oh my, this is AI vs actual pro players.
 
Go Alpha!

Also, I'm not an SC2 player, but even I can tell that AI's micro is disgustingly good.
 
So far it looks like mana can't beat the ai. Alpha star was wiping him the first two games. Funny part was in game 3 when they start talking about the value of building a wall. How impressed they were it did a wall an changed strategy, how walls keep your enemy out , etc. I wondered if some poor democrat was triggered.

You could make many bad jokes like that even an ai wants the wall etc
 
They mentioned that the trained AI runs on a single, consumer desktop GPU, which is interesting.
 
After watching several games, I am convinced. These agents/bots will be winning tournaments in a couple of years
 
are there rules around this though? Like does it have to emulate mouse movements and clicks, as well as brain latency?
 
10-0. Can't say i'm surprised since the AI can easily micro many units at once which is a huge advantage.
 
are there rules around this though? Like does it have to emulate mouse movements and clicks, as well as brain latency?
10-0. Can't say i'm surprised since the AI can easily micro many units at once which is a huge advantage.


Funny you mention that. I'm just listening instead of watching now, but they mentioned they made a separate AI that only uses the same "window" a human sees. But all the AIs are limited to the same number of clicks/key presses per minute a pro human is capable of.

They're showing off that AI now...
 
Funny you mention that. I'm just listening instead of watching now, but they mentioned they made a separate AI that only uses the same "window" a human sees. But all the AIs are limited to the same number of clicks/key presses per minute a pro human is capable of.

They're showing off that AI now...

I get what you're saying but whether or not the AI can control only one screen. They can think quicker and make better decisions instantly. RIght now they are showing the player's view and breaking it down. Why can't they show the AI's view so we can see what they are doing specifically from their pov.
 
10-0. Can't say i'm surprised since the AI can easily micro many units at once which is a huge advantage.

Just watch the APM. It doesn't initiate more actions than a human does.
 
Just watch the APM. It doesn't initiate more actions than a human does.

AI was able to macro/micro (?) control at least 3 groups of attacking units at the same time, this is a huge advantage but yeah, AI is issuing the same number of commands. We can conclude that every commands sent is optimal.
This was very interesting to watch, wish I caught it live at the beginning, it would prevent me from skipping content.

I do not know how the AI managed the visibility angle of this, was it restricted at how fast/accurate it can move the camera ? Those were really nice matches ! Please post more of this ;)
 
I get what you're saying but whether or not the AI can control only one screen. They can think quicker and make better decisions instantly. RIght now they are showing the player's view and breaking it down. Why can't they show the AI's view so we can see what they are doing specifically from their pov.
AI was able to macro/micro (?) control at least 3 groups of attacking units at the same time, this is a huge advantage but yeah, AI is issuing the same number of commands. We can conclude that every commands sent is optimal.
This was very interesting to watch, wish I caught it live at the beginning, it would prevent me from skipping content.

I do not know how the AI managed the visibility angle of this, was it restricted at how fast/accurate it can move the camera ? Those were really nice matches ! Please post more of this ;)
They mentioned that the AI is being fed the entire map visibility at once, so the AI sees everything at once (minus fog of war ). Despite that, they mentioned that when they dug into the algorithm while it was running they found that the AI was "focusing" in on specific areas similar to how the human player moves his camera to specific areas.

are there rules around this though? Like does it have to emulate mouse movements and clicks, as well as brain latency?
They didn't say it was a rule, but they mentioned that when they measured the amount of time between when the AI "sees" something and when it reacts was about 300ms, which they mentioned is generally slower than human players. However, I'm not sure what the range on that is (ex: will the AI react much faster when there is very little to think about like during quick micro operations of an army)
 
AI was able to macro/micro (?) control at least 3 groups of attacking units at the same time, this is a huge advantage but yeah, AI is issuing the same number of commands. We can conclude that every commands sent is optimal.
This was very interesting to watch, wish I caught it live at the beginning, it would prevent me from skipping content.

I do not know how the AI managed the visibility angle of this, was it restricted at how fast/accurate it can move the camera ? Those were really nice matches ! Please post more of this ;)

At some point they said bot is looking at the whole map at once. Maybe if a human were to use a mind reading controller but no, eventually these things will be better than us in every game, even Quake
 
At some point they said bot is looking at the whole map at once. Maybe if a human were to use a mind reading controller but no, eventually these things will be better than us in every game, even Quake

Yeah well aren't they designed for that ;) Now if only they could apply this at broader applications, we would be doomed as a working class.
 
Yeah well aren't they designed for that ;) Now if only they could apply this at broader applications, we would be doomed as a working class.

Well, I mentioned it before, but when the working class goes because of automation, the level of consumers will greatly diminish and cause an economic breakdown for capitalist societies. So I'm curious to see what happens at that point; do we go Star Trek level of neo-capitalism where wealth is more about self/societal enrichment or does everything fall apart because of upper-class's greed and unwillingness to adapt?
 
Just watch the APM. It doesn't initiate more actions than a human does.

while the average APM was quite normal, the peak APM at one point exceeded 1.500. that's around twice as much as i've seen human players peak at. very much extraordinary.
 
  • Like
Reactions: dgz
like this
The AI is getting by a lot due to its superior micro; while I'm not a SC2 player, I see some things the AI isn't doing that does hurt it:

1: Not walling off its mineral line
2: Charging up ramps, loosing more units then necessary
3: Using it's entire army to deal with a few units behind it

The last of these cost the AI a game, as a few units kept the AI busy enough until the human got enough units to win a fight. Humans are still superior tactically, but the AIs unbeatable micro still allows it to compete.
 
They didn't say it was a rule, but they mentioned that when they measured the amount of time between when the AI "sees" something and when it reacts was about 300ms, which they mentioned is generally slower than human players. However, I'm not sure what the range on that is (ex: will the AI react much faster when there is very little to think about like during quick micro operations of an army)

Just wondering if it can also insta-click around the map which would give it a huge advantage, even if it was learning from a crap player like me.
 
Having finally gotten around to watching the whole thing, that was a fascinating look at AI. As a person with both a huge interest in AI and a reasonable skill (Plat league, on the rare occasions I play) and good understanding of SC2, I was gripped the whole time. The insights in to how the AI was trained and how it thinks were also gripping.

Seriously worth a watch for anyone with even a vague interest in either AI or SC2. Some of its behaviour was incredibly Pro like, while other little mistakes it makes would have killed any other pro. I think the strength of not playing like a pro is certainly something that helps it win, but on the other hand it makes a lot of efficiency mistakes.

Man, I want more of this.
 
Yeah well aren't they designed for that ;) Now if only they could apply this at broader applications, we would be doomed as a working class.

Doomed as a working class? Who's going to produce all the shit for those fat AIs? Other bots? Hell no. It will be us again
 
They mentioned that the trained AI runs on a single, consumer desktop GPU, which is interesting.
I think you may have misheard. At about 1 hour 20 minutes or thereabouts, just after mana game 2, they mention it is only taking one consumer level GPU to run when asked what kind of setup the ai is using for the demo. Moments later they say something about when training and they mention using tensor flow clusters which they reference are about equivalent to 50 gpus. I was distracted by eating and didn't quite catch how long the clusters came into play but it was mentioned in reference to the training.

So in theory, any of us could run the ai if we had a second video card to dedicate in our machine and play against or alongside it. Training it might take a while longer depending on how long they ran the cluster for. As you'd need to multiply that time by 25 for the dual GPU users among us.

I think it's interesting though because we'll all probably have 50 gpus power in 12 years or 15 years and can be making AIs perform tasks better than world class human level performance at home in modest amounts of time.
 
This just serves to remind me why I hated starcraft if I had any inclination to revisit it that is now gone for a few years again.
 
Yeah you made my point lol, ever watched the matrix ?

What's your point? Humanity DID work for the machines. A battery is producing energy. See? Work. Productivity!
 
Back
Top