A human player finally beat DeepMind's AlphaStar AI at StarCraft II

Cal Jeffrey

Posts: 4,155   +1,416
Staff member
Recap: AlphaStar, DeepMind's StarCraft II playing AI, has finally been beaten by a human. Pro SCII champ Grzegorz “MaNa” Komincz prevailed in a live-streamed head-to-head match against a new version of the AI that more closely mimics a human player's "economy of attention." MaNa found a weakness and beat the AI for the first time. AlphaStar still maintains a 10-1 record.

After humiliating the world’s best Go players, AlphaGo retired, and DeepMind looked for other challenges for its AI like how to diagnose eye disease and how to make lip-reading experts look like fools.

It was not done playing games though. DeepMind’s AlphaZero taught itself chess from scratch, which only took about four hours. After beating the world’s best chess-playing program, Google and Blizzard began teaching it how to play StarCraft II.

The DeepMind AI derivative named AlphaStar went head-to-head with SCII pros Dario “TLO” Wünsch and Grzegorz “MaNa” Komincz and soundly crushed them with each losing five out of five games back in December.

TLO claimed a disadvantage since the match used the Protoss race, which he is not too fond of and admitted lacking the skill to use effectively. He was swept in a five-game series.

Acknowledging that AlphaStar had the advantage against TLO, the DeepMind team decided to pit the AI against Protoss expert MaNa during the second round. The twice champion of major StarCraft II tournaments also fell to disgrace after being shutout 5-0.

DeepMind admitted that one advantage that AlphaStar had over the human players is the ability to see the entire map at once, although it still cannot see through the Fog of War. In other words, it can pay attention to its own base and enemy units and movement without having to shift focus as a human player does. This “zoomed-out” view allows the AI to make strategic decisions more effectively.

At the end of the December event, DeepMind decided it wanted to train a whole new agent that did not have the benefit of the overhead view. Instead, it would have to “decide” where to focus its attention just like a human. It only took the AI about a week to become proficient in the new method. Of course, one week of practice for AlphaStar is equivalent to 200 years of play time for us mortals.

The puny humans got a chance to redeem themselves on Thursday when MaNa again got to face the new AlphaStar agent in mortal StarCraft combat. DeepMind and Blizzard streamed the live match. AlphaStar started out strong, but the StarCraft pro was able to recognize and exploit a weakness in the new AI and pulled off a win against the machine.

You can watch the live match above, which starts at about 2:31:07 for your convenience. The two hours before the game shows highlights from the December beatings AlphaStar doled out. DeepMind also explains quite a bit about how the AI works, which is quite interesting if you feel like rewinding.

If you play SCII and would like to analyze AlphaStar’s butt-kicking strategies for ideas, you can find videos for all the AI/Human matches at the DeepMind website.

Permalink to story.

 
In RTS it’s not all about strategy, a big factor is the speed of the commands.
If 2 players with same skills play against each other and the one use a normal mouse with 2 buttons but the other use a mouse with 12 buttons, that who has the more shortcut buttons will win most of the times.

The neural networks maybe they can efficient answer queries about a data set in which they have trained but they can’t respond to action commands.
They can recognize a butterfly but they can't draw a butterfly under command.
 
That's impressive as hell that Komincz was able to discern an AI weakness in a relatively short period of time.

No doubt the AlphaStar agent "learned" from that experience, and that's the last time for that weakness exploit.
 
That's impressive as hell that Komincz was able to discern an AI weakness in a relatively short period of time.

No doubt the AlphaStar agent "learned" from that experience, and that's the last time for that weakness exploit.
True but do we really want unwinnable AI in games/sports? Its good to drive the point that AI can surpass humans in every field but thats already a given now.
What we really need is GTA with neural net bots who behave and act like real humans. Maybe in a decade it might happen.
 
That's impressive as hell that Komincz was able to discern an AI weakness in a relatively short period of time.

No doubt the AlphaStar agent "learned" from that experience, and that's the last time for that weakness exploit.
True but do we really want unwinnable AI in games/sports? Its good to drive the point that AI can surpass humans in every field but thats already a given now.
What we really need is GTA with neural net bots who behave and act like real humans. Maybe in a decade it might happen.

It's not so much that AI can beat humans as much as it is the fact that they aren't limited by physical restrictions. Humans don't have direct access to all the information and instant control of the computer. We have to have to rely on slow analog processes to first get said information and then have to process it on the fly with the non-specialized hardware that is the human brain. It's not such a fair comparison because of the inherit advantages an AI posses in that regard.

I do like the idea of using AI in games and whatnot. It would be lightyears ahead of what we are currently using.
 
"DeepMind admitted that one advantage that AlphaStar had over the human players is the ability to see the entire map at once. In other words, it can pay attention to its own base and enemy units and movement without having to shift focus as a human player does. This “zoomed-out” view allows the AI to make strategic decisions more effectively."

I was under the impression that it means the AI could see entire map and enemy units but its limited by Fog of War as mentioned in other sites. That's a pretty drastic distinction which is not made clear in this article.
 
The thing I like about the AI is that for people who want challenging gaming experiences without having to deal with human opponents.

You just need to look at games like League of Legends and Heroes of the storm which have very poor AI and see that many players would probably prefer to play against a difficult AI. That way if they get beat they can only blame themselves.

There are also many story driven games that would greatly benefit from this AI. Can you imagine a new ALIEN game using the Deepmind AI? Or Resident Evil? It would make for some pretty amazing gameplay
 
I am so glad it is sorted out, had lost many nights' sleep because of it ...

On a more serious note - I thought the South Koreans were the bomb when it came to Starcraft - at least for starcraft 1, was like a national sport there at some point. Oh well ...
 
"Economy of attention" - so basically they made it behave more like a human without using the multitasking ability it has of a computer. No wonder it lost!
 
That's impressive as hell that Komincz was able to discern an AI weakness in a relatively short period of time.

No doubt the AlphaStar agent "learned" from that experience, and that's the last time for that weakness exploit.

Thats actually a major problem for neural network AIs: only the best of each generation of AIs get to move on; there's no learning from the loosing AIs as they get culled from the "population". Likewise, there isn't a clean way to un-ingrain bad behaviors that are inherited early in development.

The AI makes a LOT of tactical mistakes. I'm not a SC player, but some things the AI is bad at that even I can see:

*)Not walling off the mineral line (AI gets around this by over-producing workers, which is cost-inefficient)
*)Being too aggressive going up ramps (Looses more units then necessary)
*)Reacting with ALL units when any base is under attack (This is what cost the AI the game, as the AI was prevented from attacking)

From a pure tactical perspective, the AI is Silver tier at best. What makes it beat pro's is because of it's perfect micro and insane APM. Tactically, it still has a long way to go.
 
Did google and blizzard teach the AI on the possibility of time travel?

if the AI can calculate the probability/possibility of some things, I wonder what its responses will be if asked the following:
1. time travel.
2. extra-terrestrial aliens.
3. time frame of human space exploration: planetary and beyond.
4. atlantis location.
5. real builders of the pyramids.
 
Did google and blizzard teach the AI on the possibility of time travel?

if the AI can calculate the probability/possibility of some things, I wonder what its responses will be if asked the following:
1. time travel.
2. extra-terrestrial aliens.
3. time frame of human space exploration: planetary and beyond.
4. atlantis location.
5. real builders of the pyramids.

1. Yes; relativity pretty much confirms its possible, at least to go forward.
2. Yes
3. 50 years, 500 years
4. Likely the greek islands, specifically near or on Crete, since its based off the destruction of the Minoan civilization
5. The Egyptians
 
Let the AI play some top korean players, it will be smashed to pieces.
In short term, yes most likely but the thing is every defeat strengthens the neural net as it learns more strategies and the underlying patterns of build and counters. I predict within a year it will get so strong that Blizz or Google may host a human tounament with a $1-$10 entry fees (to keep away abusers) and the opponent will be Deepmind and prize will be like $100k or even more. It will be a best of 5 and I doubt if anyone will prevail against AI at that time. This will be a great news story and publicity for both Blizzard and Google.
 
In short term, yes most likely but the thing is every defeat strengthens the neural net as it learns more strategies and the underlying patterns of build and counters. I predict within a year it will get so strong that Blizz or Google may host a human tounament with a $1-$10 entry fees (to keep away abusers) and the opponent will be Deepmind and prize will be like $100k or even more. It will be a best of 5 and I doubt if anyone will prevail against AI at that time. This will be a great news story and publicity for both Blizzard and Google.

Except AI is trained with datasets not datapoints. This strategy would never work as you couldn't have it begin to replicate nor develop strategies or counter-strategies based on single matches, even over time.

AL and ML algorithms need to be trained with large, clean datasets - someone's entire streaming record, stripped of non-game play moment - not single matches. Adapting on-the-fly, or even after a single match, is what is referred to as "self-teaching", and still only exists as concept in the AI/ML research fields (not even a theory, with the kind of mathematics that goes along with theories)
 
Back