deepmind's alphastar beats pros in starcraft ii

[Uri Blass]

Anyway, I am not familiar with the game, but the argument I read sounded pretty convincing. Also, why pick a game allowing for simple inhuman exploits, doable by a simple bot from 2011? Yes, I understand, notoriety, popularity and databases, but I would think of a game not needing AI restrictions, like the one excellent turn-based I presented. Longish turns, where micro for top players is almost irrelevant and comes almost automatically, without any need of fast clicks. No need to restrict AI in any way. Maybe DeepMind AI can find an exploit even in this game, an imbalance, but it would be of a totally different kind.

They cracked Go, Shogi, Chess and now StarCraft.

I think such a round based strategy game you suggest would be no challenge for Deepmind....

AlphaStar has 200 years of accumulated StarCraft play,
with an round based strategy game, even a complex one,
they could train their agents with thousands of years,
and develop strategies humans did not come up with.

Myself would not be impressed by any round based strategy game Deepmind is able to crack,
but RTS with StarCraft was another kind of challenge.

btw: did the human players had any hardware advantage compared to AlphaStar?

--
Srdja

I am less impressed by starcraft because it is a game that not only I never played and not only I do not know the rules of the game but I also have no idea what is the answer for the following questions:
What is the initial position of starcraft?
How many option does a player have in the beginning of the game?

It is easy to answer these question in chess or go and I guess also in Shogi(never played Shogi but from the little that I read I understand that shogi is basically similiar to chess).

[megamau]

They picked Starcraft because it is the most famous "milestone" game still not conquered.
If you check https://xkcd.com/1002/ and know about Sharp, Alphago, Watson and Libratus it becomes quite clear.

The criticisms of AlphaStar are somewhat valid, but similar to the ones moved towards Alphazero.
DeepMind presents the results in a way that gets the most PR, but all the data is there for proper assesment, and the result itself is impressive even whitout the "embellishment".

There are no previous bots able to play at pro level, regardless of perfect "micro" management. And AlphaStar "micro" is not so outrageous as it's being portrayed. It is faster and more precise than human, but not by orders of magnitude.

[smatovic]

I am less impressed by starcraft because it is a game that not only I never played and not only I do not know the rules of the game but I also have no idea what is the answer for the following questions:
What is the initial position of starcraft?
How many option does a player have in the beginning of the game?

It is easy to answer these question in chess or go and I guess also in Shogi(never played Shogi but from the little that I read I understand that shogi is basically similiar to chess).

I think that is the "another kind of challenge" i meant,
StarCraft is no board-game, so there is no game tree search like AB or MCTS involved,
and therefore i am impressed that Deepmind was able to train neural networks to play RTS.

--
Srdja

[Laskos]

I am less impressed by starcraft because it is a game that not only I never played and not only I do not know the rules of the game but I also have no idea what is the answer for the following questions:
What is the initial position of starcraft?
How many option does a player have in the beginning of the game?

It is easy to answer these question in chess or go and I guess also in Shogi(never played Shogi but from the little that I read I understand that shogi is basically similiar to chess).

I think that is the "another kind of challenge" i meant,
StarCraft is no board-game, so there is no game tree search like AB or MCTS involved,
and therefore i am impressed that Deepmind was able to train neural networks to play RTS.

--
Srdja

Strategy in 20 minutes until the end of the game, with zillion of maximum useful actions involved totally? This is action-fest, and I don't care about "milestones"’, that is more a PR milestone, not a scientific one. I repeat, I am skeptical of AlphaStar strategy and balance in humanly-doable games. The whole finesse of the game strategy might be shortcutted by several "cheats".

Anyway, this "RTS" game is quite crappy for scientific purposes with AI.

[smatovic]

...
Anyway, this "RTS" game is quite crappy for scientific purposes with AI.

Then we should ask the military what they think about AlphaStar?

--
Srdja

[Laskos]

...
Anyway, this "RTS" game is quite crappy for scientific purposes with AI.

Then we should ask the military what they think about AlphaStar?

--
Srdja

The military already have for decades "Automaton 2011" cheat for each specific task, especially superhuman ones.

[Werewolf]

I am less impressed by starcraft because it is a game that not only I never played and not only I do not know the rules of the game but I also have no idea what is the answer for the following questions:
What is the initial position of starcraft?

It varies depending on map, and there are hundreds of map. But each new game has some common elements: you start with a command centre and a handful of workers.

How many option does a player have in the beginning of the game?

Quite a few and as time continues the options quickly increase. Because these options involve both a decision of what to build and where to build and then how to deploy your army and whether to be economic or military and so on....the options are literally incalculable very quickly.

Here's an analogy:

Imagine England and Germany go to war tomorrow. However, suppose neither side has an army and neither side has any money in the bank nor any factories or facilities to build with. But they are at war.
That's like the start position of Starcraft.

Then you have to decide what plan to follow, and how best to execute it. Some people like a military approach ("Quick, build tanks and attack!"); others like an economic approach ("The dollar is greater than the tank"). Both have pros and cons.

It's very rich strategically.

[Werewolf]

I am less impressed by starcraft because it is a game that not only I never played and not only I do not know the rules of the game but I also have no idea what is the answer for the following questions:
What is the initial position of starcraft?
How many option does a player have in the beginning of the game?

It is easy to answer these question in chess or go and I guess also in Shogi(never played Shogi but from the little that I read I understand that shogi is basically similiar to chess).

I think that is the "another kind of challenge" i meant,
StarCraft is no board-game, so there is no game tree search like AB or MCTS involved,
and therefore i am impressed that Deepmind was able to train neural networks to play RTS.

--
Srdja

Strategy in 20 minutes until the end of the game, with zillion of maximum useful actions involved totally? This is action-fest, and I don't care about "milestones"’, that is more a PR milestone, not a scientific one. I repeat, I am skeptical of AlphaStar strategy and balance in humanly-doable games. The whole finesse of the game strategy might be shortcutted by several "cheats".

Anyway, this "RTS" game is quite crappy for scientific purposes with AI.

No I think this is obviously wrong. The game is well recognised as being rich strategically and I'd invite you to try it before rejecting that. As said before there is a lot of micro - which I personally find annoying - but there's also a lot of strategy too and some of it is subtle and quite high level. Against the current A.I it has a clear edge over me in micro and I easily have an advantage over it in thought and planning. I usually score about 80% against the top A.I. so it's not just about executing operations quickly. Also some games can last more than an hour.

[Laskos]

I am less impressed by starcraft because it is a game that not only I never played and not only I do not know the rules of the game but I also have no idea what is the answer for the following questions:
What is the initial position of starcraft?
How many option does a player have in the beginning of the game?

It is easy to answer these question in chess or go and I guess also in Shogi(never played Shogi but from the little that I read I understand that shogi is basically similiar to chess).

I think that is the "another kind of challenge" i meant,
StarCraft is no board-game, so there is no game tree search like AB or MCTS involved,
and therefore i am impressed that Deepmind was able to train neural networks to play RTS.

--
Srdja

Strategy in 20 minutes until the end of the game, with zillion of maximum useful actions involved totally? This is action-fest, and I don't care about "milestones"’, that is more a PR milestone, not a scientific one. I repeat, I am skeptical of AlphaStar strategy and balance in humanly-doable games. The whole finesse of the game strategy might be shortcutted by several "cheats".

Anyway, this "RTS" game is quite crappy for scientific purposes with AI.

No I think this is obviously wrong. The game is well recognised as being rich strategically and I'd invite you to try it before rejecting that. As said before there is a lot of micro - which I personally find annoying - but there's also a lot of strategy too and some of it is subtle and quite high level. Against the current A.I it has a clear edge over me in micro and I easily have an advantage over it in thought and planning. I usually score about 80% against the top A.I. so it's not just about executing operations quickly. Also some games can last more than an hour.

I sure can imagine that it is rich strategically for average and even top humans, I played Civilisation awhile ago, and Total War series quite recently (still, both turn-based). Not my kind of strategy (I am always messing up against average humans with wrong strategies in the developing stages of tech lines and economy, not learning well the manual and its applications on the ground), but nevertheless, yes, rich in abstract decisions. But there were exploits some good players found, later "fixed" by the devs with patches (at least in Total War series). If a bot like this ML NN DeepMind found exploits, the whole strategic subtlety may fall apart. Exploits can be on strategical level too, not just micro. It is possible that the games I presented have strategic exploits, and all the subtle strategy I found in these games is reduced to rule of thumb "withdraw from the Dnepr as fast as possible all your units, keep the flanks and grind the attacks being well supplied" or "push onto the flanks all your units, disregard supply, cut off the enemy in half". That would be boring, but still a bit more insightful than plain precision click-fest (at superhuman rates when necessary).
Am I to disregard this passage?

"MaNa: I would say that clearly the best aspect of its game is the unit control. In all of the games when we had a similar unit count, AlphaStar came victorious. The worst aspect from the few games that we were able to play was its stubbornness to tech up. It was so convinced to win with basic units that it barely made anything else and eventually in the exhibition match that did not work out. There weren’t many crucial decision making moments so I would say its mechanics were the reason for victory.

There’s almost unanimous consensus among Starcraft fans that AlphaStar won almost purely because of its superhuman speed, reaction times and accuracy. The pros who played against it seem to agree."

[Werewolf]

Am I to disregard this passage?

"MaNa: I would say that clearly the best aspect of its game is the unit control. In all of the games when we had a similar unit count, AlphaStar came victorious. The worst aspect from the few games that we were able to play was its stubbornness to tech up. It was so convinced to win with basic units that it barely made anything else and eventually in the exhibition match that did not work out. There weren’t many crucial decision making moments so I would say its mechanics were the reason for victory.

There’s almost unanimous consensus among Starcraft fans that AlphaStar won almost purely because of its superhuman speed, reaction times and accuracy. The pros who played against it seem to agree."

Yes OK, fair point. But if you were to take the superb micro of AlphaStar and import it onto the existing SC2 A.I, it would still get crushed by any decent human.

The reason for this is that micro, like tactics in chess, can only help so much. In chess if the opponent insists on putting his pieces on bad squares he will lose; in SC2 if the opponent doesn't have a viable plan the results are similar.

I think one thing I was impressed by was that AlphaStar won some late game battles which are typically more plan-decided.

But anyway, one thing we can all agree on is that Deep Mind should reduce the operations (clicks) per second of AlphaStar and see how it does then. Then they can improve it from there.

In reality what's actually happened is they have a new AI which is vastly better than the old one in both strategy and micro, but they're only selling the strategy side and people like yourself can see through the marketing BS.