Google DeepMind takes on StarCraft II

[HEXUS]

The AI researchers take a big step beyond board games like Chess and Go.

Read more.

[shaithis]

"we're still a long way from being able to challenge a professional human player at the game of StarCraft II"

...and then we have to take it to another level to beat Koreans!

[semo]

Would the AI be able to bank/throttle actions in order to do bursts? You need to consider the latency between actions. Ordering a new unit can take only 2 actions but the single action of selecting a point on the minimap can take a lot longer (unless you count moving the cursor with your mouse an action). For the AI it would probably take the same amount of time.

[LSG501]

Only concern is this.... isn't this basically training AI in war tactics and battlecraft...

[TheAnimus]

Would the AI be able to bank/throttle actions in order to do bursts? You need to consider the latency between actions. Ordering a new unit can take only 2 actions but the single action of selecting a point on the minimap can take a lot longer (unless you count moving the cursor with your mouse an action). For the AI it would probably take the same amount of time.

That's very easy to limit, you can draw up a simple function that replicates how a user behaves, and make it so the AI 'knows' about that delay too.

[TooNice]

Only concern is this.... isn't this basically training AI in war tactics and battlecraft...

That's what I was thinking too

[Lanky123]

Would the AI be able to bank/throttle actions in order to do bursts? You need to consider the latency between actions. Ordering a new unit can take only 2 actions but the single action of selecting a point on the minimap can take a lot longer (unless you count moving the cursor with your mouse an action). For the AI it would probably take the same amount of time.

That's very easy to limit, you can draw up a simple function that replicates how a user behaves, and make it so the AI 'knows' about that delay too.

Yep - modelling the human input behaviour is the least of the problems here.

It astounds me that they can reach any sort of strategy at all given the delay between pertinent action and reward in a game like this. I guess this is where the 'curriculum scenarios' come in - to give some sort of a baseline strategy from which it can then extrapolate/explore.

[kalniel]

Only concern is this.... isn't this basically training AI in war tactics and battlecraft...

Well it's nice to know that next time we're invaded by the Zerg we've got a head start

Third party AI is great - means developers shouldn't have to worry about the hugely expensive task of trying to develop AI for their games individually. Just make your game player-controllable and set the AI to learn from the human interface.