back
10 comments
Wouldn't DeepMind have an unfair advantage, since the best human players are restricted by how fast they can use the mouse and keyboard (measured in actions per minute / APM), whereas DeepMind can easily outperform a human in APM?
So far this hasn't mattered much because it's a very difficult game to play.

Honestly, if it can play top tier StarCraft matches, I'm willing to be we could apply it to the vast majority of things humans can currently do; although I'll rule out creative tasks for the time being.

I think so. If you didn't limit the APM of the AI then I don't think it would be very hard to get it to beat the best broodwar players in the world.

So many aspects of SC:BW are currently balanced around extremely micro intensive unit matchups (mutalisk versus marine and medic, mutalisk versus templar/archon, lurker versus marine, science vessel emp versus arbiter) and so many units and abilities are actually borderline unusable at a competitive level because players don't have enough time to spare (the dark archon's feedback abilitiy is on paper probably one of the most powerful spells in the game but virtually never gets used because the micro opportunity cost is too high, medics optical flare is another good example, on paper it is an extremely useful ability but has only been used to good effect once or twice in pro games) that once you get a half decent AI playing the game won't resemble anything played by humans.

Hell, even worker rushes might be unstoppable in the hands of an fairly rudimentary SC AI. Timing attacks like the bamboo, JoyO, and bulldog coupled with a perfect worker rush backup would so radically alter the pool of viable strategies that we would essentially be starting from square one in terms of build orders.

I call massive bullshit until I see it. StarCraft was one of my favorite games ever due to good design of course but also all skills required to be good at it. There's literally so many ways to approach an AI with humans stomping them in hilarious ways basically every time.

They'd benefit best by categorizing attack strategies, counter strategies, aspects/units on map, and so on. The DeepMind system would do all that in a way that converts raw data into small, actionable information. Traditional methods could do pathfinding, buildup, time-limited decisions, micro-ing, and so on. A hybrid system seems best here.

They're not just going to throw DeepMind at it and beat good humans. We'll find its patterns, then outsmart it. Especially the pro's.

Competitions with bots and humans below:

http://webdocs.cs.ualberta.ca/~cdavid/starcraftaicomp/report...

Great survey of the problem domain and various techniques here:

http://webdocs.cs.ualberta.ca/~cdavid/pdf/starcraft_survey.p...

EDIT to add:

I suggest a title change along the lines of "DeepMind will try to beat StarCraft next." That's more accurate in terms of both reality and original article.

>We'll find its patterns, then outsmart it. Especially the pro's

And then it will learn our patterns and use them against us. AlphaGo for example, uses the patterns of millions of matches and builds on them.

Go and Atari games aren't anywhere near as complicated as a RTS. Plus, players create new patterns on the fly. Even if it wins, then it won't have beaten us by far any more than it does with Go. I didnt have to see a million pro games to win at StarCraft or any other strategy game. I improved with thought, some tips from pro's, and limited amount of experience.
Why do you think the results of competitions with toy AIs can tell you anything about DeepMind?
They showed that a complex game with complex, adaptive opponents working in real-time with limited info doesnt work with any known technique. That means my position is the status quo until DeepMind shows otherwise by toppling amateur and pro human players without full map visibility.
I have a feeling that it would just have one main strategy and micro it's way out of any situation. Imagine a computer version of polt or marineking.
Think the title means to say 'starcraft'.

Interesting how deepmind started with atari, moved to go, and is staying on mastering games.