> StarCraft is closed source, making it quite challenging to run simulations. One of the techniques used by AlphaGo is reinforcement learning, which involves a huge amount of simulation. In order for DeepMind to overcome this limitation, it’s likely that novel abstractions of the state-space will need to be developed.

For their atari work, deepmind was using raw pixel data from the screen. You could do something similar for starcraft. I suspect the actual machine learning question is 'given a sequence of n frames of the game, where should i click next? repeat.'

A training set might look something like video of a game with a text file recording X,Y positions for every click and a timestamp that can be synced to the video. They need lots and lots of those.

From there things get drastically more complicated. To go from very unstructured data like that and form higher level representations that can be used for long term planning and strategy... That would be quite a feat.