) and Blizzard released a very full-featured SDK for developing and running bots for Starcraft 2.DeepMind's press release: https://deepmind.com/blog/deepmind-and- ... vironment/
Blizzard's post about it: http://us.battle.net/sc2/en/blog/20944009
AI researchers have been working on Brood War agents for several years, since the engine was relatively primitive and easier to reverse-engineer, but controlling SC2 was basically impossible without resorting to insane hacks, reading the screen, etc. This API provides a very rich set of features (in the ML sense of the word) and an action interface (which can be capped to human-like APM levels!) to make integration simple, as well as a pared-down Linux SC2 engine for simulating games and outputting replays, so you can run it on GPU instances in the cloud.
Dave Churchill of BWAPI fame has already released a reference implementation and some higher-level utilities: https://github.com/davechurchill/CommandCenter. I guess he had early access, being one of the most well-known BW researchers.
DeepMind's own results (at least the ones revealed in the paper they just published) are very poor so far. My favorite excerpt:
Which just means this is a perfect time to get in on the ground floor of this problem if you're interested in ML and AI!Unsurprisingly, none of the agents trained with sparse ternary rewards could develop a viable strategy for the full game. The most successful agent based on the fully convolutional architecture without memory managed to avoid constant losses by using the Terran ability to lift and then move buildings out of attack range.
