Google's DeepMind AI project is certainly one of the most promising out there. The project's official scientific mission is to "push the boundaries of AI", and Google has decided that one of the best ways to accomplish that goal is by teaching it to kick human butt in games - both digital and physical.
AlphaGo, Google's DeepMind Go-playing computer program has already proven more than capable of doing just that. Through a series of "highly inventive winning moves", the AI was able to beat many Go professionals from all over the world.
Still, it seems Google isn't content to rest on their laurels. The tech giant is moving on from the slower-paced, more thoughtful nature of board games to the tense, strategic environment of StarCraft II. In partnership with Blizzard, Google is looking to turn StarCraft II into DeepMind's own personal AI training playground of sorts.
To accomplish this, Google and Blizzard are releasing a set of tools called "SC2LE", which will "accelerate AI research" in the game. These tools include, among other things, a machine learning API from Blizzard, an open source DeepMind toolset (PySC2) and a large dataset of "anonymised game replays" - which currently sits at 65,000, but is expected to reach well over half a million in the weeks following this announcement.
These tools are completely open to the public, so feel free to get involved if that's your thing.
DeepMind will undoubtedly face a number of hurdles before it will become competent at playing StarCraft II, but one of the biggest challenges lies in the game's Fog of War mechanics.
Fog of War will obstruct the battlefield for both players and AI alike, forcing the AI to explore to obtain the information it needs to defeat its opposition. This is one obstacle DeepMind hasn't had to face just yet - after all, board games like Go provide you with all of the information you need to succeed at any given time.
DeepMind's ability to play the game is still far from that of a professional-level StarCraft 2 player -- it can't even defeat the game's easiest built-in AI yet -- but that might just change a few years from now.