Researchers at Google’s DeepMind AI company are teaming up with Blizzard Entertainment to release an API tailored for AI research environments based in StarCraft II. The company plans to make the interface available to AI researchers and developers next year.
Blizzard’s release of the API will open the field for programmers to create and train their own AI agents to play StarCraft II. According to DeepMind’s Oriol Vinyals, research like this could lead to more interesting AI opponents in video games or AI coaches that help players enhance their own skills.
Opponents in StarCraft II make quick decisions, have limited knowledge of the map, and make moves simultaneously. All of these factors make the game a unique place to expand the abilities of artificial intelligence.
The complexities that make StarCraft a great learning environment also present a fair amount of challenges for AI. Ben Weber, a data science manager at Twitch, broke down some of the hurdles AI agents will have to overcome when DeepMind first announced its interest in StarCraft earlier this year.
In a blog post announcing the collaboration, a representative from DeepMind said these qualities mimic the “messiness” of the real world and could eventually translate into better AI understanding for real-world tasks.
DeepMind has been using games to teach AI agents how to function in complex and sometimes unpredictable situations for a while now. DeepMind AI agents have learned to play Atari 2600 games at the same level as human players, and have also managed to defeat a top Go player at the complex board game earlier this year.
No tags.
