Google begins training DeepMind AI to dominate players in Starcraft II

Time and again we have seen machines triumph over humans in various competitions, such as when IBM's Watson beat the best Jeopardy players of all time. More recently, Google's DeepMind AI defeated the world's best Go player. Now AI researchers are setting their sights on StarCraft II, one of the most popular multiplayer games of all time, Wired reports.

This comes with Blizzard's blessing, as it has just released the tools necessary for AI researchers to build bots capable of playing against human players. In a partnership with DeepMind that was announced last year, Blizzard is also opening a cache of data from 65,000 past StarCraft II matches that could prove instrumental in training bots to compete. And that's just the beginning—Blizzard and DeepMind will add around half a million games to the cache each month.

Why focus on StarCraft II? While popular in the esports scene, StarCraft II also provides a bigger challenge for machine learning research that goes well beyond previously conquered games. For example, games like Go and chess are so-called perfect information games—the AI bot can see all of its opponents moves and calculate different outcomes. On a Go board, the number of valid positions is a 1 followed by 170 zeroes. But in StarCraft II, researchers estimate you'd need to add at least 100 more zeroes to account for its complexity.

"It's a big step up," says Oriol Vinyals, a DeepMind researcher working on StarCraft. "This game will require us to innovate in planning, memory, and how we deal with uncertainty."

While StarCraft II may seem like an easier game to conquer compared to Go or Chess, the additional challenge comes from not being able to see everything. The bots that compete will only see and do the things human players can see and do, so there won't be an unfair advantage. This will force bots to try and anticipate player strategies, and to develop their own styles of play.

Seeing an AI compete against a human in StarCraft II is interesting in and of itself, though there are bigger goals and potential outcomes here than just fun. Google has used machine learning from DeepMind to cut cooling costs in its datacenters. By developing AI that can one day defeat humans in StarCraft II, there is hope of taking on more complex jobs.

"From a scientific point of view, the properties of StarCraft are very much like the properties of life," says David Churchill, a professor at Memory University of Newfoundland and organizer of StarCraft bot competitions. "We're making a test bed for technologies we can use in the real world."

Researchers have a ways to go before getting to that point. Early results from bots that were trained with StarCraft data shows there is a significant gap between humans and machines.