Skip to main content

Google DeepMind's next gaming challenge: can AI beat StarCraft II?

Google DeepMind's next gaming challenge: can AI beat StarCraft II?


Blizzard and DeepMind are releasing an open research environment

Share this story

DeepMind, the British artificial intelligence company that's now a part of Google, stunned the world back in March when its AlphaGo AI defeated world-champion Go player Lee Se-dol to achieve one of the most sought-after feats in computer science. The ancient Chinese board game of Go was considered the most difficult "perfect information" game for computers to crack, as the dizzying complexity produced by its simple rules requires a highly advanced degree of intuition to play at a serious level, but AlphaGo managed it with a revolutionary system built on neural networks and machine learning.

Now DeepMind is turning its attention to a game that will pose an even bigger challenge: StarCraft II. Blizzard Entertainment's real-time strategy hit is one of the most fiercely competitive games played professionally around the world, and the company is working together with DeepMind to release it as an AI research environment. The announcement is being made at Blizzard’s annual conference, BlizzCon.

"StarCraft is an interesting testing environment for current AI research because it provides a useful bridge to the messiness of the real-world," says Oriol Vinyals, a DeepMind research scientist who was once the top-ranked StarCraft player in Spain. "The skills required for an agent to progress through the environment and play StarCraft well could ultimately transfer to real-world tasks." Creating strong StarCraft AIs would obviously also have major implications for the game itself, with the potential for customized virtual coaches to raise the level of play across the board.

StarCraft is a complex strategy game that involves resource management, scouting, and battle tactics. Unlike Go, players are unable to see the entire field of play at once, meaning they have less information to work with when planning their moves; it’s a test of memory and adaptation as much as it is long-term planning. "The thing about Go is obviously you can see everything on the board, so that makes it slightly easier for computers," DeepMind founder Demis Hassabis told me earlier this year, in an interview that Blizzard executive producer Chris Sigaty says "got some discussions going" between the two parties.

StarCraft II’s present AI system operates on completely different principles to DeepMind’s, according to Sigaty — it’s hand-crafted and designed to work at different difficulty levels. It also "cheats" a little, because it has access to a degree of information that a human player just wouldn’t be able to attain; for example it’s able to issue commands to all units at any time, even ones that are "off-screen" from its point of view. The challenge for Blizzard is to create a StarCraft AI player that is stronger than a human while being bound by the same limitations.

'StarCraft II' is an unprecedented challenge

The game is an unprecedented challenge for DeepMind, too. While the company has taught AI to play simple Atari games before, Vinyals says he relishes the chance to work with a modern, more graphically intense release. There are "super exciting opportunities for pushing AI research forward thanks to the problems that StarCraft II poses," he says.

To that end, Blizzard and DeepMind plan to release an environment in the first quarter of next year that will be open to any AI researchers who want to contribute. A new interface will simplify the graphics of StarCraft II into basic visuals that are better suited to feeding into machine learning systems (see video below), and an API will allow for similar functionality to previous bots that read from game data. Blizzard will also release replays from StarCraft II games to give AIs a dataset to learn from. The aim is for bots to play from visual information, just like humans do.

Blizzard and DeepMind both caution that we’re far away from the day where an AI will be able to defeat a top-ranked StarCraft player. Neither would offer a rough date when asked, though it’s worth pointing out that AlphaGo’s victory came a lot sooner than most observers predicted. "I think pros will lean into this and take away things that we’re actually not saying," says Sigaty, who believes that the reaction among the community will be positive overall. But however long it takes for an AlphaGo-style showdown, the foundations have been laid for researchers to start their work, and StarCraft II should provide a fascinating platform for the field at large.