DeepMind's AlphaStar AI now beats 99.8% of Starcraft II players

DeepMind has announced that its AlphaStar AI played and beat human opponents online at StarCraft II's highest level for months.

AlphaStar, the AI trained to play the game, operated anonymously under the same conditions as human players -- meaning it could only see through the game camera, and was slowed down to operate at human speeds. Despite this, AlphaStar AI agents reached the game's Grandmaster level and beat 99.8% of humans on its main battle.net servers.

DeepMind, which previously trained AIs to beat top players in the board game Go, emphasized Starcraft II's immense complexity.

From DeepMind:

There are up to 10^26 possible actions available to one of our agents at each time step, and the agent must make thousands of actions before learning if it has won or lost the game. Finding winning strategies is challenging in such a massive solution space.

To improve the current AlphaStar agents over previous iterations, DeepMind trained them against other bots that didn't just seek to win, but instead worked to exploit the AI's weaknesses. DeeMind's scientists claim to have learned this strategy from the way human players train together.

"In the real world, a player trying to improve at StarCraft may choose to do so by partnering with friends so that they can train particular strategies," the AlphaStar team wrote. "As such, their training partners are not playing to win against every possible opponent, but are instead exposing the flaws of their friend, to help them become a better and more robust player."

DeepMind began work on training AI agents to play Starcraft II just over two years ago, citing the game as an invaluable area to experiment with AI technology that could later be applied elsewhere in life.

"From a scientific point of view, the properties of StarCraft are very much like the properties of real life," David Churchill, a professor and DeepMind advisor, told Wired when Deep mind announced its plans in 2017.

"We’re making a test bed for technologies we can use in the real world."

Related people

Sections