30 October 2019
By Donna Lu
An artificial intelligence can now play the real-time strategy video game StarCraft II so well that it is better than 99.8 per cent of human players.
The AI, called AlphaStar, was developed by tech firm DeepMind, which is owned by the same parent company as Google.
AlphaStar played anonymously against human players in a series of online games on the official StarCraft II game server, Battle.net, and ended up ranked in the top 200 players for each of the leagues it competed in.
StarCraft II is a popular science-fiction game that involves controlling armies and building infrastructure. Players must compromise between short-term payoffs and long-term gain. Using AI to make sequences of complex predictions may be useful for real-world applications such as climate modelling and self-driving cars.
In January, an earlier version of AlphaStar beat two of the world’s top professional players, but at the time it was given special advantages such as the ability to see the entire map of the game.
To level the playing field, this time AlphaStar was restricted to what human players are able to see. This prevented the AI from completing actions in multiple locations simultaneously, said Oriol Vinyals at DeepMind in a press conference.
The number of actions was also restricted, so it wasn’t able to rapidly click at superhuman speed.
Did You See This CB Softwares?
37 SOFTWARE TOOLS... FOR $27!?Join Affiliate Bots Right Away
StarCraft II gamers can play as one of three alien races, each of which have different characteristics. DeepMind created three versions of AlphaStar, each specialising in one of the three races, and trained them on footage of human games.
The three versions then played against each other and a league of other AIs that acted as sparring partners. The three versions were trained continuously for 44 days, equivalent to about 200 years of gaming time.
Each version consisted of a single neural network, a type of AI algorithm, that had to learn to vary its strategies and play style depending on its opponent.
AlphaStar was then pitted against human players online. The three versions all made it into the top league, which ranks the top 200 players in each geographical region. In Europe, this placed AlphaStar in the top 0.2 per cent of approximately 90,000 players.
“It’s an extremely impressive achievement,” says Julian Togelius at New York University. “It should not be understated that this is a very hard game to play well.”
StarCraft II is challenging for AI to master partly because of the sheer number of potential choices a player can make. “Players control hundreds of units at once, resulting in 1026 possible choices for every single move,” said David Silver at DeepMind, also at the press conference.
“I don’t think there is such a thing as a completely fair comparison between human and machine gameplay,” says Togelius. Unlike humans, AIs don’t get exhausted or emotionally upset, or physical ailments like cramping fingers. But under the circumstances, the restrictions on AlphaStar made the match-up as fair as possible, he says.
Journal reference: Nature, DOI: 10.1038/s41586-019-1724-z
More on these topics: