DeepMind, the artificial intelligence division of Google parent Alphabet, has announced that its AlphaStar AI has become the first to reach the top of the StarCraft II league without any artificial restrictions or ability boosts being added to the game.
DeepMind demonstrated its ability to train artificial intelligences to beat the world's best human players at previously-inaccessible games back in March 2016 when it beat go expert Lee Se-dol 3-1 before releasing an updated version which could teach itself to beat the best human players in just three days. The company's success with the notoriously difficult game of go was followed by real-time strategy StarCraft II - but with the AI having a greater insight into what's happening on the battlefield than its human competitors.
Now, the company claims it can beat the world's best players with no unfair advantage. 'AlphaStar now has the same kind of constraints that humans play under – including viewing the world through a camera, and stronger limits on the frequency of its actions,' the company explains in its latest announcement. 'AlphaStar can now play in one-on-one matches as and against Protoss, Terran, and Zerg – the three races present in StarCraft II. Each of the Protoss, Terran, and Zerg agents is a single neural network.'
The training of the AI is entirely automated, the company claims, while the matches that proved its worth took place on the official Battle.net server for the game - ending with the AI being ranked above 99.8 percent of all active players and achieving Grandmaster status across all three in-game races.
The team's work, which it says could be applied to a variety of problem domains,is detailed in the journal Nature, while replays of all the matches - which took place between July and August this year - can be found on the DeepMind website.
September 21 2021 | 14:00