“StarCraft II”, a trilogy from the US game developer Blizzard Entertainment, is a real-time strategy game. Two opponents face each other on a limited map and try to destroy the enemy base by taking control of units and buildings.

In the research of artificial intelligence, the strategy game presented a particular challenge, not least because of its complexity, according to a technical article published in the journal “Nature”. Until now, according to the authors, computer programs have never achieved anywhere near this capability.

In addition, artificial intelligence had to compete against other players and process a great deal of incomplete information. After widely acclaimed successes in the Go board game, a DeepMind AI system recently performed well in the multiplayer game Quake III Arena in May.

“Nonhuman advantages” eradicated

Back in January, DeepMind had already released a version of AlphaStar that was criticized as having “nonhuman advantages”. In the current version, care was therefore taken to ensure that the system was not able to view the entire map in any different way than human players can and that it could not perform more actions per minute.

Since July, versions of the program have been competing anonymously against human players in public competitions. However, this does not prove that AlphaStar can consistently beat the best professional players. But the experiment showed that the AI itself had learned to play well, said Marcus Liwicki, Professor at the Lulea University of Technology in Sweden.

However, Jan Peters, a professor at the Technical University of Darmstadt, said that the AI had not adapted much to the opponent in the pre-planned strategy. A top player noticed in the final game that he was playing against a machine and then tricked it. “The AI lost in a stupid way and made mistakes that even occasional players would not make.”