How artificial intelligence beat 99.8% of human players in StarCraft

How artificial intelligence beat 99.8% of human players in StarCraft 

DeepMind announced that their AlphaStar algorithm, designed to play StarCraft II, beat 99.8% of all players participating in the tournament.

This result gives the program the right to be called a “grandmaster,” the developers said.

Thus, artificial intelligence ended up in the Elite League, which includes only 200 of the best players around the world. This means that it beat 99.8% of active Battle.net users.

The AlphaStar program trained by looking at the recordings of the matches of the best e-sportsmen, playing against itself for each race and comparing the effectiveness of different approaches. The whole learning process took 44 days.

In order to balance the possibilities of the program with the people, AlphaStar has been modified accordingly. In particular, artificial intelligence was trained to play with all three races available in StarCraft II. In addition, the map review was limited to the same level that is accessible to humans. Finally, there was a limit on the number of clicks of the mouse buttons – no more than 22 non-duplicate actions every 5 seconds of the game. This corresponds to the number of standard movements when playing a person.

Реклама

“The history of progress in the field of artificial intelligence has been marked by significant achievements in games. Since computers hacked go, chess and poker, StarCraft has become the next big challenge. The complexity of the game is much greater than in chess, because players control hundreds of units; more complicated than go, because there are 10 in 26 possible options for each move; and players have less information about their opponents than poker, ”said David Silver, lead researcher at DeepMind at AlphaStar.

However, the software is still limited by the narrow discipline for which it is intended. This is because software is not programmed to replace rule sets. Instead, DeepMind and other research institutes use reinforced learning to enable agents to figure out how to play on their own, so software often develops new and wildly unpredictable styles of play.

In early 2019, HB reported that the AlphaStar program, developed by the Google DeepMind division, was able to beat two professional players in the StarCraft II strategy.

In particular, in a series of test matches AlphaStar was defeated by Grzegorz Kominz from Team Liquid, one of the most powerful professional StarCraft players in the world.