How much power do machines have?

True, no one would be offended that a car can cover a distance faster than a human. But when it comes to creativity rather than physical performance, amazement is quickly swallowed up by pessismism. This sometimes obscures the view of the quite remarkable achievements of researchers who program such smart machines. Google bought Deepmind in 2014; the search engine company is said to have paid more than $500 million for the AI start-up from London.

Now the Deepmind researchers have gone one better: For they have trained their artificial intelligence in the video game “Starcraft” and let it compete against professional video gamers. Ten games were recorded in December, and an eleventh was then played on Thursday evening and broadcast along with the others in a livestream. “Alpha Star,” as the program was called, managed to win ten out of eleven games, causing amazement and excitement among experienced commentators and the gaming community.

StarCraft II is immensely popular

Starcraft is one of the most popular computer games, in which there are also professional competitions. The opponents build a base, strengthen it and create units with which they then attack the opponents. There are many strategies for doing this, both defensive and offensive tactics. It has been played online for more than 20 years, yet it is still immensely popular: on Thursday alone, more than 253,000 games were played around the world by registered teams.

It is among the most complicated strategy games because players must consider hundreds of options for the next best move at the same time and set their strategy for a long period of time. Unlike board games such as chess or Go, where professionals also think and plan far ahead, however, there are no set moves in the computer game. So players have to react to supposedly unexpected actions of opponents, which is of course just as true for a computer.

A highly complex game

The game complexity, which describes how many positions are reachable for a player, is very high in Starcraft. Tic Tac Toe, or “Three Wins” has a game complexity of 10 to the power of 3 with its nine squares, already mastered by a computer in 1952. The 64 squares of chess are much more complicated with 10 to the power of 47, but in 1997 IBM supercomputer Deep Blue beat world champion Garry Kasparov anyway. The game complexity of Go is 10 to the power of 170, that of Starcraft is said to be 10 to the power of 1685 (these numbers are estimates and can change due to the smallest rule changes).

The superiority of the AI therefore surprised many. While it made mistakes that were easy to spot, it still dominated the game to such an extent that it was superior to professional players. This is especially remarkable because the Deepmind engineers did not intentionally make the computer unfairly superior. After all, in a computer game, it would be technically possible to ramp up the reaction time, or the actions that can be performed in a minute, to the point where it would be impossible for humans to keep up. But Alpha Star reacted more slowly than its human opponents and required less so-called APM, which stands for “actions per minute.” The audience was particularly surprised by how “human” the AI played.

#AlphaStar sees everything

The computer’s advantage consists mainly of two factors: Unlike the opponent, Alpha Star can see over the entire playing field, which in Starcraft is quite large. Also, the AI could control several units at once, which is difficult for humans. Only in the last game, which was broadcast live, did Alpha Star have to find its way around the virtual map piece by piece and adjust the camera. It was the only game the computer lost. It could have been the last.