Here’s how an AI managed to sweep a professional at StarCraft II

The AlphaStar algorithm beat a professional StarCraft II player by a total of 5 wins to 0 last month, DeepMind announced on Thursday. A feat that greatly impresses the experts of the game, but does not surprise the researchers.

AlphaStar is the latest addition to DeepMind. This Google artificial intelligence (AI) research laboratory also designed AlphaGo, the algorithm that beat the 2015 world go champion.

The new algorithm surprised Poland’s Grzegorz Komincz, aka MaNa, one of the world’s top StarCraft II players , by adopting new strategies. “It was different from all the StarCraft games I’ve played,” said Komincz, who has already won two major championships at this game and has stepped up the podium multiple times.

A remarkable feat

Unlike table games like checkers, chess and go, real-time strategy video games like StarCraft II have many obstacles that make working an algorithm difficult.

In chess, for example, the computer can see the entire game board and establish its strategies accordingly. At StarCraft II , players only see what’s close to their units. An algorithm must be able to predict what its opponent is doing from limited information, which is similar to human intuition.

The game is also played in real time, giving players a much shorter time to think about a strategy and adapt to their opponent’s actions. The quantity of different units, each with particular capacities, is also much greater than that of failures.

Finally, the number of valid positions in the game of go is 1 followed by 170 zeros, while this number is estimated at 1 followed by 270 zeros for StarCraft II , according to Wired .

In order for AlphaStar to beat a professional player, DeepMind’s engineers first decided to provide him with the data of 500,000 games played between humans. They then cloned the algorithm and played these clones together in a virtual tournament. The winners of each match then clashed, then the winners of this second wave, and so on until the algorithms reach the equivalent of 200 years of gaming experience.

AlphaStar, a limited master

Despite its formidable efficiency in its games against MaNa, AlphaStar still remains a very limited artificial intelligence system. Since he was trained on a single card, with one species ( StarCraft II has three), he is unable to play in other levels of the game or with another species. His mastery of StarCraft II is also not transferable to other real-time strategy video games, while human champions are normally able to easily adapt to other games of the same genre.

In addition, his five victories were obtained under special conditions: AlphaStar could see the entire map visible to his units at a time, while a human can normally display only a small part on his screen. The organizers of the challenge gave Grzegorz Komincz the chance to face a modified version of AlphaStar whose vision of the map was the same as a human. Despite some difficulties at the beginning of the game, MaNa finally triumphed by exploiting some errors committed by the algorithm.

Surpassing humans is a bit ‘boring’

Although Grzegorz Komincz was impressed by the game of AlphaStar, Mark Riedl, an associate professor at the Georgia Institute of Technology, told Wired that he was not too surprised by the results of the matches.

“We reached a point where it was just a matter of time,” he said. In a way, beating humans to games has become a little boring.”

Other artificial intelligences capable of competing or even beating video game professionals have appeared in recent months, especially for the Quake III Arena and Dota 2 games .

Leave a Reply

Your email address will not be published.