AI First: Google's AlphaStar Resoundingly Defeats StarCraft II Pro Players

Google DeepMind‘s AlphaStar AI has stunned ace players of the real-time strategy game StarCraft II. In what is hailed as a monumental achievement for the team behind Google’s artificial intelligence, the AlphaStar AI defeated the best StarCraft pro players in the world 10-0.

“This is, of course, an exciting moment for us,” David Silver, DeepMind’s lead researcher on AlphaGo and co-lead on AlphaStar, said in an interview. “For the first time, we saw an AI that was able to defeat a professional player.”

In the livestream of the competition, DeepMind demonstrated the progress in their gaming AI’s capabilities. The video showed the artificial intelligence system competing with two professional gamers. Each player got to play five games against AlphaStar but failed to win even just a single match.

How DeepMind Trained its AlphaStar AI

According to DeepMind, they trained their AlphaStar AI using 200 years worth of gameplay. The team reportedly created a league of AI agents relying on a neural network using the replays of matches played by humans.

For seven days, the AI agents tested the strategies from the neural network by playing against each other. From these matches, the AIs learned from the strategies that always lead to victories.

Five of the AI agents played against TLO, Germany’s top StarCraft II player, on the Catalyst map. Out of the five matches, the AI came out victorious.

After its satisfying match against TLO, the AI agents were once again trained for another week before facing Poland’s ace gamer, MaNa. Unlike TLO who’s better with Zerg race, MaNa is exceptionally talented with using the Protoss race.

However, MaNa’s efforts and gaming experience were still not enough to defeat the AlphaStar AI. Like TLO, the gamer lost against the AI, 5-0.

AI experts who watched the livestream were deeply impressed with the AI’s performance.

“I didn’t expect anything like this, especially because previous attempts of learning StarCraft end-to-end have been very far from human level,” Niels Justesen, a professor from the IT University of Copenhagen in Denmark, was quoted as saying.

The DeepMind team now hopes that the progress they demonstrated in training their AlphaStar AI could lead to further developments that will allow artificial intelligence to solve complex scientific problems in the future.