DeepMind’s StarCraft II AI Can Now Defeat 99.8 Percent of Human Players

Google’s DeepMind AI lab has contributed to some of the company’s most impressive AI feats in recent years such as the Wavenet voice engine and object recognition in Google Photos. DeepMind has also shown off its AI prowess by beating humans at games we never thought machines would be able to play. It started by besting the world’s best Go players, and then moved on to StarCraft II. The “AlphaStar” AI beat some of the world’s top players in early 2019, and now it’s playing online and crushing almost all challengers. DeepMind says AlphaStar is now the first AI to reach Grandmaster status in StarCraft II. 

In January, DeepMind streamed matches between elite human players and the AlphaStar AI. During those matches, AlphaStar showed an incredible understanding of the game, rotating damaged units out of harm’s way, baiting enemies on ramps, and utilizing special abilities to cut enemy formations in half. AlphaStar beat almost all of its human opponents that day, and DeepMind has continued improving the AI in the months since. 

The version of AlphaStar (or “agent”) playing online has undergone some changes that make it fairer to the human players. For example, the AI now only sees the section of the map in the main view, whereas before AlphaStar saw the entire map. It can also play and battle against all three races in StarCraft II (it was limited to Protoss before). Finally, DeepMind capped AlphaStar to a human-like 22 mouse clicks every five seconds. 

Even with those limitations, AlphaStar has managed to rise through the ranks and is now at the Grandmaster level. That’s the highest tier in competitive StarCraft II gameplay. It can defeat 99.8 percent of human players, and it may only be a matter of time before it can crush all human challengers. 

DeepMind decided to test its AI on StarCraft II because it’s a complex but endlessly strategic game. There are multiple paths to victory, and most players will tell you there’s a strong intuitive aspect to high-level StarCraft gameplay. DeepMind uses reinforcement learning to improve its AI, allowing AlphaStar to log years of game time every day by playing against different versions of itself. DeepMind believes that this same AI technology could one day have applications in robotics to improve motor control. Now, it’s just really good at pwning human StarCraft II players.

Now read: