An artificial intelligence program called AlphaStar now ranks among the top 0.2% of human players for the real-time strategy game StarCraft II. Reported in this week’s Nature, the algorithm represents a major achievement for machine learning that could be adapted to help solve complex problems for other applications.
StarCraft II is a science fiction strategy game that is played professionally in competitions around the world; it is one of the world’s most lucrative professional electronic sports, or esports. In the game, players compete against one another, controlling one of three different alien races that each have different characteristics and abilities. The game has emerged as a grand challenge in the field of artificial intelligence research; previous artificial agents have failed to rival top human players, despite simplifying the game rules, manually programming certain action sequences or relying on superhuman capabilities such as executing tens of thousands of actions per minute.
Oriol Vinyals and colleagues present a multi-agent reinforcement learning algorithm called AlphaStar, in which several deep neural network agents compete against each other, generating a league of continually adapting strategies and counter-strategies. AlphaStar then competed against human players in a series of online games, where it reached Grandmaster level for all three of the StarCraft races. This makes it the first artificial agent to reach the top tier of human performance in a professionally played esport without simplifying the game.