The History of AI

AlphaStar

Era 4 · Deep learning and foundation models (2012–2021) › Deep reinforcement learning

Lineage
Data and learning
Kind
Application

Who

DeepMind

What happened

DeepMind reports in Nature that AlphaStar has reached Grandmaster level in StarCraft II, above 99.8 percent of players, competing anonymously on the game's public servers with a limited number of actions.

Why it mattered

Reinforcement learning reaches the elite of a real-time game, with imperfect information and thousands of possible actions at every moment.

The full story

StarCraft II is a real-time strategy game: each player builds a base and an army, cannot see what the opponent is doing except where their own units are, and must make decisions continuously over long games. Compared with Go, there is hidden information, far more possible actions, and no turn that waits.

AlphaStar started by imitating the moves of the best players in its database of human games. Then several agents played against one another and improved through deep reinforcement learning.

In December 2018 it beat the professional player MaNa 5–0. That version saw the whole map at once, instead of looking at it through the game camera as a person does, and a version retrained under more realistic limits lost the rematch against MaNa. The next version was restricted: with a cap on actions in every five-second window, it competed anonymously against players on the game's servers from July 2019, and by the end of August it had reached Grandmaster level, among the top 0.2 percent of human players.

The Nature paper, published online on October 30, 2019, with 42 authors led by Oriol Vinyals, framed it as evidence that multi-agent reinforcement learning could compete with the best humans in an environment that complex.

Videos

Sources

How to cite

“AlphaStar.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/alphastar

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline