Skip to content
ML Atlas

AlphaStar

DeepMind's StarCraft II agent, which reached Grandmaster level for all three races on the public ladder in 2019, under human-like viewing and speed limits.

Also StarCraft II · league training

Role · kind
Real systemsFamous system
Level
Intermediate

Vinyals et al., DeepMind, 2019

A computer player for a fast real-time strategy game where you can't see the whole map and must control hundreds of units at once. It learned from human games, then from a big league of practice opponents, and beat almost all human players.

Where it fits

Reinforcement LearningReinforcement LearningActor-Critic MethodsActor-Critic MethodsLong Short-Term MemoryLong Short-Term Memo…Self-PlaySelf-PlaySupervised LearningSupervised LearningTransformerTransformerAlphaStar
+All relationships as a list

Is a kind of / part of

See it on the full map

How it works

Board games show the whole board and let you think. StarCraft II does neither: you see only part of the map, both players act at the same time, games last thousands of steps, and each moment offers a huge number of possible actions. In 2019 DeepMind’s AlphaStar reached Grandmaster level, ranking above 99.8% of active players, for all three of the game’s races.

How it learned:

  1. Imitation: first, supervised learning on human game replays gave it a reasonable starting strategy.
  2. League training: then many agents trained against each other with reinforcement learning (actor-critic methods). Besides main agents, the league included exploiters whose only goal was to find weaknesses in the main agents, which kept them from overfitting to one style. This is an advanced form of self-play.

The networks combined a Transformer-style module to process the list of units with an LSTM memory over time.

To keep the comparison fair, the final version viewed the game through a camera like humans do and had limits on how fast it could act.

Everyday use cases

  • Case 01

    How you've probably met it

    Through its 2019 matches against professional StarCraft players and its anonymous games on the public ladder.

Strengths, weaknesses and when to use it

Strengths

  • Handled imperfect information, long time horizons and huge action spaces in real time.

Weaknesses

  • Enormous compute; skills were specific to one game.

When to use itAs the landmark example of multi-agent RL in a complex real-time game.

Compare with

Go deeper

Sources

  1. [1]Vinyals et al., Grandmaster level in StarCraft II using multi-agent reinforcement learning, Nature 575 (2019)
  2. [2]Google DeepMind, AlphaStar: Grandmaster level in StarCraft II using multi-agent reinforcement learning (2019)

Facts checked as of 2026-10-12