On October 30, 2019, DeepMind announced that its AlphaStar agent had reached Grandmaster level in StarCraft II, the top rank on the game’s competitive ladder. According to DeepMind, “AlphaStar was ranked above 99.8% of active players on Battle.net, and achieved a Grandmaster level for all three StarCraft II races” (Protoss, Terran, and Zerg). The work was published in the journal Nature.
StarCraft II is a real-time strategy game that poses challenges distinct from board games: players control many units at once, much of the map is hidden behind a fog of war, decisions unfold continuously rather than in turns, and there is no single best strategy. AlphaStar learned through multi-agent reinforcement learning, training a whole league of agents that competed against one another so that the system kept improving and could not be beaten by a single fixed strategy.
Crucially, AlphaStar played under human-comparable constraints. It used a camera-style view of the map and was capped at a limited number of actions over time (22 actions per 5 seconds), limits reviewed with professional player Dario “TLO” Wunsch, so that its success came from better decisions rather than superhuman clicking speed. Alongside OpenAI Five’s Dota 2 victory the same year, AlphaStar marked the point where reinforcement learning agents reached top-tier human performance in complex, real-time strategy games.