AlphaStar

AlphaStar was developed by DeepMind to play StarCraft II, the real-time strategy game created by Blizzard Entertainment. StarCraft II is considered one of the most challenging esports for AI due to its imperfect information, enormous action space, and requirement for real-time strategic decision-making.

AlphaStar was revealed through exhibition matches in late 2018 against professional players Dario "TLO" Wünsch and Grzegorz "MaNa" Komincz. After extensive training, the agent reached Grandmaster rank on the European StarCraft II ladder, placing it among the top fraction of a percent of all active players.

AlphaStar used deep reinforcement learning combined with a multi-agent league training system, in which different agent populations specialized in different strategies and played against each other. The system was initially trained on human game replays and then refined through massive amounts of self-play.

Unlike chess and Go, which DeepMind previously mastered with AlphaZero, StarCraft II features imperfect information through the fog of war, real-time rather than turn-based play, and a vast action space involving hundreds of units. The game also requires simultaneous management of economy, military strategy, and tactical micro-decisions, making it far more complex for AI systems.