AI conquered a complex real-time strategy game once thought beyond reach. In 2019 a system reached elite human level at StarCraft II, handling hidden information and long horizons. It extended AI’s game-playing prowess.
A harder game
Complexity soared. The game hides information and spans long horizons. Search alone fails.
Elite play
The bar was high. The system reached top human ranks. It beat strong players.
Learning from games
Data mattered. It learned from human replays then self-play. Skill compounded.
Multi-agent
Dynamics were rich. Strategy emerged against varied opponents. Adaptation was key.
Constraints imposed
Fairness was tuned. Reaction speed was limited to be humanlike. Comparison stayed meaningful.
Broader lessons
Transfer is hoped. Techniques may aid real planning. Research continues.
The bottom line
An AI reached elite human level at StarCraft II in 2019, mastering hidden information and long horizons. It learned from replays and self-play. The techniques aim toward real-world planning.
