Grandmaster level in StarCraft II using multi-agent reinforcement learning

ABSTRACT

The game of StarCraft II has emerged as a grand challenge for artificial intelligence research owing to its complex, multi-agent, and partially observable environment. Here we introduce AlphaStar, an artificial intelligence that plays StarCraft II at a Grandmaster level using a multi-agent reinforcement learning algorithm. AlphaStar uses a novel neural network architecture that incorporates a transformer torso and a deep LSTM core, trained via supervised learning from human replays followed by a multi-agent league of self-play. AlphaStar reached Grandmaster level for all three StarCraft races (Protoss, Terran, and Zerg), demonstrating that general-purpose learning algorithms can scale to highly complex, imperfect-information environments.