Scalable decision-making for games of imperfect information.

Ataraxos, an AI for Stratego, achieves the first superhuman victory over the best human player, using significantly less compute and data.

• Why it matters: Hidden information complicates decision-making and has limited AI progress in complex, real-world strategic games, creating a critical challenge for artificial intelligence.
• What they did: The team developed Ataraxos using novel self-play reinforcement learning and test-time search techniques tailored for large-scale hidden information, applying these methods to multiple games.
• The result: This approach not only outperformed top human players in Stratego but also delivered superhuman and state-of-the-art results in other games, demonstrating a scalable design pattern for decision-making under hidden information.