RESEARCH · RESEARCH · #1352
Ataraxos (CMU/NYU/Stanford/MIT) defeats top Stratego player Pim Niemeijer
Researchers from Carnegie Mellon, NYU, Stanford and MIT published a Nature paper introducing Ataraxos, an AI that the authors say is, to their knowledge, the first to reach superhuman performance in Stratego; in an official 20-game match it beat Pim Niemeijer 15–1–4. The system trains purely by self-play, uses a belief network and a regularization schedule to maintain strategic diversity, and—according to the paper—was trained in about a week on 16 Nvidia H100 GPUs (plus four days on four GPUs for the belief network) with an estimated compute cost under $8,000 (2025 prices).
KEY POINTS
- Researchers from Carnegie Mellon, NYU, Stanford and MIT published a Nature paper introducing Ataraxos, an AI that the authors say is, to their knowledge, the first to reach superhuman performance in Stratego; in an official 20-game match it beat Pim Niemeijer 15–1–4.
- The system trains purely by self-play, uses a belief network and a regularization schedule to maintain strategic diversity, and—according to the paper—was trained in about a week on 16 Nvidia H100 GPUs (plus four days on four GPUs for the belief network) with an estimated compute cost under $8,000 (2025 prices).
- This matters because it demonstrates a practical, sample-efficient approach to a hard imperfect-information game—achieving superhuman play in Stratego at a tiny fraction of the compute cost reported for DeepNash/DeepMind.
WHY IT MATTERS
This matters because it demonstrates a practical, sample-efficient approach to a hard imperfect-information game—achieving superhuman play in Stratego at a tiny fraction of the compute cost reported for DeepNash/DeepMind.