Researchers from Carnegie Mellon, NYU, Stanford, and MIT developed Ataraxos, an AI capable of defeating the world's top Stratego players. In a 20-game series, the system secured 15 wins against Pim Niemeijer, a player with four world championship titles.
The training process for Ataraxos cost less than $8,000, utilizing 16 Nvidia H100 GPUs for one week. This is significantly more efficient than the DeepMind DeepNash system, which researchers estimate would cost between $3 million and $4.5 million to train at 2025 prices.
Ataraxos operates without human data, relying on self-play and a custom belief network to predict hidden pieces. The system employs a regularization technique to maintain strategic variety, preventing it from becoming predictable during gameplay.
Beyond Stratego, the team applied their methodology to other games, including the Barrage variant, the card game Hanabi, and the Chinese card game Dou dizhu. The researchers have released the Ataraxos code to the public.
