Researchers Defeat World’s Best Stratego Player With Low-Cost AI Model Ataraxos

A team from Carnegie Mellon, MIT, NYU, and Stanford developed Ataraxos, an AI that defeated Pim Niemeijer, the top-ranked Stratego player, in a series of 20 online games. The AI won 15 games, drew 4, and lost 1, marking a significant breakthrough in imperfect-information games where hidden data and long time scales previously stumped systems like DeepMind’s DeepNash.

Technical Approach and Efficiency

Unlike DeepNash, which required 1,024 specialized chips and an estimated $3–4.5 million to train, Ataraxos was trained on 16 GPUs for one week, plus 4 GPUs for four days for its belief model. The team achieved this efficiency by:

• Using a custom simulator that runs millions of moves per second on standard graphics cards.
• Implementing a belief model neural network that predicts opponent piece locations, allowing the AI to sample plausible board states rather than iterating through all 10^33 possible setups.
• Adjusting strategy aggressively early in training and conservatively later to avoid the oscillation common in hidden-information self-play.

Gameplay and Implications

Ataraxos demonstrated a calm, methodical playstyle, often recovering from low-probability positions without the emotional volatility seen in human players. The AI’s strategies have already influenced the human metagame, with players adopting its unconventional setups, such as hiding the flag in corners behind two bombs. The researchers suggest these techniques for handling hidden information and long-horizon planning could extend to real-world scenarios like negotiations and war gaming, though they note the current model lacks interpretability.