With most information hidden, the game Stratego
Summary
Ars Technica reports that a team from CMU, MIT, NYU, and Stanford developed Ataraxos, an AI that defeated Pim Niemeijer in 20 online Stratego games with 15 wins, four draws, and one loss across three weeks, using just 16 GPUs and a few thousand dollars. The approach combines a belief model to guess hidden pieces with a self-play training regime, enabling forward-planning in a large search space; it outperformed the previous DeepNash approach and even beat world champions at related variants. The article discusses implications for AI in imperfect-information domains and potential real-world applications, while noting the system's current lack of explainability.