Ataraxos AI masters Stratego by predicting hidden pieces

Researchers from Carnegie Mellon, MIT, NYU, and Stanford have developed Ataraxos, an artificial intelligence capable of mastering the board game Stratego. The system demonstrated its proficiency by defeating four-time world champion Pim Niemeijer, winning 15 games with four draws.

Unlike games with limited hidden information, Stratego presents a massive search space due to the high number of hidden pieces. Ataraxos overcomes this by utilizing a dual neural network architecture. A specialized belief model estimates the identities of an opponent's hidden pieces based on their movement patterns, allowing the AI to sample plausible game states rather than attempting to brute-force every possible combination.

The AI was trained through self-play, completing approximately 163 million games. To manage the complexity of hidden information, the researchers implemented a training strategy involving significant strategic shifts early in the process and more refined adjustments later. Notably, Ataraxos achieved this capability using significantly fewer resources than previous models like DeepNash, requiring only 16 GPUs for one week of training.