AI Cracks Stratego: Human Champion Defeated

For years, complex games like chess and Go had fallen to artificial intelligence, but Stratego remained a formidable challenge. Now, a collaboration between Carnegie Mellon, MIT, NYU, and Stanford has unveiled Ataraxos, an AI capable of reliably defeating top human players. Ataraxos decisively beat Pim Niemeijer, a legendary Stratego grandmaster, with a score of 15 games to one. This AI's development marks a significant advancement in artificial intelligence's ability to master imperfect-information games.

Stratego's complexity stems from its hidden piece identities and bluffing mechanics, which previously stumped AI systems like DeepMind's DeepNash. Ataraxos overcame these challenges by employing a novel training method that involved making large strategic adjustments early in its learning process and a unique belief model to predict opponent's hidden pieces. This allowed it to sample plausible game states rather than exploring every possibility, a key to its success.

The AI's name, Ataraxos, reflects its calm and unbothered playstyle, allowing it to make calculated moves even from disadvantaged positions. This methodical approach enables it to execute bluffs more effectively than humans. The development team highlighted the AI's cost-effectiveness, requiring only 16 GPUs and a few thousand dollars for training, a stark contrast to previous AI endeavors that incurred millions.

Beyond Stratego, the Ataraxos architecture has shown promise in other complex games, including Barrage Stratego and Hanabi. Researchers believe this technology could be applied to real-world challenges such as negotiations and financial markets, though making the AI's decision-making process interpretable remains a future goal.