Researchers from Carnegie Mellon, MIT, New York University, and Stanford University have developed an artificial intelligence system that defeats the best human Stratego players. Their model, named Ataraxos, defeated Pim Niemeijer fifteen games to one in a recent match. This achievement stands out because previous attempts by major tech firms failed to create a reliable winner despite having significantly larger budgets.
The core difficulty lies in the game’s hidden information structure. Players do not know the identity of any piece until it collides with an opponent, meaning strategies must account for uncertainty over extended periods. The new approach required only sixteen GPUs and a few thousand dollars to train, contrasting sharply with the massive resources typically needed for similar tasks. This demonstrates that high-performance agents can emerge from modest computational inputs when dealing with complex, imperfect information.
- Ataraxos defeated Pim Niemeijer fifteen games to one with four draws.
- The system required just sixteen GPUs and a few thousand dollars to train.
- The study was led by researchers from Carnegie Mellon, MIT, New York University, and Stanford.
