An artificial intelligence system, named Ataraxos, has achieved a championship level in the game of Stratego, defeating highly skilled human players. Developed by a collaborative team from MIT, Carnegie Mellon University, New York University, and Stanford University, Ataraxos represents a significant advance in AI's ability to handle games with hidden information. Stratego, a game where opponents' piece identities are concealed, has long served as a benchmark for testing AI's strategic reasoning capabilities.
The Ataraxos system integrates two key elements: a learned strategy blueprint derived from extensive self-play, and a decision-time planning technique. This latter method allows the AI to refine its choices in real-time by estimating the probabilities of an opponent's hidden pieces. Researchers published their findings in the journal Nature on September 30, 2026.
This approach proved more efficient than prior AI models. Ataraxos required substantially less training data and fewer self-play games compared to DeepNash, a previous Stratego-playing AI developed by DeepMind. The new system used less than one-hundredth of DeepNash's training examples and less than one-thirtieth of its self-play games. This efficiency suggests a more scalable method for developing AI in complex environments.
Beyond Stratego, the researchers adapted Ataraxos to perform well in other games with imperfect information, including Barrage Stratego, Hanabi, and Dou dizhu. This cross-game success indicates the generality of the AI's underlying principles. The implications of this research extend beyond gaming. The ability of AI to strategize effectively with incomplete information could aid human decision-making in real-world scenarios such as business negotiations, cybersecurity planning, and military operations.
Gabriele Farina, an assistant professor at MIT and a senior author on the paper, highlighted the challenge of imperfect information tasks. He noted that in such situations, enumerating all possibilities is often infeasible due to their sheer volume. Farina stated that AI algorithms capable of general-purpose performance on these difficult tasks represent a substantial step forward.
The development of Ataraxos addresses a long-standing challenge in artificial intelligence. Games like Stratego, with their hidden elements and strategic depth, have proven more difficult for AI than games with perfect information, such as chess or Go. While systems like AlphaZero have excelled in perfect information games, their methods are not easily transferable to scenarios where crucial information is concealed. The complexity of Stratego, with its vast number of possible piece configurations, has historically made it a formidable test case.
Future work aims to incorporate interpretability features into Ataraxos, enabling the system to explain its decision-making processes in a way that humans can understand. This step is considered important for applying the AI's capabilities to real-world problems where human oversight and comprehension are essential.