Researchers from MIT, Carnegie Mellon University, New York University, and Stanford University have developed an artificial intelligence (AI) system that can defeat elite human players at Stratego, a board game that has long served as a difficult test for AI. The system, named Ataraxos, achieved a record 15-1-4 victory against the world's top-ranked Stratego player and posted a 39-2 record against top competitors at the Stratego world championship. No previous AI had reached this level of performance.
Why the game matters
Stratego is what researchers call a game of imperfect information, meaning each player knows things the opponent does not. In this case, the identity of each piece remains secret until it collides with another piece. This creates what researchers describe as an explosion of possibilities. The number of possible piece arrangements exceeds 10 to the 66th power, far more than chess. Past AI systems, including one developed by Google DeepMind, required millions of dollars in computing costs yet still failed to beat top human players. The new system achieved strictly stronger results using less than one hundredth of the training examples and less than one thirtieth of the self-play games.
The researchers built Ataraxos using self-play reinforcement learning, a method where the AI plays against itself many times to learn effective strategies. They combined this with decision-time planning, a technique that lets the system evaluate possible moves during actual gameplay rather than relying only on pre-learned patterns. During play, Ataraxos uses a generative model to estimate the likely identities of hidden opponent pieces, then chooses moves based on those probabilities. The system remains calm under pressure and avoids overcorrecting when valuable pieces are threatened, a weakness common in human players.
The same approach worked for other hidden-information games. When adapted for Barrage Stratego, Hanabi, and Dou dizhu, Ataraxos again reached superhuman performance. The researchers hope to add interpretability measures, meaning ways for the system to explain its decisions in understandable terms, before it could be used in real-world settings like military planning, business negotiations, or cybersecurity defense.
This matters because decisions in the real world rarely involve complete information. Traders do not know why others are buying or selling. Military commanders do not know enemy positions. Anyone facing an opponent who keeps secrets must reason under uncertainty, and this AI offers a more efficient way to do so than previous methods.
This research is published in Nature.