Ataraxos AI defeats leading human Stratego players
Researchers combined self-play training with planning during play; applications beyond games remain possibilities, with work on explaining decisions still ahead.
AI illustrationAn AI system called Ataraxos recorded 39-2 against leading human players at the Stratego world championship, researchers reported in work published in Nature. The system tackles games in which an opponent’s information remains partly concealed.
Adam Zewe described the research in MIT News. The team brought together researchers from MIT, Carnegie Mellon University, New York University and Stanford University, with Samuel Sokota as lead author and Gabriele Farina as senior author.
Ataraxos learns by repeatedly playing against itself, building an initial strategy before it faces an opponent. During a game, it revisits its choices rather than relying solely on that earlier training. A generative model estimates which identities are likely for the opponent’s hidden pieces, helping the system assess possible moves.
This combination matters in Stratego because each player positions 40 pieces whose identities stay secret until pieces meet. The number of possible arrangements makes considering every situation impractical.
Farina said Ataraxos surpassed DeepMind’s DeepNash in playing strength while using less than one hundredth of its training examples and less than one thirtieth of its self-play games. The researchers also adapted their approach to Barrage Stratego, Hanabi and Dou dizhu, reporting performance above leading human players in each.
The researchers see possible uses in decisions involving incomplete information, including business negotiations and cybersecurity. Those remain potential applications of the approach, rather than deployments established by the game results.
A next step is to make the system’s reasoning understandable to people. Farina said decisions would need to be auditable before adoption, with humans retaining authority over whether to follow a recommendation. The team plans to develop ways for Ataraxos to explain its choices.