Updated
Updated · MIT News · Sep 30
Ataraxos Beats Top Stratego Players 15-1-4, Using Under 1/100 of DeepNash Training Data
Updated
Updated · MIT News · Sep 30

Ataraxos Beats Top Stratego Players 15-1-4, Using Under 1/100 of DeepNash Training Data

2 articles · Updated · MIT News · Sep 30

Summary

  • MIT-led Ataraxos posted a record 15-1-4 against the world’s strongest Stratego player and went 39-2 versus top humans at the Stratego world championship, marking the first clear superhuman result in the hidden-information game.
  • Less than 1/100 of DeepNash’s training examples and less than 1/30 of its self-play games were enough because the system pairs self-play reinforcement learning with decision-time planning that estimates opponents’ hidden pieces before each move.
  • 10^66-plus possible piece configurations make Stratego far harder to brute-force than chess, and earlier systems including DeepMind’s remained too costly and too weak to beat elite human players.
  • The researchers also adapted Ataraxos to Barrage Stratego, Hanabi and Dou dizhu, where it again reached superhuman performance, suggesting broader use in negotiations, cybersecurity and military planning.
  • Nature published the work as the team said the next step is interpretability, so humans can audit and overrule the model before any real-world deployment.

Insights

How did Ataraxos crush world-class Stratego players while using far less training than DeepMind’s DeepNash?
If AI can infer hidden pieces and outbluff champions in Stratego, how close is it to real-world planning under uncertainty?
Why do researchers say interpretability—not raw strength—is the next obstacle for superhuman AI in hidden-information games?