Researchers built Ataraxos, an AI for Stratego. It beat the most decorated human player by a large margin. The win marks the first superhuman result in Stratego. Ataraxos used orders of magnitude less compute and data than past efforts. The same techniques also produced superhuman AIs for Barrage Stratego and dou dizhu. They set a new state of the art for Hanabi. The approach works across adversarial, cooperative and team games. It relies on self-play reinforcement learning and test-time search. A policy–value network proposes decisions and predicts outcomes. A belief network models hidden information. Search refines decisions at test time. The work establishes a design pattern for hidden-information decision-making.
What changed
Previous AI efforts for Stratego required multimillion-dollar compute budgets and still fell short of top human play.
What it unlocks
Superhuman performance in Stratego and other hidden-information games with far less compute and data.
- 15 wins, 1 loss, 4 draws in 20-game series
- over 10^33 possible piece configurations in Stratego
- few thousand dollars to train Ataraxos
Sources