An AI system has beaten the strongest human players of Stratego, the board game where you cannot see what your opponent's pieces are. Ataraxos won 15 games, lost one and drew four against the world's top-ranked player, and went 39-2 against top players at the world championship. Researchers from MIT, Carnegie Mellon University, New York University and Stanford described it in Nature on September 30.
Chess and Go show both players the whole board. Stratego hides the rank of every enemy piece until it fights, so a player has to guess, bluff and plan around what might be there. The game has more than 10^66 possible piece setups, far more than chess.
Less practice, stronger play
Ataraxos also plays better than DeepNash, Google DeepMind's earlier Stratego AI, while using far less training, according to the team.
"Our system reaches strictly higher playing strength than DeepNash while using less than one hundredth of the training examples and less than one thirtieth of the self-play games," says Gabriele Farina of MIT.
Planning on the fly
Ataraxos first learns a basic strategy by playing against itself, a method called reinforcement learning. During a real game it adds a second step. A generative model guesses where the hidden pieces probably are, and the system plans its next move around those guesses.
"With Stratego, there is an explosion of possible universes you might have to deal with," Farina says. Methods built for poker, another game of hidden information, could not handle that scale.
The researchers say the approach could help in other settings where information is hidden, such as negotiations or cybersecurity. Those uses have not been tested.
