Ataraxos AI defeats world's best Stratego player, overcoming hidden information challenge

AI Conquers Stratego

A research team from Carnegie Mellon, MIT, New York University, and Stanford University has created an AI named Ataraxos that successfully defeated Pim Niemeijer, considered the best Stratego player globally. Ataraxos won 15 games, drew four, and lost only one, marking a significant milestone in AI game-playing capabilities.

This achievement follows previous AI successes in games like chess (Deep Blue in 1997) and Go (AlphaGo in 2016), but Stratego presented unique challenges that earlier AI systems struggled to overcome.

The Challenge of Imperfect Information

Stratego is an imperfect-information game, similar to poker, where players do not know the identity of their opponent's pieces until they engage in battle. Each player has 40 pieces, ranging from marshals to spies, bombs, and a flag. The objective is to capture the opponent's flag.

The game's complexity stems from the vast amount of hidden information. Unlike Texas Hold'em poker, which has 1,326 possible hands, Stratego involves more than a decillion possible board setups due to the 40 hidden pieces. This makes it difficult for an AI to weigh all potential outcomes.

Long Games and Bluffing

Another factor contributing to Stratego's difficulty for AI is its length. A typical Stratego game can last up to 2,000 moves, significantly longer than the 40 moves common in chess. This extended duration requires the AI to manage hidden information and strategic planning over a prolonged period.

Stratego also incorporates bluffing, where players might move a weak piece to deceive an opponent. Balancing bluffing frequency—not bluffing too often to avoid predictability, but also not never bluffing to maintain threat credibility—was a key hurdle that stumped prior AIs, including DeepMind's DeepNash, introduced in 2022.

Training Efficiency

Ataraxos was trained using 16 GPUs, with the training process costing only a few thousand dollars. This demonstrates that advanced AI capabilities in complex imperfect-information games can be achieved with relatively modest computational resources compared to some larger AI projects.