AI Masters Imperfect Information in Stratego and Diplomacy
Artificial intelligence achieves new milestones by conquering Stratego and Diplomacy, two complex board games that rely heavily on hidden information and human-like negotiation rather than pure calculation.
Artificial intelligence achieves a major milestone by mastering Stratego and Diplomacy, two complex board games that rely on imperfect information. Unlike chess or Go where all pieces are visible, these games require players to hide their identities, form secret alliances, and deceive opponents. This shift forces AI to move beyond raw calculation and instead learn softer skills like bluffing and predicting human behavior.
DeepMind tackles Stratego with a new model called DeepNash, which plays in an unpredictable way that opponents cannot easily exploit. By using a novel algorithmic method called Regularised Nash Dynamics, the system wins 84% of its matches against experienced human players. DeepNash even employs complex tactics like sacrificing high-level pieces as a calculated risk to trick opponents into revealing their strongest assets.
Meta and CSAML take on Diplomacy with an AI named Cicero, which plays the game at a human level by integrating natural language processing with strategic planning. Cicero navigates the game's core mechanics of forming agreements, building alliances, and executing vendettas against other players. Together, these breakthroughs show that AI is evolving to handle the nuanced, psychological elements of human interaction.