This game-playing AI is the new champ at Stratego

A new artificial intelligence system, developed by researchers from MIT, Carnegie Mellon University, New York University, and Stanford University, has achieved superhuman performance in the challenging board game Stratego. This victory marks a significant step forward for AI systems capable of handling tasks involving hidden information, which has potential applications in various real-world scenarios such as business negotiations and cybersecurity.

Stratego is a two-player game with imperfect information, wherein players hide the identities of their pieces until they come into contact. The game's complexity stems from the vast number of possible piece configurations, numbering over 10 to the 66th power. Previous attempts to create AI systems with superhuman Stratego skills, such as DeepMind's DeepNash, relied on computationally expensive techniques that were still unable to surpass top human players.

To overcome these challenges, the researchers created Ataraxos, an AI system designed to be both efficient and effective, named after the Greek term for someone unbothered or free from anxiety. Ataraxos uses a combination of self-play reinforcement learning and decision-time planning to excel at Stratego. By employing self-play reinforcement learning, Ataraxos learns a strong "blueprint strategy" by playing against itself repeatedly, and it then refines its choices using decision-time planning, which estimates the likely identities of hidden pieces and evaluates future moves before selecting the optimal strategy.

The AI's innovative use of decision-time planning allows it to navigate the complex web of possibilities in Stratego, demonstrating the potential of AI systems to handle tasks with hidden information more effectively than ever before. In a recent contest, Ataraxos defeated the world's strongest Stratego player 15-1-4 and achieved a 39-2 record against top human players in the Stratego world championship.

This groundbreaking achievement sets the stage for further advancements in AI technology that can tackle complex problems involving hidden information.