AI Wins Over Human Champion in Game of Stratego

In the realm of gaming AI, milestones have been achieved with machines defeating human champions in various classic games. From Deep Blue defeating Garry Kasparov at chess in 1997 to AlphaGo surpassing Lee Sedol at Go in 2016, the triumphs have been remarkable. However, one game has remained a challenge for AI developers – Stratego.

Despite DeepMind’s substantial resources, they were unable to create a machine capable of consistently outplaying top human players in Stratego. Now, a collaborative team of researchers from Carnegie Mellon University, MIT, New York University, and Stanford University has accomplished this feat. Their AI, named Ataraxos, managed to defeat Pim Niemeyer, one of the greatest Stratego players with an impressive record of 15 wins, 1 loss, and 4 draws. What makes this victory even more astonishing is that it only took 16 GPUs and a modest investment to train the AI.

The Complexity of Stratego

Stratego presents a unique challenge to AI due to its intricate gameplay mechanics. Each player is provided with 40 pieces representing different military ranks, a flag, and a bomb. The objective is to capture the opponent’s flag while concealing the true identity of the pieces. The hidden information aspect makes Stratego akin to a poker-like imperfect information game, which has posed difficulties for AI systems for years.

Unlike games with limited hidden information like Texas Hold’em, where there are only two hidden cards, Stratego offers a vast array of possibilities. With 40 pieces on the board that can be arranged in numerous configurations, Stratego boasts over 100,000 possible setups. Additionally, the length of a Stratego game can extend up to 2,000 moves, prolonging the engagement and strategic depth of the game.

The element of bluffing further complicates the gameplay, as players can deceive opponents by moving weaker pieces as powerful units. Finding the right balance between bluffing and strategic play is crucial in Stratego, as observed in the shortcomings of earlier AI models like DeepMind’s DeepNash.