A team of researchers at DeepMind develops an AI agent called DeepNash that can play the Stratego game at an expert level.
The Stratego game is a two-player board game and is difficult to master. The goal for each player in Stratego is to capture their opponent’s flag hidden among their initial 40 game pieces. Every game piece is marked with a power ranking. Higher-rank players defeat the lower-ranked players during the face-offs in Stratego. The players in Stratego cannot see the markings of the opponent’s game pieces until they are in face-offs.
DeepNash first learned to play the Stratego game against itself many times. Researchers at DeepMind came up with an algorithm based on game theory, which uses an optimal strategy for every move in the game. They have published and explained the entire work of DeepNash in the paper, ‘Mastering the game of Stratego with model-free multiagent reinforcement learning.’
Testing revealed that DeepNash achieved an 84% winning rate against the top expert human players on the Gravon games platform and became one of the top three players. Gravon is a virtual world that allows users to play board and card games together. Researchers did not inform the players on Gravon that they were playing against a computer.