Mastering Stratego, the classic game of imperfect information

The paper “AlphaZero: A Learning Agent That Conquers Multiple Attrition Games” by researchers from MIT, Google, Harvard, University of Adelaide, and the Australian National University discusses AlphaGo, a deep learning AI system developed by Google’s DeepMind subsidiary. The team introduces AlphaZero, an analogous but fundamentally different system that uses a series of games to teach itself to play the game Go, which was previously thought to be impenetrable for human players. The researchers conducted several benchmark experiments on AlphaGo, where they compared its performance with other AI systems and humans in Go. AlphaZero’s performance was impressive, showing a 70% improvement over the best existing human algorithm, proving that AlphaGo’s deep learning approach is indeed capable of learning complex game-playing skills. The paper also provides insights into how AI systems interact with other systems in real-time strategy games like StarCraft II and how such interactions can be used to test and evaluate AI systems. Overall, the research team’s efforts demonstrate the power and potential of deep learning techniques in game-playing and reinforcement learning applications.