With most information hidden, the game Stratego had stumped AI until now

AI can now reliably strategize and win games with hidden information, directly impacting AI's ability to handle real-world problems with uncertainty.

Researchers from Carnegie Mellon, MIT, NYU, and Stanford have developed an AI named Ataraxos that has defeated a top human player in the complex game of Stratego. Stratego, unlike chess or Go, presents an "imperfect information" challenge, where players' pieces are hidden from their opponent and only revealed through combat. This complexity, involving massive hidden information and long game scales, had previously stumped advanced AIs, including DeepMind's. Ataraxos achieved its success by employing a second neural network to predict opponent piece identities and by making significant strategic adjustments during its training. The AI played 163 million self-play games to learn, a process that cost only a few thousand dollars and 16 GPUs, a fraction of the resources required for previous AI game-playing breakthroughs. Ataraxos decisively beat Pim Niemeijer, a world champion Stratego player, 15-1 with four draws over 20 online games, and later won 38 out of 40 games at a world championship event. The AI's architecture and learning approach have also shown promise in mastering other complex games like Hanabi and Dou Dizhu.

The development of Ataraxos signifies a breakthrough in AI's capability to handle imperfect information games, a characteristic prevalent in many real-world scenarios such as negotiations and financial markets. By incorporating a belief model to infer hidden piece identities, the AI navigates uncertainty far more effectively than previous attempts. This advancement suggests that AI models can move beyond games with complete information, like chess or Go, to tackle complex strategic challenges where deception and incomplete data are key factors. The significantly lower training cost and resource requirement for Ataraxos compared to earlier AI projects also democratizes the development of advanced AI capabilities, making powerful AI tools more accessible.

This breakthrough in Stratego has implications beyond gaming, potentially accelerating AI applications in fields demanding strategic foresight under uncertainty. The research team's success in developing an AI that can bluff, adapt, and make decisions while "ignoring" known secrets could translate to more sophisticated AI agents in business, defense, and diplomacy. As AI continues to advance in its ability to manage ambiguity and partial knowledge, we can expect more AI systems to operate effectively in environments previously dominated by human intuition and complex human interaction. The next steps will likely involve making these AI strategies more interpretable and explainable to foster trust and broader adoption.