Stratego has long been a bastion of human skill in the age of game‑playing AI. While computers conquered chess, Go and even poker, the classic board game—where each player commands 40 hidden pieces whose identities are revealed only when they clash—remained out of reach. The enormous hidden‑information space (more than a decillion possible piece arrangements) and games that can stretch to two thousand moves made traditional reinforcement‑learning approaches falter; even DeepMind’s DeepNash, launched in 2022 with massive computational resources, could only achieve amateur‑level play.<p>Now a multi‑institutional team has broken that barrier. Their creation, Ataraxos, faced off against Pim Niemeijer, widely regarded as the strongest Stratego player alive and a multiple‑time world champion. Over three weeks of online play Niemeijer managed just a single win, while Ataraxos prevailed in fifteen games with four draws. The victory margin is striking given that top‑level human contests are usually decided by narrow margins.<p>The researchers attribute the success to two key innovations. First, Ataraxos learns through self‑play, reinforcing winning moves and discarding losing ones, much like DeepNash. Second, and critically, it adds a belief‑model neural network that estimates the opponent’s hidden pieces from observed behavior. Rather than exhaustively searching the astronomical game tree, the belief model samples plausible enemy configurations, simulates candidate moves in each, and selects the action with the best expected outcome. This look‑ahead capability, absent in DeepNash, lets Ataraxos plan several turns ahead despite the hidden information.<p>Training was remarkably efficient. Ataraxos required 16 GPUs for a week of core self‑play, plus an additional four GPUs for four days to train the belief model—altogether a few thousand dollars in cloud compute. By contrast, DeepNash reportedly consumed 1,024 of Google’s specialized TPU‑like chips for two to three months, a cost estimated at $3‑4.5 million. The new algorithm also needed about thirty‑four times fewer training games to reach a stronger performance level.<p>Beyond Stratego, the same architecture has already beaten world champions at faster variants such as Barrage Stratego and mastered cooperative games like Hanabi and the Chinese card game dou dizhu. The researchers see the approach as a step toward tackling real‑world problems that involve hidden information and long horizons—negotiations, financial markets, or military planning—by first constructing a simplified, game‑like model and then applying the belief‑enhanced reinforcement‑learning loop.<p>Although Ataraxos plays with a calm, methodical style that avoids impulsive gambles, it still cannot explain its decisions in human‑readable terms. The team acknowledges that making the strategy interpretable remains an open challenge. For now, the result stands as a clear milestone: AI has finally conquered a game that once seemed immune to automation, doing so with far less firepower than earlier attempts.
This image depicts a Stratego bomb piece, helping readers visualize the hidden information aspect of the game that the AI had to infer.