Ataraxos AI Achieves Superhuman Performance in Stratego with Low Compute Cost

Stratego is an "imperfect-information" game, meaning players cannot see the identities of their opponent's pieces until they engage in battle. This creates a massive state space and requires complex bluffing and long-term strategic planning. Unlike perfect-information games like Chess or Go, where all information is visible, the hidden nature of Stratego has historically made it difficult for AI to achieve superhuman performance.

Ataraxos overcomes these challenges through a novel design that combines self-play reinforcement learning with test-time search. The system utilizes two interdependent Transformer-based networks: a set-up network to determine initial piece arrangements and a move network to select actions. A key innovation is the inclusion of a "belief model," a second neural network trained to estimate the likely identities of an opponent's hidden pieces based on their movement patterns. This allows the AI to perform a search by sampling plausible game states and selecting the best move across them.

Notably, Ataraxos achieved these results with extreme efficiency. While previous AI efforts, such as DeepMind's DeepNash, required massive computational resources—estimated at $3 million to $4.5 million for training—Ataraxos was trained using just 16 NVIDIA H100 GPUs for one week and additional GPUs for the belief model, costing less than $8,000.