MIT-led team's AI Ataraxos beats top Stratego player Pim Niemeijer 15-1-4

The team reported that Ataraxos was trained for less than $8,000 using academic computing resources — specifically one week on 16 Nvidia H100 GPUs and four additional days on four GPUs for its belief network — which they estimate is roughly 1/500th the cost of DeepNash's training. The researchers highlighted innovations in self-play reinforcement learning and decision-time planning using a generative model to estimate hidden opponent pieces, enabling efficient learning without relying on vast amounts of data or compute.

Beyond Stratego, the researchers adapted Ataraxos to other imperfect-information games such as Barrage Stratego, Hanabi, and Dou dizhu, achieving superhuman results in each. They argue that their approach shows large amounts of hidden information are no longer a barrier for reinforcement learning, potentially expanding AI's applicability to real-world strategic decisions in fields like business and cybersecurity.