Reference
Sources and reading guide
| Source | What it contributes | Boundary of the evidence |
|---|---|---|
| Sokota, Vinitsky, Hu, Kolter and Farina, Superhuman AI for Stratego Using Self-Play Reinforcement Learning and Test-Time Search, 2025 | The Ataraxos method: setup, move and belief transformers, dynamically damped self-play, test-time search by update equivalence, the GPU simulator, and the 15-1-4 series against Pim Niemeijer | Its numbers come from 16 H100s for a week plus 4 H100s for 4 days; nothing in this notebook has reproduced them yet |
| Perolat et al., Mastering the game of Stratego with model-free multiagent reinforcement learning, Science 2022 | DeepNash: Regularised Nash Dynamics with a reward transformation and replicator dynamics, the 10^535 state count, 84% on Gravon and top-3 rankings | Trained on 1024 TPU nodes according to the Ataraxos authors; the code is no longer functional, so no direct comparison exists |
| International Stratego Federation rules, chapter 10 (extract in Han Wolf, A model for capture by the two-squares rule, 2012) | The two-square rule as three non-stop moves between the same two squares, and its extension to Scouts by the squares they pass over | A static look-ahead model of captures under the rule, not a general engine |
| Strategus rules | The online platform's wording of the two-square, continuous-chase and 200-move rules used for the Ataraxos evaluation | The chase rule text is informal; the engine implements the paper's formal definition |
| Gravon rules | Classic rules and the platform DeepNash was evaluated on | Its two-square wording (five consecutive turns) differs from the ISF wording the engine follows |
| Cowling, Powley and Whitehouse, Information Set Monte Carlo Tree Search, 2012 | Determinization and its known weakness, strategy fusion, which the expectimax rung inherits | Results in other games do not transfer to this implementation |
| Zhang and Sandholm, General search techniques without common knowledge for imperfect-information games, and application to superhuman Fog of War chess, 2025 | Obscuro: knowledge-limited unfrozen subgame solving over a sampled information set, one-sided GT-CFR, predictive CFR+ on the last iterate, purification; 80% against the top human | Fog of War chess enumerates its information set (up to a million positions); Stratego's must be sampled, which is the open question of the subgame-solving approach |
| Liu, Fu, Fu and Yang, Opponent-Limited Online Search for Imperfect Information Games, ICML 2023 | Safe-1-KLSS and the internal gift; OLSS-I and OLSS-II, which limit the opponent's reach and strategy set to shrink the subgame | Evaluated in poker variants and two-player Mahjong with tabular blueprints; no result in a game of Stratego's size |
| Holeček, NashDreamer: Model-Based Reinforcement Learning for Zero-Sum Imperfect-Information Games, 2026 | A centralised recurrent state-space world model with per-player infoset models and R-NaD in imagination; sample efficiency in Goofspiel and Battleship | Games of a few dozen plies with full unrolls from the first move; the notebook's environment is nearly free, so only the infoset model is tested here |
Read the Ataraxos paper alongside the concept pages from
self-play reinforcement learning to
test-time search, and DeepNash alongside
DeepNash and regularised Nash dynamics. The
engine's rule choices are in rules and anti-chase.
The PDFs are vendored under paper/. Sources were checked on
September 18, 2026.