With most information hidden, the game Stratego had stumped AI until now
AI researchers have built a Stratego-playing system, Ataraxos, that decisively beats one of the strongest human players while using far less compute and training data than DeepMind’s earlier 2022 effort. Commenters dig into why imperfect-information games like Stratego, Hanabi, poker, bridge and trading card games are hard for machines, debating how techniques such as reinforcement learning, search, and Nash-equilibrium approximations scale to large hidden-information spaces and frequently changing rulesets. Others reflect on how superhuman game AIs reshape human play and culture—citing chess and Go—while sharing nostalgia for physical board games and variants like Electronic Stratego.
Comparison with earlier Stratego AIs
- Several comments contrast the new Ataraxos system with DeepMind’s 2022 DeepNash work.
- Key claimed improvements: roughly two orders of magnitude less training data, far fewer self-play games, yet stronger play against top humans.
- Some see this as “actual mastery” of Stratego, suggesting the earlier “mastering” label was premature.
- The relatively low training cost (16 GPUs, a few thousand dollars) is highlighted, especially versus DeepMind’s budget.
Hidden-information and variant rules
- Stratego’s fog-of-war memory demands (humans forgetting revealed pieces vs AIs never forgetting) are noted.
- People are curious how the new method would handle harder variants like “silent defense” and Electronic Stratego, where defenders don’t reveal piece strength and information uncertainty is higher.
- A linked game archive is mentioned for studying Ataraxos’s play.
Broader game-AI landscape
- Comparisons to StarCraft, Dota, Advance Wars, poker, bridge, Hanabi, Hearthstone, MTG, and custom card games.
- Ongoing debates:
- Whether hidden-information, stochastic card games with evolving “metas” are significantly harder for AI than static board games.
- Counterarguments that generalization over card interactions and Nash-equilibrium approximations (as in poker) should scale to MTG-like games, given enough engineering.
- Bridge and complex partnership communication are flagged as a particularly interesting unsolved AI challenge.
Skill, randomness, and fairness
- Discussion of how imperfect information and randomness allow novices to occasionally beat experts (poker, rock-paper-scissors) vs deterministic depth in chess/Go.
- For RTS bots, action-speed and precision advantages (APM/EPM, micro) are seen as making “fair” human–AI comparison tricky.
Impact on human play and game culture
- Mixed reactions: some are excited about stronger tools and new strategies; others worry that AI transforms competitive games (as in chess and Go) and diminishes the “journey.”
- Several nostalgic accounts of childhood Stratego, preferred setups, psychological play, and house variants.
- Some note that deep board games often lose casual opponents once one player becomes much stronger, and that online/digital implementations partially mitigate this by offering better matchmaking and rules enforcement.