With most information hidden, the game Stratego had stumped AI until now
Researchers built an AI called Ataraxos that defeated the world's best Stratego player using significantly less compute than previous attempts.
summary is enough Because the source article is unavailable, the comment thread and title provide a clear overview of the Ataraxos AI milestone and its efficiency gains.
The article could not be read automatically; this analysis is based on the title and discussion.
summary
Although the linked article text is unavailable, discussion details show that researchers from CMU, MIT, NYU, and Stanford developed an AI named Ataraxos that defeated top human Stratego player Pim Niemeijer 15 to 1. Improving on DeepMind's 2022 DeepNash effort, Ataraxos achieved superior performance while requiring roughly 34 times less training data and running on just 16 GPUs. Commenters analyze how the algorithm navigates hidden information and psychological bluffing, noting unconventional tactics like tucking the flag into a corner behind two bombs.
key points
- Ataraxos defeated top human Stratego player Pim Niemeijer in a match score of 15 wins, 1 loss, and 4 draws.
- The model required 34 times fewer training games than DeepMind's DeepNash and operated on a modest setup of 16 GPUs.
- The AI adopted rare human strategies, such as protecting its flag in a corner with only two bombs.
who should read it
AI researchers and board game enthusiasts interested in reinforcement learning for imperfect-information environments.
discussion positive
consensus
Commenters are impressed by the dramatic drop in compute requirements compared to DeepMind's previous attempts and dive deep into how reinforcement learning handles imperfect information and bluffing.
pushback
A few participants express disappointment that solving these games alters and commodifies the organic human experience of competitive board gaming.
> The algorithm also learned far faster—it played about 34 times fewer games than DeepNash, and still ended up much stronger.
Imo, this is the critical piece and what makes the AI work at all.
With hidden information games, the best move depends on information you don’t have. So a move could be good or bad, it just depends on something that’s impossible to know.
You’d like to search ahead, meaning “if I do this they will do that” but that’s impossible since you don’t even know what the opponent can do because you don’t know their hidden state.
If the possible hidden states are randomly …
It breaks down the precise mathematical and strategic difficulty of predicting moves in hidden-information games where standard tree searches fail.
Generated by an LLM from the article and thread. Check the sources before relying on it.