Computers mastered chess and Go years ago. Stratego, the classic board game where you can’t see what your opponent’s pieces are, held out much longer. Not anymore.
A team from MIT, Carnegie Mellon, NYU and Stanford has built an AI called Ataraxos that beat Pim Niemeijer — described in the research as the best Stratego player ever — with 15 wins, one loss and four draws over 20 games. The work was published in Nature on Sept. 30.
Why Stratego is so hard for AI
In chess, both players see the whole board. In Stratego, every piece starts face-down, so you have to guess what you’re attacking, bluff about your own pieces, and keep updating your guesses as the game unfolds.
“With Stratego, there is an explosion of possible universes you might have to deal with,” said senior author Gabriele Farina, an assistant professor at MIT, in the university’s announcement.
How it works
Ataraxos combines a few ideas:
- Self-play: it learned a basic “blueprint” strategy by playing itself over and over.
- Belief modeling: it estimates what the opponent’s hidden pieces probably are, based on what has happened so far.
- Planning at move time: before each move, it samples possible versions of the hidden board and refines its choice.
According to MIT, the system also went 39–2 against top championship players and plays stronger than DeepNash, the Stratego AI Google DeepMind unveiled in 2022.
The real headline: cost
DeepNash took a huge amount of computing power. Ataraxos was trained on 16 Nvidia H100 chips for about a week — a few thousand dollars of compute, according to the Nature paper. MIT says it used less than a hundredth of DeepNash’s training examples.
The same techniques also reached top results in other hidden-information games, including the card games Hanabi and dou dizhu.
Why it matters beyond games
Real life is full of hidden information: negotiations, cybersecurity, military planning. The researchers see those as possible future uses. Farina also noted that Ataraxos is good at weighing risk in a way people often aren’t, because it doesn’t panic.
The catch
The researchers acknowledge that Ataraxos can’t explain its decisions in terms people can follow. Before anything like it is used in high-stakes settings, they say, it would need ways for humans to audit and understand its choices.
More on Contoh
- A Light-Powered AI Can Screen Deepfake Videos With Almost 98% Accuracy
- Why AI Chatbots Make Things Up — and How to Catch Them When They Do
- Free Gemini Users Will Be Limited to Google’s Smallest Model Starting Oct. 9