An AI Went 15-1 Against a Top Stratego Player in a Game of Hidden Pieces
Ataraxos infers what the opponent's concealed pieces are before choosing its next move. The researchers say it reached a higher level of play than the earlier DeepNash while using fewer training resources.
Researchers from MIT, Carnegie Mellon University, New York University and Stanford University have unveiled “Ataraxos,” an AI that plays the board game Stratego. The headline result in the paper, released on September 30, comes from a match against one of the world’s strongest human players. Across the 20 games the team published, Ataraxos went 15-1-4 against Pim Niemeijer.
How do you play without knowing the identities of your opponent’s pieces?
In Stratego, the identity of your opponent’s pieces is hidden. The AI has to choose its next move without knowing which pieces the other side is holding. Ataraxos was built for exactly this kind of game, where part of the information needed to make a decision is invisible. Its central innovation is the addition of a separate AI model that infers what the hidden pieces are.
Ataraxos first learns a baseline strategy by playing against itself repeatedly. Once an actual game begins, it estimates probabilities for what type each of the opponent’s pieces might be. It then works through the likely consequences of each available move before settling on its next action. In other words, it layers information gathered mid-game and fresh calculation on top of a strategy learned in advance.
What changed from DeepNash?
This is not the first time an AI has played Stratego at a high level. DeepNash, released by Google DeepMind in 2022, also learned its strategy through self-play. It ranked among the top three in the all-time rankings of an online Stratego platform.
The Ataraxos team added a layer that infers hidden information during play and then re-examines candidate moves. The result, they report, is stronger play than DeepNash achieved while using fewer training examples and fewer self-play games. Alongside the win-loss record, the team points to that reduction in training resources as a key result.
According to the researchers, training Ataraxos cost in the thousands of dollars. They also tested the same approach on Barrage Stratego, Hanabi and Doudizhu.
What you need to run it yourself
The materials needed to examine or reproduce the work are publicly available. The paper is open to anyone, and a public repository hosts the code along with the training method. The repository specifies a runtime environment of Linux and a GPU with CUDA 11.8 support. Assembling that kind of research setup is the starting point for anyone who wants to run it directly.
The significance of the work lies in combining inference with move evaluation to raise performance in situations where information is incomplete. The domain in which those gains have been demonstrated, however, is games with clear rules and clear outcomes. The researchers note that applying the approach to real-world decision-making will require further work on making the AI’s reasoning understandable and verifiable by people.