Researchers from Carnegie Mellon, MIT, NYU, and Stanford developed Ataraxos, an AI capable of playing the board game Stratego. The system defeated Pim Niemeijer, a four-time world champion, in 15 out of 20 games, with four draws and one loss.
Ataraxos utilizes a self-play training method, having played 163 million games against itself. Unlike previous models like DeepNash, it incorporates a belief model that samples plausible opponent setups, allowing the AI to think ahead before making a move.
The training process for Ataraxos required only 16 GPUs over one week, plus four additional GPUs for four days to train the belief model. This is significantly more efficient than DeepNash, which required 1,024 specialized chips for several months.
The AI demonstrated versatility by mastering other games, including Barrage Stratego, Hanabi, and dou dizhu. Researchers suggest the underlying techniques could eventually be applied to real-world scenarios like negotiations, financial markets, and military wargaming.
Future development will focus on making the AI's decision-making process more interpretable. Currently, Ataraxos can execute complex strategies, including effective bluffing, but it cannot explain the reasoning behind its specific moves.
