Researchers say their AI beat top Stratego champion on a small budget

In short: Researchers built an AI called Ataraxos that beat a leading Stratego player, and it cost only a few thousand dollars to train.

What happened

A research team from Carnegie Mellon, MIT, New York University, and Stanford University reported that their AI system, Ataraxos, defeated Pim Niemeijer in online Stratego matches. Niemeijer is widely seen as one of the best Stratego players ever, with four world championships and more than 600 weeks ranked number one.

Across 20 games played over three weeks, Ataraxos won 15 games, lost 1, and drew 4. Niemeijer was paid $100 for each win. The researchers say some losses can still happen because Stratego includes luck, like how your pieces are arranged at the start.

Stratego is hard for computers because you can see where your opponent’s pieces are, but not what they are until they fight. It is like playing poker where you can see where the other person placed their cards, but they stay face down for a long time. There are also an enormous number of possible starting setups, and games can last up to thousands of moves.

The key change was adding a second neural network, a “belief model,” that guesses which hidden pieces the opponent probably has based on their moves. This let Ataraxos “think ahead” by testing a smaller set of realistic possibilities, instead of trying to consider every possible board.

Why it matters

This result suggests strong game-playing AI may not always require huge budgets. The team says Ataraxos trained using 16 GPUs for about a week, far less than earlier systems that used large, expensive computing clusters. That could make this kind of research more accessible to universities and smaller labs.

Source: Arstechnica