Ataraxos AI beats the world's top Stratego player by record margin
Researchers from MIT, CMU, NYU and Stanford report in Nature that their AI Ataraxos beat the strongest Stratego player 15-1-4 using far less training data.
A team of researchers has built an AI called Ataraxos that defeated the world's strongest Stratego player with a record score of 15 wins, one loss and four draws. The work was published in Nature and announced by MIT on 30 September.
Stratego is a board game of hidden information: each player knows where their own pieces are but not what the opponent's pieces are. That makes it far harder for computers than chess, where both sides can see the whole board.
Key details
- Ataraxos also posted a 39-2 record against top championship players.
- The researchers say it needed less than one hundredth of the training examples and less than one thirtieth of the self-play games used by DeepMind's earlier Stratego system.
- It combines self-play reinforcement learning with planning at decision time, using a generative model to guess where the opponent's pieces are likely to be.
- The same approach was adapted to other imperfect-information games, including Barrage Stratego, Hanabi and the Chinese card game Dou dizhu.
The project brought together researchers from Carnegie Mellon, MIT, New York University and Stanford. Samuel Sokota of Carnegie Mellon is the lead author and Gabriele Farina of MIT is the senior author.
Why it matters
Hidden-information games such as card games have long been tougher for AI than chess or Go. A method that works with much less computing power could, over time, make stronger and more human-like opponents more practical in everyday card and board games, including on mobile.
Source: MIT News
