The games AI conquered

Deep Blue beat Garry Kasparov at chess in nineteen ninety-seven. AlphaGo beat Lee Sedol at Go in two thousand sixteen. Poker bots have beaten professionals for years. But Stratego held out, even against DeepMind.

Meet Ataraxos

A team from Carnegie Mellon, MIT, New York University, and Stanford University built Ataraxos. It beat Pim Niemeijer, arguably the best Stratego player of all time, fifteen games to one, with four draws.

Why Stratego was hard

In Stratego, each player gets forty pieces. You win by capturing the opponent's flag, but you don't know what pieces they have. Identities reveal only when pieces collide in battle.

The hidden information problem

In Texas Hold'em poker, there are just one thousand three hundred twenty-six possible hands. In Stratego, forty pieces could be in any order, creating more than a decillion possible setups.

Games last much longer

Chess games usually last forty moves. Stratego games can easily last two thousand moves. The game also requires bluffing: moving weak pieces as if they were strong to scare opponents.

How Ataraxos learned

Ataraxos learned by playing against itself one hundred sixty-three million games in total. Winning moves were reinforced and played more often. Losing moves were played less often.

The belief model breakthrough

Ataraxos used a second neural network, a belief model, trained to guess the opponent's hidden pieces based on their moves. Instead of checking every possible arrangement, it sampled plausible ones and picked the best move.

Calm and unbothered

Ataraxos comes from the ancient Greek word for a state of calm. The AI doesn't react impulsively like humans do. It works methodically back into games from significant deficits without emotional desperation.

“It means somebody that's calm and unbothered.”

The match with Niemeijer

Over three weeks, Pim Niemeijer played twenty online games against Ataraxos, earning one hundred dollars for each win. He had four world championships and more than six hundred weeks as the world's top-ranked player. He won just once.

The cost difference

DeepNash was trained for two to three months on one thousand twenty-four of Google's specialized chips, costing three million to four point five million dollars at two thousand twenty-five prices. Ataraxos needed sixteen GPUs for a week, plus four GPUs for four days.

Efficiency and speed

The Ataraxos team wrote a simulator running millions of moves per second on graphics cards. The algorithm learned far faster, playing about thirty-four times fewer games than DeepNash while ending up much stronger.

Beyond board games

The Ataraxos architecture also beat three world champions at Barrage Stratego, mastered the cooperative card game Hanabi, and beat the best bots at the Chinese card game dou dizhu. The team now wants to apply it to real-world problems.