An artificial intelligence just beat 8 world champions at bridge

An artificial intelligence has bested human champions in a modified version of bridge — a card game that's notoriously difficult for machines to master.

The tournament was hosted by the French company NukkAI, which invited eight bridge champions to pit their skills against NooK, an AI it designed specifically for that purpose.

"At the beginning of the challenge, they arrived and they were very confident," NukkAI CEO Jean-Baptiste Fantun, a mathematician and prominent bridge player himself, told As It Happens guest host Gillian Findlay.

"After one or two sets of 10 deals, they really began to worry a little bit."

Watch: 8 bridge champions compete against an AI (commentary is in French)

The bridge champions had every right to be cocky. While computers have long been able to beat humans at a variety of games — including chess, checkers and Go — Fantun says nobody has managed to create an AI that can defeat a bridge champion before now. 

"What we've seen represents a fundamentally important advance in the state of artificial intelligence systems," Stephen Muggleton, a professor of machine learning at Imperial College London, told the Guardian newspaper.

The reason computers excel at chess but struggle with bridge is twofold, says Fantun. Firstly, bridge is a very social game that can involve both collaboration and deceit. Secondly, it has an element of mystery.

For example, in chess or Go, the game is one-on-one, and both players can see everything on the board. There are no secrets. 

But in bridge, your opponents can work together against you, and you can't see their cards. That makes it hard for an algorithm to make predictions about its opponent's next move. 

"That's something that makes bridge very close to … real life," Fantun said. "So it makes it a very difficult game for the computer."

Not your grandfather's bridge game

The tournament used a modified version of bridge. 

Normally, bridge is played between two pairs of partners, each with their own set of cards. But in this case, each human champion played alone, controlling both sets of cards, against a pair of AI bots called Wbridge5, which weren't created by NukkAI. 

The NooK AI, meanwhile, took on the exact same role as the human champions, playing the exact same cards against the exact same opponents.

Each of the eight human champions played 10 sets of 10 games, while Nook AI played 80 sets of 10 games, for a total of 800 consecutive deals each. The difference between the scores of the human and the AI were averaged over each set.

By the end, NooK won 67, or 83 per cent, of the 80 sets, making it the victor.

Despite the win, Fantun admits NooK hasn't mastered all elements of game. The tournament didn't include betting, which is usually a key component of bridge. What's more, unlike real bridge players, NooK didn't collaborate with a partner.

"So even in bridge, there are other things to be solved," Fantin said. "We still have a roadmap in front of us."

How did it win?

Most AIs that have beaten humans at games rely on what's called "black box" systems. They take in a lot data, then use that to make accurate predictions. 

But because so much of the "data" in bridge is hidden from the opponent, NukkAI designed NooK as a "white box" system. Instead of playing billions of games and learning to predict outcomes, the AI simply learned the rules of bridge and got better over time with practice. 

"Also, we have some small neural networks that help mimic the the opponent's behaviour. So, it's very different from what we have seen in other games," Fantun said.

This type of AI is, in many ways, less complex. But it has an important feature known in the tech world as "explainability."

"We can show why the bots made such and such a decision. When it's made the brilliant play, for example, we are able to show everybody, the bridge community, what was in the bot's minds, if you excuse my expression," Fantun said.

It's something that has implications way beyond bridge, he said. His company aims to build AIs that can work with, rather than replace, human beings. That requires AIs that can explain what they're doing and why. 

"We believe it will have a huge impact in all the areas where the human must remain in control for responsibility or ethical reasons … for example, defence, cybersecurity, health care," Fantun said. "All these are areas where we cannot leave the machine in control."

Written by Sheena Goodyear. Interview produced by Kate McGillivray.