An artificial intelligence has beaten eight world champions at bridge, a game in which human supremacy has resisted the march of the machines until now.
The victory represents a new milestone for AI because in bridge players work with incomplete information and must react to the behaviour of several other players – a scenario far closer to human decision-making.
In contrast, chess and Go – in both of which AIs have already beaten human champions – a player has a single opponent at a time and both are in possession of all the information.
“What we’ve seen represents a fundamentally important advance in the state of artificial intelligence systems,” said Stephen Muggleton, a professor of machine learning at Imperial College London.
French startup NukkAI announced the news of its AI’s victory on Friday, at the end of a two-day tournament in Paris.
The NukkAI challenge required the human champions to play 800 consecutive deals divided into 80 sets of 10. It did not involve the initial bidding component of the game during which players arrive at a contract that they must then meet by playing their cards.
Each champion played their own and their “dummy” partner’s cards against a pair of opponents. These opponents were the best robot champions in the world to date – robots that have won many robot competitions but that are universally acknowledged to be nowhere near as good as expert human players.
The AI – called NooK – played the same role as the human champion, with the same cards and the same opponents. The score was the difference between those of the human and the AI, averaged over each set. NooK won 67, or 83%, of the 80 sets.
Jean-Baptiste Fantun, co-founder of NukkAI, said he had been confident the machine – which the company has been developing for five years – would triumph in thousands of deals, but with only 800 it was touch-and-go.
Announcing the results, the mathematician Cédric Villani, winner of the Fields medal in 2010, called NukkAI “a superb French success story”.
AI researcher Véronique Ventos, NukkAI’s other co-founder, calls NooK a “new generation AI” because it explains its decisions as it goes along. “In bridge, you can’t play if you don’t explain,” she says.
The game relies on communication between partners.
Explainability is a hot topic in AI. “Most of what the general public have heard in recent years about machine learning is based on black box systems such as AlphaGo, which is unable to explain to human beings how decisions are being made,” said Muggleton.
Instead, NooK represents a “white box” or “neurosymbolic” approach. Rather than learning by playing billions of rounds of a game, it first learns the game’s rules and then improves its play through practice. It is a hybrid of rules-based and deep learning systems. “The NooK approach learns in a way that is much closer to human beings,” Muggleton said.
“The pendulum is swinging towards these kinds of methods,” says Michael Littman, a professor of computer science at Brown University in Rhode Island. “Not being able to tell people what’s going on just doesn’t work in our societies.”
Even if a person or an AI can’t explain in words what they are doing, Littman says, their behaviour needs to be “legible” to others – enacting rules they understand.
This will be critical in domains such as health and engineering. Self-driving cars negotiating a junction will need to be able to read each other’s behaviour, for example.
Littman was disappointed the challenge didn’t include bidding, which is where much of the most interesting communication – and deception – happens in bridge.
But Nevena Senior, a many-times world bridge champion for England and one of NooK’s challengers, said the contracts the humans and NooK were given to play were sufficiently variable that the card game became as important as the bidding.
She said NooK’s creators had done a “magnificent” job. She found that it read its opponents better than the humans did, and was better able to exploit their mistakes.
“This is something that humans do after enough experience and I was pleasantly surprised that a robot mimics typical human skills,” she said.
Other AI triumphs
1996: IBM’s Deep Blue chess machine wins a game against world chess champion Garry Kasparov but loses the match 2-4. A year later, Kasparov loses the rematch.
2007: Checkers is solved by researchers at the University of Alberta in Canada. After sifting through 500bn positions, they build a checkers-playing computer programme that can’t be beaten.
2011: IBM’s Watson computer defeats TV gameshow Jeopardy! champions Brad Rutter and Ken Jennings, claiming the $1m first prize.
2016: Google DeepMind’s AlphaGo defeats Korean Go champion Lee Sedol 4-1. The Korea Baduk Association awards AlphaGo the highest Go grandmaster rank, an honorary 9 dan.
2022: NukkAI’s bridge-playing computer NooK defeats eight world bridge champions in Paris.
