At the end, Jean-Baptiste Fantun was almost apologetic. His technology company, NukkAI, had just wiped the floor with a gathering of the world’s elite bridge players, using only an artificial intelligence algorithm. Now he was summoning the defeated champions onto the stage. Above their heads, a projector beamed the final result onto a giant screen: Human Score: 5,238. Robot Score: 6,147. “Well, it’s just a game,” said Fantun.
Everyone knew that was a lie.
Instead, the Nukkai Challenge, held recently in Paris, marked the breaching of another of those thresholds which we humans hold so dear, another puncturing of our collective desire to believe that our species is unique and will always be capable of tasks that computers can never accomplish.
It was a process that began, famously, in 1996, when IBM’s Deep Blue beat then world chess champion Garry Kasparov. The game of Go followed two decades later, when Lee Sedol was humanity’s humbled representative. In the meantime, many were the flesh and blood frontiers to fall to the might of the silicon chip. From poker to quiz games, computers are now routinely better than us at pursuits and pastimes that used to seem quintessentially human.
Now bridge has fallen. Not bridge! Even Microsoft founder Bill Gates – a keen player – considered bridge well-protected from the brute calculating power of the machine. After all, isn’t it a social game, a game of partnership and instinct as well as raw probabilities? Isn’t it a game where, unlike chess and Go, most of the information – the cards – is not open and available on the board, but hidden from view? Surely, not just calculation but experience and intuition mattered?
And so it turned out. Even as computers mastered countless other domains, they proved no match for the best human bridge players. But then four years ago Fantun and NukkAI’s co-founder Veronique Ventos set about changing all that. They focused on what is called in the AI field a “neuro symbolic” approach, which combines two very different paths to learning. The first makes essentially random choices, stumbling to success in countless games, and learning as it goes. This was the approach taken by Google’s subsidiary, Deepmind, in its AlphaGo algorithm which beat Lee Sedol. The second, “symbolic” approach, is rules-based, and is effectively how children learn at school. “Someone explains it to us,” says David Beavan, principal research software engineer at The Alan Turing Institute, before adding: “The combination is important.”
According to Stephen Muggleton, professor of machine learning at Imperial College London, the blended approach allowed NukkAI’s algorithm to grasp more than just the statistical realities of bridge. Brute probabilistic calculation was, in other words, being refined to cope with the game’s ambiguities and uncertainties by using “background knowledge much in the way that we augment our own learning with information from books and from our previous experience”.
“This is learning based on incomplete information,” Muggleton added, as the scores were revealed. “What we’ve seen today represents a fundamentally important advance in the state of artificial intelligence systems.”
Suddenly, the nuances and hidden information that makes bridge so compelling are no barrier to computer mastery. Still, as the eight champions gathered late on March, some could still not believe that they were about to meet their match.
“I have played against robots and they have been good, but I always think I can beat them,” said serial English champion Nevena Senior ahead of the challenge. “I’ve not seen a robot that’s better than me.” In the end, she didn’t win a single hand.
French champion Thomas Bessis described the carnage. “It’s pretty desperate for the humans,” he said. “There are just times that we don’t understand why the AI is playing better than us – but it is. It’s very frustrating.”
At least he managed to keep a broad smile on his face as he related details of the drubbing. The Russian player Anna Gulevich, by contrast, looked shattered. Alone among the eight champions, she seemed to be on the brink of beating the AI going into the final hand, only to be pipped at the last. “It was very emotional,” she said. “[The AI] lacks strategy, and humans have strategic vision.” This she had attempted to display by playing an unorthodox 10 of diamonds at a crucial moment. But it didn’t come off as she’d hoped. “I need to take my revenge,” she sighed.
“It’s a significant result,” says Tom Townsend, The Telegraph’s bridge correspondent, speaking from Rome where he is competing in the World Championships. There, the result has created remarkably little stir, as competitors focus on going head to head, human-to-human, in the traditional way. But he knows the result’s significance will be impossible to ignore completely. “If you’d asked me before, I would have bet on the humans,” says Townsend. “These are all world-class players.”
Nonetheless, there is hope for mankind. Perhaps the most quixotic part of bridge – bidding – was excluded from the challenge, with each auction set instead at three no trumps. Whereas chess computers are now utterly indomitable, once bidding is factored in, says Townsend, “we’re nowhere near that stage in bridge”.
That may have pros and cons for the average player. Part of what has made Deepmind chess and Go computers so devastating is that they learn without reference to established examples and norms. But this means that on occasion they produce utterly new moves or tactics. Hence the now celebrated “Move 37”, with which the AlphaGo computer completely baffled onlookers against Lee Sedol. Was the outlandish feint a good or bad move? No one knew at the time. But it turned out to be a bravura turn. Essentially, a new way of playing the 3,000-year-old game had been created. Far from “killing” Go with its processing power, AlphaGo had helped rejuvenate it.
Such benefits to bridge may be to come. But NukkAI’s approach is already driving progress elsewhere. That’s because, unlike many AI algorithms, known as “black boxes”, it is able to explain why it has made the move it has. Such “white box” techniques are likely to be critical for the widespread adoption of certain technologies, like driverless cars. If there’s not to be a big pile-up, for example, it’s essential that one AI-piloted autonomous vehicle at a junction is able to explain to other vehicles why it is taking the path it is.
More importantly still, say experts, is that AI machines can explain to humans why they are doing what they are doing. “AI systems need to win our trust, just as other humans do. And one of the best ways of doing that is explaining decisions,” says Beavan. “The more they explain, the more we can trust them. Until they can do that, AI integration into human society might be quite limited.”
No wonder, then, that as Fantun sheepishly brought the curtain down on proceedings in Paris, half-thanking his human participants, half-apologising to them, he conceded that it’s a “small step for bridge. But it’s a big step for next-generation artificial intelligence.”
The Daily Telegraph
