A hybrid AI just beat eight world champions at bridge – and explains how it did it

Bridge champion Sharon Osberg once wrote, “Playing bridge is like running a business. It’s about chase, chase, nuance, deception, reward, danger, cooperation and on good days, victory.”

While it’s no surprise that chess fell to number-crunching supercomputers long ago, one would expect humans to retain a more unassailable advantage in bridge, a game of imperfect information, cooperation, and shrewd communication. Over thousands of years, our brains have evolved to read subtle facial features and body language. We have built sprawling societies that depend on the competition and collaboration of millions. Surely such abilities are beyond the reach of machines?

First yes. But maybe not forever. In recent years, the most advanced AI has begun to encroach on some of our proudest territories; the ability to navigate an uncertain world where information is limited, play is infinitely nuanced, and no one succeeds alone.

Last week, French startup NukkAI took another step when its bridge-playing AI NooK ousted eight world bridge champions in a competition in Paris.

The game was simplified, and NooK didn’t exactly compete against the human players – more on that below – but the performance of the algorithm was otherwise spectacular. In particular, NooK is a type of hybrid algorithm that combines symbolic (or rule-based) AI with the deep learning approach that is dominant today. Also, unlike its pure deep learning peers, NooK is more transparent and able to explain its actions.

“What we have seen represents a fundamentally important advance in the state of artificial intelligence systems,” said Stephen Muggleton, professor of machine learning at Imperial College London The guard. In other words, not bad for a cold, calculating computer.

Black box, white box

To play bridge, perhaps the most challenging card or board game ever mastered by AI, the NukkAI team combined deep reinforcement learning with symbolic AI, an approach notoriously used by IBM’s Deep Blue to Garry Kasparov in the 90s years to beat at chess.

Deep reinforcement learning algorithms consist of a network of interconnected artificial neurons. To learn a game, an algorithm plays itself billions of times, assessing its performance after each round and incrementally improving by tuning and retuning its neural connections until it finally masters the game.

Symbolic AI, on the other hand, is rule-based. Software engineers hardcode the rules that AI must know in order to be successful. This can be, for example, that a bishop can move any number of squares diagonally on a chessboard, or that if an opponent is pursuing a certain strategy, a counter-strategy increases the chances of winning. This approach is fine for the finite, but as the space of all possible moves in complex games increases, it becomes untenable.

That’s why Go World Champion Lee Sedol’s loss to DeepMind’s AlphaGo in 2016 was a big deal. Back then, experts didn’t expect that the AI ​​would beat the best Go players for a decade. AlphaGo showed the surprising power of deep learning compared to “good old fashioned AI”.

But deep learning has its downsides. One of them is that it is a “black box”. How the billions of nodes in a neural network perform a specific task is mysterious.

AlphaGo’s move 37 against Lee Sedol was a decision no human would make – it calculated the odds that a pro would have picked that move at 1 in 10,000 – but it made the move anyway and won. Still, the algorithm couldn’t explain it What in his training conveys his trust. This opacity is a problem when the stakes are higher than in a board game. To trust self-driving cars or medical algorithms making life-or-death decisions and diagnoses, we need to understand their reasons.

One potential solution, championed by researchers like NukkAI, would blend deep learning and symbolic AI, exploiting each individual’s strengths in what is known as a “neurosymbolic” approach.

NooK, for example, first learns the rules of the game and then playfully improves its skills. The combination refines the probabilistic ‘brain’ of the algorithm,” Muggleton said The Telegraphtaking it beyond statistics. NooK, he said, “uses background knowledge in a similar way to how we augment our own learning with information from books and past experiences.” This allows the algorithm to explain decisions: It is a “white box” AI.

Because of this, Bridge – a communication and strategy game that has resisted AI conquest – is a great test of this approach. “You can’t play bridge if you don’t explain it,” said Véronique Ventos, co-founder of NukkAI The guard.

There are bridge game algorithms, but they don’t hold the water against the best people. At the NukkAI competition in Paris a little over a week ago, the situation seems to have changed.

fun and games

In the NukkAI Challenge, NooK faced off against eight world bridge champions.

Each champion played ten sets of ten games, while NooK played 80 sets of ten games, or 800 straight deals. Instead of playing against each other, the human and AI played the same hands against the same opponents, a pair of bridge bots (not built by NukkAI) called Wbridge5.

A bridge game begins with players bidding how many tricks or turns they think they can win. The highest bid is called the contract, and the one who sets the contract is the declarer. Declarer’s partner or dummy places their hand face up on the table and leaves the game. The declarer now plays against his opponents with both hands, trying to win enough tricks to fulfill his bid.

The NukkAI Challenge removed the bidding to simplify play, and both humans and NooK took on the role of declarer in each game, with the bridge-bot pair acting as opponents (or defenders). The difference between NooK’s score and each human player’s score was averaged over each set. NooK beat its rivals in 67 or 83 percent of the 80 sets played.

“It’s pretty desperate for people,” said French champion Thomas Bessis. “Sometimes we don’t understand why the AI ​​plays better than us – but it is. It’s very frustrating.”

NooK’s win is an impressive feat, but there are caveats. Skipping the bidding process and playing only the declarer role removed challenging and nuanced parts of the game where partners must communicate with each other and deceive their opponents. Also, staying focused for 100 straight hands is challenging for a human, but not so for a computer. Finally, Jean-Baptiste Fantun, co-founder of NukkAI, said he was confident the machine would pull through on thousands of deals, but was less confident about its prospects above just 800. In other words, the more he plays, the better his chances of winning. Playing many hands in a row might have helped the AI ​​crowd out the humans in this case.

“There are other things to solve for bridge as well,” said Fantun. “We still have a roadmap ahead of us.” That said, it’s going too far to say that bridge has fallen victim to AI like chess or go. But having the AI ​​outperform the best human players in one part of the game is a major milestone on Fantun’s map. And while larger AI algorithms like OpenAI’s GPT-3 continue to impress, NukkAI’s performance in the bridge space may lend more weight to the case for a hybrid approach.

Next, they need to show that NooK can play and win – no disclaimers needed.

Credit: TATIANA/Unsplash

Are you looking for ways to stay ahead of the pace of change? Reconsider what is possible. Join a highly curated, exclusive cohort of 80 executives for Singularity’s flagship Executive Program (EP), a five-day, fully immersive executive transformation program that disrupts existing mindsets. Discover a new way of thinking, a new toolset and a network of fellow futurists dedicated to finding solutions to the rapidly changing world. Click here to learn more and apply today!

Leave a Reply

Your email address will not be published.