Alphago And AI: The State Of Play

In what many are seeing as a major breakthrough, Google Deepmind’s AI system, AlphaGo, beat Lee Se-dol, the world’s greatest Go player of the last decade on Tuesday 15th March 2016. The final score was 4-1 to AlphaGo. Cath Elliston talks about what this means for humanity.

Backgammon, draughts, chess, and now the ancient Chinese board game of Go: developers of artificial intelligence have often turned to games. But, far from being a trivial pursuit, we should see this preoccupation as a powerful sign of increased AI capability.

AlphaGo is an AI system, designed by 15-20 people at Google Deepmind, the British artificial intelligence company acquired by Google in 2014. There are just two rules to Go, but it is considered one of the hardest games for machines to play. It combines logic with intuition and there are more possible configurations of the board than there are atoms in the universe.


How Does Alphago Work?

AlphaGo differs from other AI systems that have been used to beat human professionals in the way it was trained. First, AlphaGo was shown 1000 games played by human experts and trained to mimic and predict their moves. Then, by playing against older versions of itself 30 million times, it learnt to improve incrementally. The result was a new version that could beat its old self 80-90% of the time.

If Deep Blue, the chess-playing AI that beat world champion Garry Kasparov in 1997, was told to play another game – even a far easier game than chess – it would not know where to start. In contrast, the algorithms used in AlphaGo are more general purpose, and could potentially be applied to other games, and even other fields.


How Significant Would A Victory For Alphago Be?

When AlphaGo secured a victory over European Go Champion Fan Hui in October, Fan Hui himself expressed surprise. He was not the only one. In 2014, when Remi Coulom’s Go-playing program beat grandmaster Norimoto Yoda (with a four move head start) he predicted that it would be another ten years before machines beat the best players fair and square.

Experts are hugely excited about this development. In an excellent Newsnight piece, Rohan Silva called AlphaGo ‘the dawn of machines that think like humans’. Deepmind co-founder and CEO, Demis Hassabis, has similarly termed it ‘the first significant rung of the ladder towards proving a general learning system can work’.

Lee Se-dol felt optimistic ahead of the game. ‘I am confident I can win’, he says, but adds, ‘at least this time’. But his confidence was misplaced. AlphaGo even played ‘surprising and beautiful moves’ that astonished even Hassabis to secure its victory.


Where Do We Go From Here?

The next step, according to Hassabis, is applying the algorithms used in AlphaGo to fields like healthcare and science where they can help human experts to achieve more. In fact, Deepmind is currently working on health initiatives to improve the efficiency of the NHS, including an app to help nurses and doctors detect cases of acute kidney injury.

So, while artificial intelligence mastering a game of Go might seem unimportant, AlphaGo is surely another sign of progress that could bring radical changes to our society, sooner than you might think.

And this is where Future Advocacy comes in. As artificial intelligence increasingly permeates every aspect of our lives, we will be focusing on developing an advocacy agenda to make sure that the great opportunities heralded by AI are maximised, while the risks are minimised. Make sure you follow us on Twitter to keep up to date with our progress.


By Cath Elliston