Google’s DeepMind has mastered another game
Every parent is secretly a little scared of the day when they no longer have to let their children win at a game – be it chess, Scrabble or even Hungry Hungry Hippos – but Google researchers have now had the bittersweet pleasure of seeing their DeepMind AI learn the ancient Chinese game of Go and promptly destroy some of the world’s finest players.
If you’re not familiar with Go, all you really need to know is that, while it looks pretty simple, it’s actually pretty complicated. So complicated that the world’s experts estimated we were still a decade away from an artificial world champion. As Google points out in
If you’re not familiar with Go, all you really need to know is that, while it looks pretty simple, it’s actually pretty complicated. So complicated that the world’s experts estimated we were still a decade away from an artificial world champion. As Google points out intheir blog on the victory, there are actually more potential positions on the Go board than there are atoms in the universe.
“The reason it was quicker than people expected was the pace of the innovation going on with the underlying algorithms and also how much more potential you can get by combining different algorithms together,” DeepMind’s chief executive Demis Hassabis told the BBC.
As you can see in the video above, DeepMind’s bespoke Go program, AlphaGo, beat the European champion Fan Hui 5-0. Ouch.
The delight doesn’t need to stop there for Google either. It also means they’ve bested Facebook’s AI to mastering Go:
Big deal, you may think: Deep Blue beat Kasparov at chess in 1997, so why is this still news? Well, two main reasons. For one, Go is a lot less logical than chess and with a lot more “intuition” at its heart. For another, it’s more complicated, with there being around ten times as many possible moves in your average Go position than in chess.
It required a different approach and that’s what Google did. “We built a system, AlphaGo, that combines an advanced tree search with deep neural networks,” writes Hassabis. “These neural networks take a description of the Go board as an input and process it through 12 different network layers containing millions of neuron-like connections. One neural network, the ‘policy network,’ selects the next move to play. The other neural network, the ‘value network,’ predicts the winner of the game.
“We trained the neural networks on 30 million moves from games played by human experts, until it could predict the human move 57% of the time (the previous record before AlphaGo was 44%). But our goal is to beat the best human players, not just mimic them. To do this, AlphaGo learned to discover new strategies for itself, by playing thousands of games between its neural networks, and adjusting the connections using a trial-and-error process known as reinforcement learning.”
Still, AlphaGo isn’t the world champion yet, even if it has conquered Europe. In March, it will face an even tougher match in Seoul against Lee Sedol, the best player in the world over the last ten years. I don’t fancy Sedol’s chances.
If all of this is mildly terrifying, take some comfort in this: we’re still a way off an artificial football player who would get picked for five-a-side ahead of me, and I’m terrible at football.
There’s something quietly reassuring about that.
Image: Makia Minich, used under Creative Commons