How Artificial Intelligence Playing Atari Could Help Self-Driving Cars

Image for article titled How Artificial Intelligence Playing Atari Could Help Self-Driving Cars

When Google announced its new artificial intelligence system could play, learn, and beat some old-school Atari games, my first reaction was, "Cool. Call me when it can play Spy Hunter." Actually, what about Gran Turismo? Because that's obviously part of the plan.


Google dropped $648 million on London-based AI company Deep Mind last year, and for that kind of scratch we expected more than a computer that could play 22 Atari 2600 games better than any human. If Watson can beat us at Jeopardy and Deep Blue can kick Kasparov's ass at chess, what's the big deal with a computer mastering Space Invaders?

Because the crew at Deep Mind didn't teach the AI the rules. It just learned.

In a paper published in Nature, Deep Mind expanded on similar research carried out in 2013, tasking the AI with playing seven times more games than it did two years ago. The results were mixed, but the goals are clear.

"Ultimately the idea is that if this algorithm can race a car in a racing game then also essentially with a few extra tweaks it should be able to drive a real car," said Demis Hassabis, one of the paper's authors, during a press conference.

The biggest problem for Google's self-driving cars – beyond just basic issues like dealing with weather – is the need for detailed mapping. Currently, the cars can't travel anywhere that Goggle hasn't scanned using its range of LIDAR and sensor arrays. That's where the AI would come in, learning different environments on-the-fly to handle everything from construction zones to idiotic pedestrians. But it's still a long ways out.


The AI in this latest paper was only able to beat less than half of the games researchers had it play because it's stuck in the here and now. According to MIT's Technology Review:

The classic game Ms. Pac-Man neatly illustrates the software's greatest limitation: it is unable to make plans as far as even a few seconds ahead. That prevents the system from figuring out how to get across a maze safely to eat the final pellets and complete a level. It is also unable to learn that eating certain magic pellets allows you to eat the ghosts that you must otherwise avoid at all costs.


Deep Mind is out to fix this by boosting the software and hardware to increase its memory span in both directions – in the past and towards the future – and after it works through the 2D worlds of Atari, the next step is modern 3D games.

But Hassabis admits that still at least five years away, and using that technology to drive a car is, "again, even further away then that."


Contact the author at
Public PGP key
PGP fingerprint: 7301 D7FC 2FF6 D437 E5A7 0568 3A14 624A 1800 4C85


Ash78, voting early and often

The ultimate test will be how it does in a Testarossa. Will it hit a billboard or palm tree, killing the driver and his yellow-haired girlfriend?