25 Feb 2015

Computers learning to play video games: why it matters

A computer programme that can play a computer game? So far, I hear you yawn, so exciting. But this latest discovery could be a major milestone in artificial intelligence research.

One of the holy grails of artificial intelligence (AI) research is “generic” intelligence that’s “non-trivial”. In other words, a computer that can solve a problem it’s never encountered before and isn’t programmed to solve. But also an architecture that’s written in such a way that it could be applied to many different types of problem. From planning the best route from A to B to solving the deepest mysteries of biology or physics.

What makes this new work so important is that the research team (from UK-based start-up DeepMind, recently acquired by Google) created a programme capable of playing a large number of different computer games, looking only at the information on the screen, without any prior programming on how to do so. Not a single byte.

For those of us of a certain age it has extra appeal, because the research team used some of the earliest computer games (remember the Atari 2600?) to demonstrate their system.

Games represent convenient “problems” for an artificial intelligence to solve. And the graphics of those made in the 1970s are simple enough to allow the programme to provide results without requiring too much computational power.

With just a few hundred “gos” the system could beat the best 13-year old at games like Pong and  Breakout. In fact, faced with 49 different Atari games, it was 75 pr cent better than a human professional games tester in more than half of them.

Key to its success, said the team behind it, was the combination of  “reinforcement learning” and “deep neural networks”in the system – both inspired by an understanding of how our brains organise learning and memory functions.

“This doesn’t imply that all you have to do is scale up this system to get human-level AI,” said Murray Shannahan, an AI researcher at Imperial College in London, who wasn’t involved in the research. “But it’s a step in that direction.​”

​The system wasn’t bad at the more unpredictable game Asteroids,  but other game classics like Pac Man and Montezuma’s revenge were more of a challenge. These games require strategy and forward planning beyond the reach of the system as it is currently developed.

The machines aren’t rising up just yet – but watch this space.

Follow @TomClarkeC4 on Twitter