Google's AI DeepMind can now travel the London Underground

System taught itself how to get around

Google's artificial intelligence (AI) system DeepMind has learned how to navigate the London Underground.

The keyword here is 'learned'. The system was not taught the thousands of stations and routes, but was given an external memory of them and left to learn by itself. This marks a major milestone in the capabilities of AI.

DeepMind is the same intelligence that has recently thrashed grand masters at fiendishly complex board game Go. This is a stunning achievement, but the majority of the decision making was based on logic alone.

The Tube task means that DeepMind has to adopt a strategy from a context, such as the quickest route, the one with the least walking or the one with fewest delays.

The secret lies in the external memory that allows DeepMind to tuck away tasty tidbits of knowledge to draw on later, much as a human does, at least theoretically.

However, Alex Graves from the Google DeepMind team in London, has made it clear that he doesn't see this as a quantum leap.

"I'm wary of saying we now have a machine that can reason. We have something that has an improved memory - a different kind of memory that we believe is a necessary component of reasoning. It's hard to draw a line in the sand," he told The Guardian.

A study in the journal Nature showed that DeepMind can work out the quickest route between two underground stops. But it can also provide contextual information such as where you would be if you went two stations north of Victoria. This relies on understanding the physics of the Underground, rather than looking for two stations due north as the crow flies.

DeepMind was recently taught the concepts of curiosity and reward to finish classic video game Montezuma's Revenge.