London Cab Drivers Outsmart AI in Complex Route Planning

New research unveils how expert London taxi drivers rely on spatial awareness and memory to navigate the city's intricate streets, showcasing a sophisticated cognitive strategy that AI systems have yet to match.

​​​​​​​Research: Expert navigators deploy rational complexity–based decision precaching for large-scale real-world planning. ​​​​​​​Image Credit: mikecphoto / Shutterstock

An international team of scientists, including a neuroscience principal investigator from the Champalimaud Foundation in Lisbon, has analyzed how expert London taxi drivers plan their routes whenever they pick up passengers. This is the first study about human planning to be performed in a real-world, large-scale environment. Their results are published today (January 23rd) in the Proceedings of the National Academy of Sciences. 

London's licensed taxi drivers are not unknown to research: some 25 years ago, they were found to have a distinctive brain feature due to their navigation expertise. Indeed, the volume of their hippocampus, a structure in the brain that plays a vital role in learning and memory, is different from that of people who do not drive taxis.

The new study focused on their planning behavior when doing their job. "Expert navigators can plan routes efficiently and quickly in enormous and intricate environments, such as cities," the authors write in their paper. "[But] most studies to date have employed small-scale and/or abstract environments with naive participants. Here, we surmount these challenges by asking London taxi drivers – famous for their expert knowledge of the London street network, composed of over 26,000 streets – to plan routes through London."

"Cities epitomize the intricacy of planning in real-world environments," says Pablo Fernandez Velasco, from the University of York, the joint first author of the study. "If we considered planning a path across a large city that, between start and destination, involved 30 streets, using a minimal tree-search algorithm would involve an evaluation of over one billion potential street sequences." If anything, this makes it clear that humans do not use such an algorithm to choose their route.

Time to think, time to name streets

The researchers analyzed the drivers' response times – "as a proxy of thinking times" – to see how these drivers rationally organized their route-planning process. 

"Our goal was to understand how expert humans make plans in very complex situations," says Daniel McNamee, Group Leader of the Natural Intelligence Lab at the Champalimaud Foundation and one of the study's senior authors. "We also wanted to compare expert humans versus AI algorithms. We hypothesized that humans were somehow doing something different. AI algorithms couldn't explain the behavior expert humans exhibit in this sort of situation."

London taxi drivers are a unique population of humans, in particular, because when they train for their license, they are not allowed to use tools such as Google Maps to help them; they have to rely solely on their memory storage and retrieval skills to navigate tens of thousands of city streets efficiently. 

A figurative depiction of the mental representation of London in the minds of expert navigators. The height of each yellow "building" depicts the thinking time of taxi drivers at that location.​​​​​​​

"They are a unique group when you want to answer questions such as how do expert humans organize their memory, what internal representations do they use or how do they retrieve pieces of information from their memory", McNamee explains. "And that gives us a unique window into the highest level of human decision-making, knowledge representation and processing."

A critical difficulty was getting access to taxi drivers. Hugo Spiers, Professor of Cognitive Neuroscience at University College London (UCL) – the other senior co-author of the study – has spent 20 years studying their brains, but this is the examination of their planning. "These taxi drivers are famous for their larger posterior hippocampus," says Spiers. "Now, thanks to collaboration across UCL and the Champalimaud, we have revealed how they put that enhanced brain to use."

The team gave a group of 43 taxi drivers a start location in London and a "goal" location. The drivers were then asked to verbally call out the sequence of turns and streets they would take to go from the route's origin to the destination. "This is a very familiar task for these expert taxi drivers because, during their specialized training for the job, they're tested on many of these start-goal combinations before they get their taxi driver license," McNamee points out.

For each route (there were 315 such possible "runs," as they say in taxi driver-speak), the researchers measured the response time between the task presentation and each driver's first call-out of a street name. This, they called the "offline thinking time," corresponds to the phase during which the taxi driver silently considers the route. Next, they measured the times between each call-out of street or turns by the drivers, which they called the "online phase". 

Organizing decisions about a route

One of the things the team wanted to study was what decisions the drivers made about their route during their offline thinking time. "One possibility," says McNamee, "was that they organize their decisions in the sequence of streets and junctions that they are going to physically encounter during the ride – that the first thing they think about is the first street they're going to take, then the second street, the third street, and so on and so forth." This means they will be slower in naming subsequent streets as they go along the route.

The results contradicted this hypothesis. The results suggest that "during the offline thinking time, the taxi drivers prioritize high-complexity, non-local parts of the London space," McNamee continues. "It seems that they immediately jump to figuring out a higher-order global structure [of the route], based on critical junction states, using the physical structure of London." The technical expression for this is "non-local precaching of critical junctions." And then, "once they've done that, they start verbalizing and filling in the rest of the details."

The drivers also took longer to consider the longer streets in London. "This suggests they have spatial awareness of the actual physical structure of London, as opposed to a more abstract kind of representation where distances are not taken into account," McNamee points out. 

"I was surprised by this result," he says. "I expected the taxi drivers would have a less spatially embodied sense of the London state space. I did not expect that the length of the street in the real world would slow them down but that they would have a more abstract notion [of their city]. But maybe having this more embodied spatial awareness is useful for more flexible planning, where they can consider if one street is usually more busy than another at a given time of day. They're also encoding all sorts of additional spatial variables and integrating them into their thinking. However, this is not something we specifically addressed in our study."

"This result is consistent with reports from previous research," says Pablo Fernandez Velasco. "Taxi drivers have reported that when they have to plan a route, they sometimes imagine going through it, both from an aerial and from a first-person perspective. If they imagine the physical space to some degree, it makes sense that the length of the street segment impacts the length of their street planning".

A different order of priorities in "London space"

In sum, expert London taxi drivers prioritize important streets and junctions to determine a more global structure of the run. "What we see with the taxi drivers is that they consider space out of the order in which they would actually experience it in the real world. ".

"If they first take time offline to think about it differently, then they'll be much faster to call out streets during the online phase," adds McNamee. This is clearly a much more efficient way of thinking than a simple sequential strategy. 

What does McNamee think is the most interesting result of this work? "I think it is the fact that we were able to identify and predict that, during the offline thinking time, when the taxi drivers are in deep silent thought, they prioritize highly complex, spatially remote junctions and streets in London," he answers. "That had never been shown in human experts – decision-makers or planners – in what I would refer to as an ecologically valid task for them, that is, a task they accomplish every day." 

"In fact, what we were studying here is a kind of expert human intuition, an innate sense of what to think about before you even start thinking about it, in a sense," he ponders. "That's a key area of interest in other work we do in my lab: intuitive human reasoning."

According to McNamee, one thing seems inevitable: human experts have a qualitatively different algorithmic strategy than AI systems. Nonetheless, AI algorithms could benefit in the future from results about how human experts think. "They could benefit in, I'd say, two ways," he explains. One has to do with more flexible planning, mainly when there are a lot of environmental features and dynamics that must be considered." 

"The other would be to integrate the information about human experts into AI algorithms designed to collaborate with humans. This is a very important point, because if we want to optimize how an AI algorithm interacts with a human, the algorithm has to 'know' how the human thinks."

"And that's something to consider," McNamee concludes.