Discover more from Apperceptive
The story Waze tells about problems with autonomous cars
The problems caused in cities by ubiquitous GPS navigation are really problems caused by automating part of driving
I’m not alone in complaining about ubiquitous GPS navigation. Google Maps and Waze (in particular) are causing changes in traffic patterns that have unpredictable knock-on effects. For drivers, this has the effect of helping find shorter travel times (sorta, mostly; this can be importantly false but that’s a whole other topic). But for urban residents and transportation planners, it's been a mixed blessing at best and is often frustrating, dangerous, or infuriating. The ways in which GPS navigation interacts with traffic networks are problematic, difficult to solve, and provide some interesting lessons about what happens when algorithmic decision-making interacts with human-designed systems, lessons which are applicable to the question of what is likely to happen with broad deployment of autonomous cars.
The big innovation of Google Maps and Waze is to use real-time traffic information to do routing. Every car using one of these apps is reporting its position and speed back to the mothership. So if there is traffic congestion, Google Maps and Waze know that in real time, and can route around it. In a city, the most common reason for congestion (that is, a bunch of cars in a line not moving) is a traffic signal. So Google Maps and Waze treat traffic signals as problems to be routed around. From the perspective of a single vehicle trying to optimize the travel time from point A to point B, this makes sense. Sitting at a light slows you down; if there's an adjacent street with no lights, all else being equal, you'll get through that segment of your trip faster. So far, so obvious.
The issue becomes apparent when you consider why there are traffic signals and why they are where they are. The purpose of traffic signals is to provide order to intersections that see more traffic than can safely interact without explicit turn-taking. If you have a quiet four way intersection that sees very little traffic, stop signs are perfectly sufficient. If you have a minor road intersecting with a major road, maybe the minor road gets stop signs and the major road gets nothing. If you have an intersection between two roads that each carry a lot of traffic, then you need a traffic light: the vicissitudes of social interactions at a stop sign would lead to backups or accidents and that intersection would become a source of traffic snarls. So as you’re driving, if you pick a route that has fewer traffic signals (and thus, as Google Maps and Waze can see, less congestion), you will most likely be driving on roads that were designed to handle less traffic.
If the route-picker is an individual driver who happens to know a back route along roads with fewer traffic signals, there isn't any obvious systemic problem. That driver might go too fast, since they're using a road that was primarily designed for local access as a through-route, but most of the traffic on that road will behave as the planners who designed the roads expected, and the road design will be safe and suited to those vehicles.
The problem comes when a meaningful percentage of the traffic from the main route gets diverted to the alternate route. Now you have roads that were not designed for through traffic carrying trunk road levels of traffic. Roads which lack signalized intersections, pedestrian accommodations, bike lanes, and visibility are suddenly being predominantly used for high volume through-traffic. This traffic tends to be faster, because the drivers are optimizing to get past this part of their trip as quickly as possible. When this happens, those roads become much more dangerous for local users, including residents, pedestrians, and cyclists.
I see this all the time in my neighborhood. There's a street that intersects with the main road through town. It's one block away from, and parallel to, a larger cross street. The intersection of the larger cross street and the main road has a light, and the larger cross street is one way and relatively wide. The larger cross street has a separated bike lane and good visibility at the corners. The smaller cross street is much narrower, enough so that two cars can't really pass each other in the areas where there are parked cars. It's also two way. The turn from the main drag onto the smaller cross street has no traffic signal; cars entering it from one direction are taking a left turn across relatively fast traffic and a crosswalk. I cross that crosswalk most days when walking my daughters, and it is a far more worrisome crossing than the larger intersection down the road. Cars making the left are trying to go quickly, to beat oncoming traffic. They aren't looking at the crosswalk, and even if they were, visibility of the ends of the crosswalk is obscured by parked cars. An extremely straightforward bit of road design—a crosswalk across a minor side road next to a main road—has become a dangerous and under-controlled infrastructure failure.
This has caused some notable problems across the US. In Los Angeles, residents of one incredibly steep and narrow street were at their wits’ end trying to get Google to stop sending oblivious commuters flying down their 35% grade hill.
At its core, these problems are about a mismatch. Urban planners are trained to design city streets for navigation by people using road signs and maps. The models of traffic flow they use are models of minimally-aided human navigation. If your navigation technology is restricted to road signs and maps, a few things are true. First of all, you're going to tend to use main roads. Main roads have more signage. Main roads are clearly designed to be through-routes. Main roads—if you look at a map—will be the obvious choice. Side roads bring more uncertainty. If you're looking at a map and trying to plot a route yourself, the question of whether a side road might be faster is very difficult to answer. The road could be blocked, or difficult to navigate, or an unmarked one-way street. It could simply be a slow and difficult road to traverse. Over time, as you learn an area, you learn its side roads. You navigate based on your experience and history with trying to follow these roads. You make rough estimates of what route seems faster. Sometimes you try a route but avoid it in the future because the road turns out to be a small residential street where you don't feel comfortable going as fast as you might elsewhere. That's the human navigational baseline that cities are designed around.
What Google Maps and Waze have done is completely change the nature of the agent that is interacting with the street grid. Instead of a human with some level of local knowledge, maps, and the ability to read signs, you have an agent with perfect knowledge of how long it takes to traverse every street right now, no need for signage, and no intuitive sense that big roads should be privileged over small roads. That's a very different kind of agent from our human baseline. Once a significant number of road users are that type of not-totally-human agent rather than the other one, the human baseline, the planning underlying the road network breaks down, and you have endemic infrastructure failures.
This sort of mismatch between the expected (human) agent and the actual (machine) agent is what can be expected with autonomous cars. In a sense, GPS navigation is the first point of interface between systems that automate part of the driving task and human-designed road networks. But really everything about roads, from when to make a signalized intersection to how to decide whether to make a left turn to insurance liability, is designed around humans being the operators of vehicles. When you change the nature of the agent – the operator and decision-maker acting in the system – you reveal weak points in the system that were not apparent before. You introduce failure modes that had not been considered, because they're outside of the space of what was possible when the system was designed.
The solutions to the problems introduced by GPS navigation are mostly non-obvious and non-trivial, and I won’t go into them here, with one exception. As with many if not all of the problems related to an auto-centric transportation network, there is a technically trivial but politically complicated solution available: reduce the number of cars on the road by making driving harder and other modes of transport easier. The subtle technical issues I described above, like a universe of other related subtle technical issues around automobiles, AI, and autonomy, are really just problems with cars-as-daily-urban-transportation made manifest.