I live in an area strongly impacted by Google making ‘individually optimal’ decisions for each driver, and actually leaving those drivers in a dramatically worse situation.

I live in a rural area, between a major population center and a major resort area, with one major highway and a few small back roads that provide alternate paths for part of the highways route. Every summer weekend the highway becomes highly congested.

Google quickly starts routing people down the back roads because of a 30 minute delay on the highway. A sudden crush of cars hits these back roads, and they end up gridlocked for 3-4 hours. Google then realizes traffic is literally stopped on these roads, and stops sending new traffic down those routes. But the people already on them are still stuck for hours.

It gets smelly when a bunch of drivers take a shit on the side of the road because they can’t go anywhere else, and leave it there.

All because Google simultaneously made an ‘individually optimal’ decision for a whole bunch of individual drivers at once.

——

Another example in the same area actually causes a backup on the highway itself. Google started suggesting one back road that required an unprotected left turn across oncoming traffic on the highway, to avoid a 10-15 minute delay further on the highway. Drivers dutifully followed directions by getting into the left turn lane.

The drain rate of the left turn rate is slow because oncoming traffic is also high. The left turn lane fills up, and one driver with directions to turn then stops in the traffic lanes to wait for room to get into the turn lane. And suddenly the highway is now encountering 2-3 hour delays that don’t clear for most of the day.

In theory, the learning based approach in this blog post could solve this without any architectural changes. They mention predictions spilling to nearby roads, learning the pattern that "if the highway is backed up now, the little road will be too in a few minutes". Once that is baked into the model, Google Maps won't direct (as many) people onto the site road because they would get stuck in the predicted traffic there. The nice thing about this is it also handles the case where other mapping applications direct drivers there, and the "smartest" mapping application will still get you there fastest.

I wonder if this will be improved with more realistic cost functions that adequately model those localized effects.

Probably impossible at a global level but I wonder if it’s possible to eventually model and update those cost functions periodically to represent local maps

This happens sometimes on I80 in the Winter (usually btwn Auburn and Tahoe). Snow conditions can wreak havoc on weekends. People are routed to backroads --but not everyone has chains on (so they may have to fumble and get them on while on a road without wide shoulders to pull over), or clearance that will enable to take the back roads --so they get stuck in even bigger snarl.

It's evident they (G) need to work with Traffic Engineers and not cowboy it.

Surprisingly, it isn't always the case that each driver optimizing their own route will lead to a global optimum - see Braess's paradox. "If every driver takes the path that looks most favourable to them, the resultant running times need not be minimal." [0]

[0] https://en.wikipedia.org/wiki/Braess's_paradox

That's definitely not how it works. Traffic is a nonlinear dynamical system with all sorts of non-intuitive coupling between various components.

What's the difference between a dynamic system and a dynamical system?

A dynamical system (https://en.wikipedia.org/wiki/Dynamical_system) is basically a bunch of (continuous / discrete) variables whose evolution is governed by a set of (differential / difference) equations. E.g. a pendulum is a dynamical system whose state can be described with a single variable 𝜃 (the angle of the pendulum), governed by the differential equation:

𝘥²𝜃/𝘥t² + g/l sin(𝜃) = 0

(where g is the gravitational acceleration and l is the length of the pendulum)

Dynamical systems whose behaviour is linear are kinda well-behaving and easy to analyse. (linear in this context means that the system's response to a linear combination of inputs will be the same as the linear combination of its responses to the individual inputs) Non-linear systems on the other hand behave chaotically, producing wildly different responses to slight differences in their inputs.

I'm not sure what a "dynamic system" is though.

In some cases that’s exactly how it works. There’s a four way stop on a major commuter route where I live. Most of the traffic passes through it in a straight line and there is a major build up, like 20-30 min long during rush hour. In one direction there is a side road that loops around off the main road for about 2 miles in a big U and joins back up again at that 4-way stop. Before Waze became popular you could shoot around this bypass and be the only car on that side of the stop sign. Now there is a pretty significant amount of traffic there every rush hour thanks to the GPS apps. Lowell road and Barnes Hill road in Concord, MA for anyone curious.

I think that modelling is to simplistic. Travel time is often times dependent on traffic load, with a stark discontinuity reaching 100% (stop and go traffic).

Consider a contrived example of two similar bridges gapping some body of water. Bridge A is fed by a major highway, Bridge B is downstream a couple of miles and connected to the highway on both sides. For most drivers, if there are no other cars on the road, the preferred route is Bridge A.

The global optimum would be to divert some percentage of highway traffic to Bridge B, _before_ we saturate Bridge A. But for each individual on the highway this would mean a detour and would prefer someone else to take Bridge B.

That's how it theoretically works as a free service where all users are equal. It's not difficult to image a tiered subscription model which finds a sub-optimal 70/30 split more profitable. It's also important to note it's optimizing for time, not fuel usage (a shorter path may require expensive elevation changes, for instance), traffic noise for neighborhoods, safety, services access, etc.

This is the correct answer. Optimizing for the user is optimizing for the conglomerate. By giving each user the path of least resistance, the highest systemic throughput is achieved.

This is correct. The optimal routing that improves transit time for individual travelers would also minimize global transit time. However, it is true that there could be externalities to this behavior, like cars being routed through neighborhoods with higher likelihood of accidents.