Monday, May 6, 2024

Waze Hallucinations

]Kaiser Fung WAZE weighs in here.]

 

I suspect everyone reading this has had moments while driving where their phone has taken them off what seemed to be the quickest and most obvious route. The natural assumption at those times is that the navigation system knows something that we do not, such as a traffic slowdown just beyond the next hill.

For a long time, I thought that way too, until a few years ago when, while still looking for an apartment close to work, I found myself making the daily drive from South Central LA to an office building across from the Burbank Airport.

I was living just off of the 110 interstate, so the shortest and generally fastest route was to take it up to the five then take that to Burbank. Not only was that the most obvious route, it was generally what my phone would recommend when I first got in the car. Occasionally, though, particularly going through downtown, my phone would tell me to get off of the interstate and jump onto surface streets. This can be a smart move under certain circumstances, there has to be a substantial loss in highway speed to justify the detour onto surface streets when you have a more or less direct shot on the freeway.

Rather than fighting my way across lanes of traffic to make an exit under these circumstances when there appeared to be no slowdown in traffic ahead of me, I started ignoring these suggestions, and I discovered that in pretty much every case, there was no slowdown, my original time to arrival remained more or less accurate, and I almost certainly would have lost time doing what I was told.

Since then, I will start out a trip by checking the times on the different routes according to my phone or laptop and pick the one the algorithm recommends, but if I am familiar with the streets between me and my destination, I will usually ignore these last minute suggestions. 

I'm not saying the navigation systems are always wrong or even that simple and intuitive routes consistently beat them. Sometimes it turns out that my phone was right and I was wrong, that traffic that was moving fine suddenly snarls up and I probably would have been better off jumping on residential streets for a weird double dogleg.

What I am saying is the algorithms are sometimes clearly wrong, and, more importantly, wrong in an interesting way. It's not hard to imagine a system producing sub-optimal routes due to a lack of data or the need to reduce complexity and computation time, but those things should tend to produce very different mistakes than the ones we're seeing. In the cases I described, the algorithm is generating convoluted routes sending us to streets with little (and sometimes virtually no) real-time data, routes that would be difficult to justify without plugging major slowdowns into the algorithm, slowdowns we know don't exist. That's the part that raises the cool questions.

I know we have someone in the audience who has worked with these or some other related system. What's going on?

3 comments:

  1. Hi Mark - I'm glad someone tried to validate these route suggestions. I have long suspected they work less often than we assume them to. My impression is that many such algos are too myopic.
    I like the example of having to weave through traffic in order to exit the highway and use smaller roads. To model that, one has to not just look at the map but also consider the congestion level around your car on the path to the exit. I doubt that any real-time algo is that sophisticated.

    ReplyDelete
  2. Not to mention the unprotected lefts onto busy 4 lane LA streets.

    ReplyDelete
  3. Maybe Waze is trying to collect unobserved data. These diversions are streets with little real-time data. How can Waze get more coverage? Well, if the static route time for the unobserved route is nearly the same as your well-observed route, then perhaps a nudge will send you there to collect more data for your benevolent Overlord. Its not helping you, its helping itself and perhaps the future you.

    ReplyDelete