Thursday, July 5, 2018

Autonomous cars and robots beating clothes on rocks

Here's yet another article on the problems with self-driving cars:

This article focuses on the problem of relying on using machine-learning to generalize from a corpus of real-world data to figure out how to build safe autonomous cars.  Oh, I have an example. When I was teaching my 16-year-old daughter to drive, we were on a standard 4-lane road and I told her to slow down because we were approaching a shopping district. She asked why, when out of the blue she just passed a man in a mouth-propelled wheelchair going the wrong way between the parked cars and her car. And I seized the teachable moment to explain that slowing down gives you more time to anticipate the unexpected.

Machine-learning algorithms are crap when it comes to anticipating the unexpected. We humans excel at that, as slow as we are.

I was lucky to have studied computer science at the university that had hired David Parnas shortly after he resigned from Reagan's Star Wars panel. Word is that Parnas agreed to join the panel so he could publicly denounce it on his first day, claiming that since it would be impossible to test, it would be guaranteed to fail. That was the rumor. What I know for sure is that at one department colloquium he explained one problem with AI (at least as it was done in the late '80s) is that practitioners were focused on automating existing practices rather than looking to shift the paradigm. The example he gave is that if an AI researcher had designed the washing machine, we'd have a machine with a robot arm that would move clothes up and down a washboard.

Self-driving cars remind me of that washboard robot. People act like the problem they solve is that drivers are fallible, and would rather spend their time watching videos or working on their devices than pay attention to the road. One solution to that problem is to have robots drive the car, but it's not a good one. We've already seen that these robots will kill people in ways that no awake, sober human will (and those other humans aren't legally allowed to drive, so let's ignore them).

If you rise up one level, you see that this is just a system-level problem: how can we move people in large around a city, safely and cheaply? And if you go to most cities in northern Europe or Japan, you can see that problem was solved long ago, through ubiquitous, efficient public transportation combined with bicycle routes in some of those cities. Sure, those cities were designed around streetcars and peddlers pushing carts, not free-ranging cars. So North American cities, especially the ones that were built in the 50s and 60s assuming $2500 cars and 25 cent gallons of gas, need some adjusting, but it's happened in core Vancouver, and other cities are jumping on that bandwagon. And improving transportation and bicycle networks should be much cheaper than getting self-driving cars out of the lab and into widespread use.

So think of the self-driving car as the one-armed washing machine: it's a start, but we can do much better.

No comments: