3 ms·
The number that people drive isn't important, what's important is the accident rate: do humans get into enough minor accidents over 1.7m vehicle-miles that 14 i
by gwern 12y ago
The number that people drive isn't important, what's important is the accident rate: do humans get into enough minor accidents over 1.7m vehicle-miles that 14 is clearly bigger or smaller? Your answer is going to be driven almost entirely by what human-rate you select.
For example, if I use https://www.dot.ny.gov/divisions/operating/osss/highway/accident-rates https://www.dot.ny.gov/divisions/operating/osss/highway/acci... , then I see that most of the quoted accident-rates are 1.8-3.29 per million miles. If I use a rate of 2.2/1m and use the simple binomial approach of x events per million miles, then the self-driving cars are statistically-significantly higher accident rates (since 14>3.74, and `binom.test(14, 1700000, p=(2.2/1000000)) ~> _p_=3.778543e-05`).
But what kinds of accidents are those? If I instead look at https://www.dot.ny.gov/divisions/operating/osss/highway/accident-costs https://www.dot.ny.gov/divisions/operating/osss/highway/acci... and look at "Property Damage Only" accident-rates, it seems the human rate per million mile is more like 65-85. Now, unsurprisingly, the robots look superior (14<110, `binom.test(14, 1700000, p=(65/1000000)) ~> _p_=2.2204e-16`).
And if you try to argue apples and apples and use fatal accidents from the same source, noting a range of 0-1.33 fatal accidents per million miles, and 0 fatalities due to the robots' 1.7m so far then the difference is unclear (0<0.85 but `binom.test(0, 1700000, p=(0.5/1000000)) ~> _p_=1`) because the usual tests don't like 0-counts and the robots have not driven enough to clearly distinguish themselves from normal human rates.