Abstract:
|
The imminent launch of highly automated vehicles (HAVs) into the general driving environment calls for close attention to analysis of their safety performance. A report by Kalra & Paddock (2016) evaluated the miles required to reach p=0.05 significance using a one-sample Poisson compared to a fixed population rate based on national crash datasets. Our analysis explored the effect of making different underlying assumptions and using alternative methods to evaluate HAV safety. First, we establish a crash rate that is based on likely initial HAV release in a lower-speed, rideshare environment and is represented as Poisson-gamma, which has greater variance than the Poisson. Using this rate, we explore the efficiency of Bayesian one-sample comparisons and Extreme Value Analysis, the latter using a measure of forward proximity where a value of zero represents a crash. Finally, we propose some alternative ways to think about this problem more holistically.
|