So How Likely Is a Car Crash? Estimating Rare Failure Events in Simulated Scenarios

Abstract: Scenario-based Simulated Testing is a common paradigm for testing complex cyber-physical systems like autonomous vehicles. This is particularly true for systems containing black-box ML components, which make formal verification or traditional engineering analysis difficult. However, answering the simple question of "how likely is it that the system will fail to meet its specification?" can actually be a significant challenge, especially when the true failure probability is rare.

In this talk, I will cover some of my recent work on failure estimation. It will touch on techniques such as adaptive importance sampling, multi-level splitting, and the use of surrogates in experiment design.