Transfer Importance Sampling - How Testing Automated Vehicles in Multiple Test Setups Helps With the Bias-Variance Tradeoff

Diverse test setups are used to ensure the safe operation of automated vehicles (AV). Field tests fully represent the complexity of AVs and their environment but are costly. Conversely, purely virtual tests are scalable but may not reflect all real-world phenomena sufficiently well. To reliably estimate risks despite the deficiencies of individual test setups, this paper proposes transfer importance sampling (TIS), a risk estimation method linking two test setups. TIS uses a scalable, cost-effective test setup to comprehensively explore an AV's behavior. The insights gained are then used to parameterize tests in a more trustworthy test setup accurately reflecting risks. Our experiments show that given a sufficient correlation between the test setups' results, TIS can reduce the effort of risk estimation compared to methods using one test setup. Contrarily, discrepancies may prevent efficiency gains but reveal deficiencies of the AV and test setups. Thus, TIS allows to compensate for and mitigate test setups' deficiencies, enabling more efficient and accurate risk estimation for safer AVs.