Determining the Validity of Simulation Models for the Verification of Automated Driving Systems

Abstract

As the verification of automated driving systems poses an immense challenge, recent approaches aim for a virtualization of such efforts using computer simulations. This goal, however, motivates a strong need for trustworthy simulation environments and models. As to assess the modeling quality, this work proposes a process to measure the difference between the behaviors of several models. To achieve this, we consider sets of discretized simulation runs to be modeled by time-homogenous Markov chains and under this assumption derive a computable distance measure between sets of simulation traces. If it can be assured that all relevant variables may be observed and no crucial hidden factors are left out, the method can be extended to compare real-world traces with their simulated counterparts.

Keywords

• Markov chains
• maximum mean discrepancy
• simulation
• two-sample tests
• validity