Popis: |
Simulation-based methods promise to be part of the solution to the challenge of validating highly automated driving functions (ADF). However, the application of simulations introduces a variety of problems itself. One of them is efficiency: When performing statistical simulations, the simulated scenarios are usually drawn by a distribution which represents reality. Only a very small share of these scenarios will be relevant/challenging since a good performing ADF can handle most parts of reality. Hence, a large number of simulation runs will be required to collect enough information for stable statistical validation. Importance sampling (IS) tackles this by drawing the scenarios from a distribution which is skewed onto the critical parts of reality. In order to draw these critical scenarios, IS requires scene models which allow selecting scenes of varying criticality. This work proposes an adequate scene model which is based on sum-product networks. It proves experimentally, that this scene model can be used with IS. |