This booth provides technical details of the scenario exploration procedures presented in the talk "Methods for Parameter Space Exploration" in Session 4 of the event. The procedures address different aspects of an important problem encountered in applying simulation in the development of a highly automated vehicle: There is a huge number of scenarios that might be relevant, so there might be many simulations to perform. For instance, by varying the behavior parameters of even a small number of traffic participants, billions of concrete instances of a (logical) scenario are generated.
The first procedure serves to find "interesting" scenarios fast. Based on an evaluation of the concrete instances which have already been simulated, it extrapolates the results to the full parameter space. The extrapolation guides the choice of further concrete scenarios for simulation. Technically, this is a Bayes optimization procedure which employs Gaussian processes to model the transition from prior to posterior estimations of the evaluation function on the parameter space. This procedure has been implemented and successfully applied to a set of logical scenarios.
The second procedure is concerned with guarantees for finding all interesting scenario instances. This is important e.g., for assuring that no critical situation is overlooked. The basic idea is to use continuity properties to argue that some concrete, simulated instance is representative for a local neighborhood, in that criticality does not change significantly in that neighborhood. Covering critical (and uncritical) regions by such representative concrete scenario instances provides the desired guarantee that nothing of importance has been overlooked. This procedure has been defined conceptually and is currently under implementation.
