Sub Use Case UUV. 1: Scenario Generation/Optimization Usually, for testing vehicles on a testbed, a test plan has to be created to reflect the needs of the function/component under test. The resulting test data gathered often contains many datapoints with little new information. The UUV.1 use case aims to improve data quality by adapting the test plan in a way that every tested scenario contains as much information as possible. To achieve this higher information density, certain elements of the scenario, like the time of day, placement of cars on the roadside or density of trees are varied automatically to create a scenario that fulfils the requirements. The two main ingredients for evaluating the data are novelty and criticality. Ideally, the gathered data should be as novel and therefore as dissimilar to already seen data as possible, while also containing a traffic situation that is quantified as critical, by a combination of safety Key-Performance-Indicators (KPIs). The optimization will be done using a Reinforcement Learning (RL) agent that proposes modifications to a basic scenario as actions and receives an estimate of the novelty value as well as the criticality of the modified scenario as a reward. An adaptation to also include comfort KPIs regarding the driving function besides criticality KPIs is possible. The basic structure of this use case can be seen in Figure 16. Figure 16 Simplified overview of the UUV.1 use case, that features a Reinforcement Learning-based Automatic Scenario Variation to find critical versions of a Scenario As the character of the state and action vector, the transfer of learned techniques to propose the wanted scenarios, and the amount of generated data highly influence the ideal RL algorithm to solve this optimization problem, T3.3 shall be used to develop concepts which fit the use case. To further elaborate on the use case, the following section will expand the view on the way the RL agent shall interact with its environment.
