Application of Evolutionary Algorithms to Analyze Criticality in Urban Traffic Scenarios

Abstract

As automated driving system equipped vehicles (AVs) join traditional vehicular traffic, a safety analysis of their programmed behavior is essential. This holds for a variety of domains including highways, rural roads, and particularly, urban traffic. In contrast to other traffic domains, AVs in urban traffic settings must interact with vulnerable road users (VRUs) such as bicyclists and pedestrians. In order to ensure that AVs do not endanger VRUs, it is crucial to identify all potential traffic situations in which criticality between AVs and VRUs could arise. This masters thesis examines how the application of evolutionary algorithms (EAs) can identify critical situations in urban traffic scenarios. To this end, three separate urban traffic scenarios are designed. Each scenario models different criticality phenomena and has a different complexity. In order to enable high-frequency repetitions of the scenarios without endangering real humans, the scenarios are implemented in the urban traffic simulator Carla. The aim of the EAs is to learn how to derive concrete scenarios that yield a high criticality from the three designed scenarios. Hence, as a fitness function of the EAs, a criticality metric that quantifies the occurring criticality in a concrete scenario is used. Specifically, a new criticality metric called the Predictive Conflict Index (PCI) is proposed and validated in this thesis. With the use of this fitness function, six different variants of EAs are implemented in Python. These are the basic mu+lambda and mu-lambda algorithms, mu+lambda in combination with Rechenberg, mu-lambda in combination with Rechnberg, mu+lambda in combination with self adaptation and mu-lambda in combination with self adaptation. The two basic variants and mu+lambda in combination with self adaptation are able to generate a large amount of critical concrete scenarios. Moreover, the learning process of the algorithms identify different properties that are linked to high criticality, such as high target speed of AVs or close distance of parked vehicles to an intersection. Furthermore, utilizing methods from data analysis such as clustering, different types of critical situations in urban traffic are revealed. Ultimately, the examined approach is highly promising as a method to improve the development of AVs.