Abstract :
[en] Improving road safety is a major concern for most car manufacturers. In recent years, the development of Advanced Driver Assistance Systems (ADAS) has subsequently seen a tremendous boost. The development of such systems requires complex testing to ensure vehicle’s safety and reliability. Performing road tests tends to be dangerous, time-consuming, and costly. Hence, a large part of testing for ADAS has to be carried out using physics-based simulation platforms, which are able to emulate a wide range of virtual traffic scenarios and road environments. The main difficulties with simulation-based testing of ADAS are: (1) the test input space is large and multidimensional, (2) simulation platforms provide no guidance to engineers as to which scenarios should be selected for testing, and hence, simulation is limited to a small number of scenarios hand-picked by engineers, and (3) test executions are computationally expensive because they often involve executing high-fidelity mathematical models capturing continuous dynamic behaviors of vehicles and their environment.
The complexity of testing ADAS is further exacerbated when many ADAS are employed together in a self-driving system. In particular, when self-driving systems include many ADAS (i.e., features), they tend to interact and impact one another’s behavior in an unknown way and may lead to conflicting situations. The main challenge here is to detect and manage feature interactions, in particular, those that violate system safety requirements, hence leading to critical failures. In practice, once feature interaction failures are detected, engineers need to devise resolution strategies to resolve potential conflicts between features. Developing resolution strategies is a complex task and despite the extensive domain expertise, these resolution strategies can be erroneous and are too complex to be manually repaired. In this dissertation, in addition to testing individual ADAS, we focus on testing self-driving systems that include several ADAS.
In this dissertation, we propose a set of approaches based on meta-heuristic search and machine learning techniques to automate ADAS testing and to repair feature interaction failures in self-driving systems. The work presented in this dissertation is motivated by ADAS testing needs at IEE, a world-leading part supplier to the automotive industry. In this dissertation, we focus on the problem of design time testing of ADAS in a simulated environment, relying on Simulink models.
The main research contributions in this dissertation are:
- A testing approach for ADAS that combines multi-objective search with surrogate models to guide testing towards the most critical behaviors of ADAS, and to explore a larger part of the input search space with less computational resources.
- An automated testing algorithm that builds on learnable evolution models and uses classification decision trees to guide the generation of new test scenarios within complex and multidimensional input spaces and help engineers interpret test results.
- An automated technique that detects feature interaction failures in the context of self-driving systems based on analyzing executable function models typically developed to specify system behaviors at early development stages.
- An automated technique that uses a new many-objective search algorithm to localize and repair errors in the feature interaction resolution rules for self-driving systems.
