研究開始時の研究の概要 |
The use of "black-box" components (e.g. AI modules) adds new, unprecedented difficulties to the verification of cyber-physical systems (CPSs). Nonetheless, their use in modern applications such as autonomous vehicles demands rigorous system testing and verification. Existing verification techniques do not scale well and rather focus on verifying individual components in isolation. This project describes aims to develop an efficient approach to lifting existing verification methods to the integration level and thereby coming one step closer to CPS verification at the full system level.
|
研究実績の概要 |
In the course of the project, I primarily investigated scenario generation for ADS. Specifically, I identified a major issue with noisy/non-deterministic simulators such as Autonomoose. To overcome, I developped kNN-Averaging, a search-based algorithm that improves current scenario heuristics. Furthermore, I investigated the complexity of scenario generation and managed to proposed a hierarchical scenario definition framework, formalising the this complexity. Next, together with colleagues, we participated in two editions of the SBST CPS challenge, where we developed and submitted Frenetic, one of the top-performing competitors. We are in preparation of a empirical study of scenario/road diversity and its impact on ADS behaviour diversity.
|