Data-driven environment model generation using imitation learning for efficient cyber-physical system goal verification

Abstract

Cyber-Physical Systems (CPS) continuously interact with their physical environment through software controllers that observe the environment and determine actions. Engineers can verify to what extent the software controller under analysis can achieve given goals by analyzing its Field Operational Test (FOT) logs. However, repeating many FOTs to obtain statistically significant results is expensive in practice. Simulation-based verification is an efficient alternative for reducing the FOT cost for CPS goal verification. However, it requires an accurate virtual environment model that can replace the real environment interacting with the CPS, and it is challenging to craft the environment model manually. This dissertation proposes a novel data-driven approach that automatically generates the virtual environment model from a small amount of FOT logs. It generates an environment model that mimics the behavior of the real environment using Imitation Learning (IL). Specifically, this dissertation provides 1) a systematic and comprehensive survey on environment modeling, 2) a formal framework of CPS goal verification and a formal problem definition of environment model generation, 3) a data-driven environment model generation approach using IL, and 4) an empirical evaluation based on case studies of an autonomous driving system goal verification and reusable datasets. The evaluation results show that the approach can generate accurate virtual environment models for CPS goal verification with small FOT log data. Therefore, CPS software engineers can automatically obtain accurate virtual environment models and efficiently verify the controller based on the simulation.