Fragility-Oriented Testing with Model Execution and Reinforcement Learning

摘要

Self-healing is becoming an essential behavior of smart Cyber-Physical Systems (CPSs), which enables them to recover from faults by them-selves. Such behaviors make decisions autonomously at runtime and they often operate in an uncertain physical environment making testing even more challenging. To this end, we propose Fragility-Oriented Testing (FOT), which relies on model execution and reinforcement learning to cost-effectively test self-healing behaviors of CPSs in the presence of environmental uncertainty. We evaluated FOT's performance by comparing it with a Coverage-Oriented Testing (COT) algorithm. Evaluation results show that FOT significantly out-performed COT for testing nine self-healing behaviors implemented in three case studies. On average, FOT managed to find 80% more faults than COT and for cases when both FOT and COT found the same faults, FOT took on average 50% less time than COT.