RQ2: Monitoring

Table 5 shows the results of the analysis of online safety monitoring, where three key observations are identified. First, Mosaic is able to improve the safety of the system while keeping a similar functional performance. It can be observed that, in 5 experiments, both the safety and performance are increased. Second, for ACC with DNN1 and DNN2, we notice that although the safety is indeed increased with the proposed online safety monitoring, the performance decreases. The safety and performance metrics of the traditional controller, DNN1, and DNN2 are 0.9744, 0.5109, 0.5311 and 0.6475, 0.9101, 0.8157, respectively. In other words, the AI-based controllers lean towards the performance metrics while failing at the safety of the system, while the traditional controller possesses an opposite control logic. One reason for this could be that the used safety query is more prone to safety metrics. Thus when switching to the traditional controller for safety, the performance of the system inevitably decreases. Third, for AFC with DNN1 and DNN2, Mosaic performs worse than AI controllers. The main reason could be that traditional controllers have worse safety and performance compared to AI controllers. The safety and performance with traditional PI and feedforward controllers are 0.7177 and 0.6143, which are lower than AFC with DNN1 and DNN2. In the case of a worse traditional controller, the switching control strategy could not help to improve the safety of the system.