From Section RQ1 and RQ2, we find that traditional and AI controllers, respectively, have their own advantages in satisfying different requirements. In this RQ, we consider combining different controllers in a single system to perform a modular redundancy, and we evaluate if this design can improve the quality and reliability of our systems.

We select 4 systems, namely ACC, AFC, WTK, and CSTR, to deploy the hybrid control schema, since the controllers in these systems show unique strengths on different evaluation metrics and in falsification. We design hybrid controllers with various combination methods which have been discussed in our paper.

1. Adaptive Cruise Control (ACC)

Although the traditional controller in ACC has the best performance on all metrics in RQ1, it still has a countable number of falsification rates in RQ2. It points that the traditional controller can stably operate under regular conditions, but when an emergency situation happens, it may fail to secure the safety baseline. However, the DRL controller with the DDPG agent has significantly better performance in falsification. In contrast, this DRL controller has poor performance on S2 MAE. Namely, the DRL controller takes more attention to maintain the safety distance and carefully control the acceleration of the ego car.

For the Scenario-dependent-based approach, we set a threshold on the relative distance. If d_rel is greater than the threshold, the control authority is given to the DRL controller which pays extra attention to maintain the safety distance. When d_rel is below the threshold, the traditional controller takes charge to minimize the average error.

All the hybrid controllers perform better in falsification than the traditional controllers, and the scenario-dependent hybrid controller performs well in MAE compared to the traditional controller. In contrast, the performances of the random-based or the average-based hybrid controllers are not as good as the scenario-dependent ones. Thus, the scenario-based controller takes the advantage of the constituent controllers and minimizes the drawbacks. Hybrid controllers constructed by the average-based method and random-based method do not effectively utilize the advantages from the traditional controller to minimize the average error.

2. Abstract Fuel Control (AFC)

We take the DDPG-based DRL controller to combine with the traditional controller in AFC. Since DDPG controller has unique strength to minimize the maximum error and the traditional controller is good at handling the average error. A threshold value is set on μ as the controller switch in the scenario-dependent-based approach.

For the Scenario-dependent-based approach, we set a threshold on the relative distance. If d_rel is greater than the threshold, the control authority is given to the DRL controller which pays extra attention to maintain the safety distance. When d_rel is below the threshold, the traditional controller takes charge to minimize the average error.

Among the 4 types of hybrid controllers we deployed, 3 of them have similar or better performance than the traditional one. From the falsification aspect, only the random-sampled controller HR_1 has been falsified. This indicates that a random control switch with a long sample interval can possibly take down the system performance.

3. Water Tank (WTK)

The hybrid controllers in WTK are slightly different from other systems, as WRK takes 2 DRL controllers to construct the hybrid controller. The results in RQ1 and RQ2 show the traditional PID controller in WTK acts poorly to meet the system requirements. It is not a good way to combine the traditional controller with others, since the PID controller most likely will drag down the overall performance. We also notice that, in DRL controllers, the DDPG controller has advantages on regulating maximum error and the TD3 controller is good to maintain the steady state.

It is not a good way to combine the traditional controller with others, since the PID controller most likely will drag down the overall performance. We also notice that, in DRL controllers, the DDPG controller has advantages in regulating maximum error and the TD3 controller is good to maintain the steady state. The scenario-dependent and the average-based hybrid controllers are significantly better than the random-based controllers, and none of the hybrid controllers have been falsified.

4. Exothermic Chemical Reactor (CSTR)

The hybrid controllers in CSTR have a traditional PID controller and a TD3-based DRL controller. Although the PID controller has terrible results on falsification, it behaves well in reducing the average error. On the other hand, TD3-based DRL controller has not been falsified and gets the best score on maximum error reduction. The condenser pressure deviation is used to switch between two controllers for the scenario-based hybrid controller.

We find that the random-based method controller HR_0.1 and the average-based controller are not falsified. Moreover, the average-based controller also performs well in MAXERR as its constituent TD3 controller does. The scenario-based controller does not have special characteristics among evaluations, this could be a result of inappropriate switching logic.