Summary

In this section, we discuss the challenges in our experiments and opportunities for future work.

State space reduction. In order to narrow the behaviour space and transparentize the dynamics of a system, we utilize a semantics-oriented abstraction to describe and predict the system semantics regarding major specifications. Based on the results from five experiment systems, our abstraction method can significantly reduce the size of the state, action and transition space over 99%, and restrict the semantics errors within an acceptable range. We notice that the reduction level of a system is not determined by the dimensions of the state space but by the diversity of its behaviour in terms of the specifications. For instance, BBC has a much larger state space than LKA, but its abstract state space is conversely smaller. This finding fits the practical situation that many subsets in the state space can have diverse and distinct values but represent a similar level of satisfaction regarding the system specifications. Therefore, abstracting the system based on concrete state values can be expensive and intricate; but the corresponding abstracted model may not effectively scale down the system and accurately conclude the system dynamics. Thus, our semantics-based abstraction tackles this problem from the semantics aspect in order to effectively and precisely decompose complex CPSs.

Guidance from semantics predictions. We investigate whether the semantics can provide guidance to empower the controller fusion strategies within the ensemble framework. To do that, we utilize the abstract model to predict the incoming semantics and monitor the possible outcomes of choosing different transitions. In this way, the ensemble method can avoid dangers in advance and drive the system toward optimal performance. Table 5 demonstrates the ability of the semantics prediction that the abstract models have low prediction errors on the validation dataset.

Enhanced ensemble strategy. We propose two comparison sets between: 1) individual controllers and traditional ensemble methods, and 2) newly proposed ensemble methods and the classical ones. Also, we apply multiple major and minor specifications to obtain a detailed and comprehensive evaluation for each controller. From the enriched evaluation results, we discover that the ensemble controllers can deliver better performance than individual controllers, the semantics prediction can reinforce the ensemble strategies, and finally, the Coordinator methods can outperform any others in each system in terms of major specifications. Besides the Majority-Voting and Averaging methods, some advanced decision fusion strategies have been proposed to enhance the overall ensemble performance, such as Bayes Optimal Classifier, Stacked Generalization, Super Learner, Consensus, and QueryBy-Committee. We do not apply these methods in our study since 1) unfeasibility in a large dataset and complex environments, and 2) unavailability for tasks in CPSs. Particularly, Bayes Optimal Classifier, Stacked Generalization, and Super Learner are not applicable to systems with multiple sub-learners and large datasets. Methods like Consensus, and Query-By-Committee, are specialized for unsupervised learning and active learning, respectively, that do not fit the context of CPSs.

We present SIEGE, a semantics-guided ensemble control framework for AI-CPS, which aims to construct an efficient, robust, and reliable control system for multi-objectives, and complex CPSs. A semantics-based abstraction is used to decompose and describe the system status in real-time and predicts the incoming semantics regarding the specification satisfactions.

We propose a series of ensemble methods that leverage the semantics predictions to optimally generate synergistic control signals from multiple DRL controllers. Further, we adopt an additional DRL agent as a coordinator to form an end-to-end DRL hierarchical control framework to perform a more flexible ensemble strategy. We performed comprehensive evaluations to investigate the performance of different individual controllers and ensemble methods on five industry-level, complex CPSs. The results demonstrated that SIEGE outperforms the state-of-the-art individual controllers and traditional ensemble methods. Our framework is also helpful in delivering a more robust, efficient, and safety-assured control system. To facilitate further research along this direction, we made our source code and experimental details publicly available on our project website.

With the increasing trend of CPSs adopting AI components in the loop, we hope our early exploratory work in this paper could inspire further extensive research along this direction towards providing better quality, safety and reliability assurance techniques for the upcoming era of AIenabled cyber-physical systems.