CPS: Small: Learning How to Control: A Meta-Learning Approach for the Adaptive Control of Cyber-Physical Systems

物联网（IoT）使能的制造系统形成了一类特别重要的网络物理系统（CPS）。支持物联网的制造系统具有由在工厂车间运送和处理材料的机器的异构混合物编织而成的物理织物。这些系统的网络结构是有线和无线数字通信网络的异构混合体，能够实现用于管理物理结构工作流的数据流的全球可见性。这些支持物联网的系统是复杂的CPS，具有很大的建模不确定性。物理和网络结构对外部环境是开放的，外部环境可能以突然和不可预测的方式发生变化。这种转变可能是由于客户工作订单的变化或由于导致网络结构的无线网络中的业务拥塞的环境变化。两个结构的动态是耦合的，因为物理结构中的拥塞可能会在网络结构中产生拥塞，反之亦然。这种复杂性和不确定性是美国制造商更广泛接受物联网技术的主要障碍。为了降低采用物联网技术的风险，该项目提出开发元学习方法，学习如何控制物联网制造中的复杂CPS。 该项目将开发控制CPS的元学习方法的算法和软件实现。 该项目将在一个测试平台上对该方法的性能进行基准测试，该测试平台捕捉物联网制造系统的物理和网络结构之间的复杂交互。该项目使用元学习算法来控制复杂和不确定的网络物理系统。该方法采用了一种新型的机器学习模型，称为行为有序抽象（BOA）。与其他深度学习方法相比，该模型具有更强的跨任务泛化能力、更好的样本效率和更高的可解释性。这种建模方法允许该项目通过将元学习嵌入到广义调节器中来解决有关深度强化学习的鲁棒稳定性的问题，该广义调节器学习“如何”在所有任务中配置控制器合成。该项目将在一个多机器人测试平台上评估该项目的“学习如何控制”框架，该测试平台模仿使用WIFI连接的机器人在工厂地板上移动材料。 该项目将研究如何将在试验台上学习到的模型和政策转移到当地制造设施中的物联网工厂。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估而被认为值得支持。

Internet-of-Things (IoT) enabled manufacturing systems form a particularly important class of cyber-physical systems (CPS). IoT-enabled manufacturing systems have a physical fabric woven from a heterogeneous mix of machines carrying and processing materials across the factory floor. The cyber fabric for these systems is a heterogeneous mix of wired and wireless digital communication networks enabling the global visibility of the data streams used to manage the physical fabric’s workflows. These IoT-enabled systems are complex CPS with a great deal of modeling uncertainty. The physical and cyber fabrics are open to an external environment that can shift in an abrupt and unpredictable manner. Such shifts may be due to changes in customer work orders or due to environmental changes that cause traffic congestion in the cyber fabric’s wireless networks. The dynamics of both fabrics are coupled since congestion in the physical fabric may create congestion in the cyber fabric and vice versa. This complexity and uncertainty stand as major obstacles to the broader acceptance of IoT technologies by U.S. manufacturers. To lower the risk in adopting IoT technologies, this project proposes developing meta-learning methods that learn how to control complex CPS found in IoT-enabled manufacturing. This project will develop algorithms and software implementations of the meta-learning approach to controlling CPS. The project will benchmark the method’s performance on a testbed capturing the complex interactions between an IoT-manufacturing system’s physical and cyber fabrics. This project uses meta-learning algorithms for the control of complex and uncertain cyber-physical systems. The approach adopts a new type of machine learning model called a behaviorally ordered abstraction (BOA). This model has greater cross-task generalization capacity, better sample efficiency, and greater interpretability than other deep learning methods. This modeling approach allows the project to address issues regarding the robust stability of deep reinforcement learning by embedding meta-learning in a generalized regulator that learns “how” to configure controller synthesis across all tasks. This project will evaluate the project’s “learning-how-to-control” framework on a multi-robotic testbed mimicking the use of WIFI connected robots moving materials across a factory floor. The project will investigate how to transfer the models and policies learned on the testbed to IoT-enabled factories found in local manufacturing facilities.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.