CMfg-SC(云制造服务组合)、DRL、DQN的定义

cloud manufacturing

Information Technology (IT) is continuously growing so that it is the origin of many of core technologies such as cloud computing, Internet of Things (IoT), embedded systems, semantic web, service-oriented

         ——《Service load balancing and logistic optimization in cloud manufacturing using genetic algorithm》

---

CMfg-SC (cloud manufacturing service composition):

1. 云制造服务组合是将一系列较小粒度的云服务按照一定流程和规则组装起来构成较大粒度的增值服务来协同完成一个任务。云服务组合是实现云制造按需分配和使用资源的有效途径。

                                  ——《云制造特征及云服务组合关键问题研究》

2. Service composition is one of the core functions of CMfg service
   platform. In the face of diversified demands of users, the premise of CMfg service composition optimization is to analyze different situations and problems in multitask requests accurately. The key to cloud manufacturing service composition is selecting the best service to meet user requirements from massive and aggregated candidate services

         ——《Service composition model and method in cloud manufacturing》

3. CMfg service composition is an important link in the entire CMfg platform operation and an effective approach to realize on-demand distribution and use of resource services. The composition process is based on task demand mode. A series of CMfg composition services with logical relationships is formed by selecting functionally aggregated resources from CMfg service pool, which allows task publishers to rent service sources on demand. This process is closely related to every stage of the entire life cycle and many key technologies of CMfg services.

         ——《Service composition model and method in cloud manufacturing》

---

DRL(Deep reinforcement learning):

深度强化学习是一种结合了深度学习和强化学习的方法。前者的基本思想是通过多层网络结构和非线性变换，组合底层特征，形成抽象的、易于区分的高层表示，以发现数据的分布式特征表示。后者主要思想是智能体通过与环境交互和试错得到的评价性的反馈信号来实现决策的优化。深度强化学习通过使用深度学习网络作为强化学习的非线性函数逼近的方式，将深度学习的感知能力与强化学习的决策能力结合在一起，用来解决状态和动作空间的高维问题。

DRL is a method that combines deep learning and reinforcement learning. The basic idea of the former is to combine the low-level features through multi-layer network structure and nonlinear transformation to form an abstract and easily distinguishable high-level representation, so as to discover the distributed feature representation of data. The main idea of the latter is that the agent interacting with the environment, learning an optimal policy, by trial and error. By using deep learning network as the nonlinear function approximation of reinforcement learning, deep learning combines the perception ability of deep learning with the decision-making ability of reinforcement learning to solve the high-dimensional problem of state and action space.

---

DQN(Deep Q-Network):

which utilizes a convolutional neural network to learn successful control policies in complex RL environments. The network is trained with a variant of the Q-learning [Watkins C J C H, Dayan P. Q-learning[J]. Machine learning, 1992, 8(3-4): 279-292.] algorithm, with stochastic gradient descent to update the weights. Moreover, To alleviate the problems of correlated data and non-stationary distributions, usually use an experience replay mechanism which randomly samples previous transitions, and thereby smooths the training distribution over many past behaviors.

---

cloud-based manufacturing should comply task as service oriented, customer centric, and demand driven manufacturing, in a way to enable industrial control systems, service composition, and flexibility