Building a Flexible Framework Towards Autonomous Networking Using Machine Learning Techniques

机器学习在语音识别、图像处理等领域取得了前所未有的进步。ML在众多学科中扮演着越来越重要的角色。在对服务质量、性能和安全性有极高要求的下一代自主网络中，ML也扮演着关键角色。ML用于组网的例子包括ML在网络异常检测、网络流量预测、预防性维护等方面的应用。ML是数据驱动的，这比基于复杂算法或模型的网络解决方案更有效。然而，ML也是非平凡的，分布式网络系统极大地阻碍了ML的有效性，因为每个节点只有有限的局部信息，并且向其他节点学习很复杂。软件定义网络(SDN)将网络设备上的主要控制功能转换为逻辑上的中央控制器，从而实现了重大的范式转变。网络功能虚拟化(NFV)通常被网络功能虚拟化所采用，这些网络功能为网络运营商在管理资源、提高性能和部署创新服务方面提供灵活性。此外，带内网络遥测(INT)仍然是一项新兴的技术，它可以通过提供对ML至关重要的高网络数据可见性来改善当前的网络监控。除了ML，SDN还对INT产生了实质性的影响，因为SDN可以高效地向INT请求和收集关于整个网络的网络状态和数据，从而实现有效的网络分析，并极大地简化ML的复杂性。另一方面，NFV可以通过在单独且强大的设备上运行ML函数来促进ML。然而，ML仍然面临着挑战，例如ML选项多、培训时间长、针对不断变化的环境进行重复培训。建议研究的主要目标是建立一个集成SDN/NFG、ML和INT的灵活的系统框架。这些技术的无缝融合在简化采用ML进行联网方面具有很大的潜力。该框架将被设计为便于(由操作员或自动地)选择最适合于特定问题的适当ML技术，例如交通控制或安全。该框架的制定将基于对网络数据的特征和模式以及不同应用问题的特点的透彻分析。此外，提出的研究还超越了当前的网络范式和ML技术，考虑了物联网(IoT)和以信息为中心的网络(ICN)，并跨问题域迁移学习。较长期的目标是调查拟议的综合框架对新兴范例的适应情况。迁移学习的目的是减少ML所需的训练时间，并在网络变化或跨相似领域时支持知识重用。

Machine learning (ML) has made unprecedented advancement in various areas, e.g., voice recognition and image processing. The role of ML is becoming more and more important in numerous disciplines. ML also plays a key role in next generation autonomous networking which has extreme high requirements in quality of service, performance, and security. Examples of ML for networking include applications of ML for network anomaly detection, network traffic prediction, preventive maintenance, etc. ML is data driven, which is more effective than complex algorithm- or model-based approaches to network solutions. However, ML is also nontrivial, and the distributed network systems tremendously hinder the effectiveness of ML, as each node only has limited local information and learning from other nodes is complex. Software-defined networking (SDN) has made a major paradigm shift by converting main control functions on network devices into a logically central controller. Network function virtualization (NFV) has often been adopted by virtualizing network functions that provide flexibility for network operators in managing resources, improving performance, and deploying innovative services. Further, in-band network telemetry (INT), still a developing technique, can improve current network monitoring by providing high network data visibility which is crucial for ML. In addition to ML, SDN also has a substantial impact on INT, as SDN can efficiently request and gather network states and data from INT about the entire network, which enables effective network analytics and greatly simplifies the complexity of ML. On the other hand, NFV can facilitate ML by running ML functions on a separate and powerful device. However, there are still challenges for ML, such as many ML options, long training time, and repetitive training for changing environments. The main objective of the proposed research is to build a flexible system framework that integrates SDN/NFG, ML, and INT. The seamless fusion of those technologies has substantial potential in simplifying the adoption of ML for networking. The framework will be designed to facilitate the selection (either by a human operator or automatically) of an appropriate ML technique that is best suitable for a specific problem, e.g., traffic control or security. The development of the framework will be based on thorough analyses of the characteristics and patterns of network data and the features of different application problems. Moreover, the proposed research is also looking beyond the current network paradigm and ML technologies by considering Internet of Things (IoT) and Information-centric Networking (ICN), and transfer learning across problem domains. The longer-term goal is to investigate the adaptation of the proposed integrative framework for the emerging paradigms. Transfer learning is targeted to reduce the training time needed for ML and to support knowledge reuse when network changes or across analogous domains.