Smart Optical Networks Enabled by Machine Learning

随着业务量的不断增加，波分复用（WDM）传输系统的数据速率、容量和灵活性不断提高。视频和云应用，以及未来的5G和物联网（IoT），要求光网络的流量和动态性更高。这使得链路和网络级别的性能下降和故障的潜在影响更加严重，对灵活和自主的网络管理的需求比以往任何时候都更加重要。机器学习（ML）作为一种将人工智能（AI）引入无线网络以优化频谱和网络资源的方法在过去十年中得到了研究，但其在光网络中的应用仍处于起步阶段。具有监控功能的灵活转发器的可用性使得利用ML的潜力在软件定义网络（SDN）环境中设计和管理日益异构、动态和复杂的网络成为可能。*******随着我们进入知识定义网络（KDN）时代，提出的研究计划旨在通过自适应和学习机制研究认知概念，以提高光网络的灵活性和鲁棒性。该项目将探索基于机器学习的方法，用于光学层的QoT估计和光信噪比（OSNR）预测，以及网络层的异常检测和主动故障管理。有监督和无监督的机器学习，以及强化学习（RL）技术，将使用合成和现场数据进行研究。拟议研究计划的长期目标是执行KDN的测试平台演示，展示ML在性能预测方面的优势，以及在光路供应和故障管理方面主动优于传统反应方法的优势。*******提出的研究计划的原创性和科学影响在于机器学习在光网络领域的巨大潜力，这为创新开辟了极其丰富的机会，以及用于训练机器学习模型并使其更适合现实世界应用的现场数据的可用性。该研究计划的结果将有助于制定预防策略，使高容量光网络更可靠，更易于管理。其他一些创新包括：用于QoT估计的新ML模型；性能演化与异常检测的新预测方法新的主动方案，简化和自动化的网络运行和更高的网络可靠性。培养涉及先进网络和机器学习多学科研究的高素质人才（HQP）的潜力是该研究计划的主要资产。三名博士生，四名硕士生和几名本科生将参与研究计划，预计在五年的时间框架内完成。******

The continuous traffic increase has led to wavelength division multiplexing (WDM) transmission systems with ever-increasing data rates, capacity and flexibility. Video and cloud applications, as well as future 5G and Internet of Things (IoT), call for even higher traffic volumes and dynamicity in optical networks. This makes the potential impact of performance degradation and failure at the link and network levels more severe and the need for flexible and autonomous network management more important than ever. Machine learning (ML) has been researched in the last ten years as a way to bring Artificial intelligence (AI) in wireless networks for optimizing spectrum and network resources, but its application in optical networks is still in its infancy. The availability of flexible transponders with monitoring capabilities makes it possible to leverage the potential of ML to design and manage increasingly heterogeneous, dynamic and complex networks in a software-defined network (SDN) context.*******As we are entering into the era of knowledge-defined networking (KDN), the proposed research program aims at investigating cognitive concepts for improving the flexibility and robustness of optical networks through adaptive and learning mechanisms. The program will explore ML-based methods for QoT estimation and optical signal to noise ratio (OSNR) prediction in the optical layer, as well as for anomaly detection and proactive fault management at the network level. Supervised and unsupervised ML, as well as reinforcement learning (RL) techniques, will be investigated using both synthetic and field data. The long-term objective of the proposed research program is to perform a testbed demonstration of a KDN showing the benefits of ML for performance prediction and the advantages of proactive over conventional reactive approaches for lightpath provisioning and fault management.*******The originality and scientific impact of the proposed research program lies in the huge potential of ML to be exploited in the field of optical networking, which opens extremely rich opportunities for innovation, as well as on the availability of field data for training the ML models and making them more suitable for real-world applications. The results of this research program will enable the development of preventive strategies for making high capacity optical networks more reliable and easier to manage. Some additional innovations include: new ML models for QoT estimation; new predictive methods for performance evolution and anomaly detection; new proactive schemes for simplified and automated network operation and higher network reliability. The potential for training highly qualified personnel (HQP) involving multidisciplinary research on advanced networking and ML is a major asset of the research program. Three Ph.D. students, four Master's students and several undergraduate students will be involved in the research program to be completed in an expected time frame of five years.******