Transforming networks - building an intelligent optical infrastructure (TRANSNET)

光网络是全球数字通信基础设施的基础，它们的发展同时刺激了数据需求的增长，并通过释放光纤通道的容量来应对这一需求。事实证明，UNLOC方案赠款内的工作成功地了解了点对点非线性光纤通道容量的基本限制。然而，下一代数字基础设施需要的不仅仅是原始容量--它需要提供通道和灵活的资源和容量，并结合具有最大数据吞吐量的低延迟、简化和模块化的网络架构，以及结合整体网络安全的网络弹性。如何建设这样一个智能和灵活的网络是一个具有全球重要性的重大问题。为了应对网络内对延迟敏感的需求的日益动态变化，并使技术互联网成为可能，目前的光网络容量过剩，导致过度设计和未利用的容量。因此，一个关键的挑战是了解如何智能地利用有限的光网络资源来动态最大化性能，同时还提高对未来未知需求的健壮性。Transnet的目标是通过创建能够在需要时随时随地动态提供容量的自适应智能光网络来应对这一挑战-下一代数字基础设施的主干。我们的愿景和雄心是将智能引入所有级别的光通信、云和数据中心基础设施，并开发能够以最佳方式动态响应容量、覆盖范围和延迟的不同应用需求的光收发器。我们预计，机器学习(ML)将在未来的光网络中变得无处不在，从数字编码、均衡和损伤缓解，到监控、故障预测和识别，以及信号恢复、交通模式预测和资源规划，都将在未来的光网络设计和运营中无处不在。Transnet将专注于机器技术的应用，以开发新的光收发技术系列，以适应新一代自x(x=配置、监控、规划、学习、修复和优化)网络体系结构的需要，能够在优化资源使用的同时考虑物理信道特性和高级应用。我们将应用ML技术将物理层和网络结合在一起；光纤的非线性在网络环境中带来了特别复杂的挑战，因为它在所有传输的波长通道的信号质量之间产生了相互依赖的关系。在对数百个独立频道、数千公里范围内的数十种可能的调制格式进行优化时，强力优化变得不可行。特别是大规模网络的异构性和优化网络拓扑和资源分配的计算复杂性，以及对未来网络的动态和数据驱动的管理、监测和控制，这需要一种新的思维方式和定制的方法。我们建议通过机器学习和概率技术的结合来降低网络设计的复杂性，使其能够自学习网络智能和适应。这将导致创建计算高效的方法来处理具有记忆和噪声的新兴非线性系统的复杂性，用于在不同的时间和长度尺度上动态运行的网络。这是一种全新的光网络设计和优化方法，需要基于对非线性物理、信号处理和光网络的理解来推进机器学习和启发式算法设计的跨学科方法。

Optical networks underpin the global digital communications infrastructure, and their development has simultaneously stimulated the growth in demand for data, and responded to this demand by unlocking the capacity of fibre-optic channels. The work within the UNLOC programme grant proved successful in understanding the fundamental limits in point-to-point nonlinear fibre channel capacity. However, the next-generation digital infrastructure needs more than raw capacity - it requires channel and flexible resource and capacity provision in combination with low latency, simplified and modular network architectures with maximum data throughput, and network resilience combined with overall network security. How to build such an intelligent and flexible network is a major problem of global importance. To cope with increasingly dynamic variations of delay-sensitive demands within the network and to enable the Internet of Skills, current optical networks overprovision capacity, resulting in both over- engineering and unutilised capacity. A key challenge is, therefore, to understand how to intelligently utilise the finite optical network resources to dynamically maximise performance, while also increasing robustness to future unknown requirements. The aim of TRANSNET is to address this challenge by creating an adaptive intelligent optical network that is able to dynamically provide capacity where and when it is needed - the backbone of the next-generation digital infrastructure.Our vision and ambition is to introduce intelligence into all levels of optical communication, cloud and data centre infrastructure and to develop optical transceivers that are optimally able to dynamically respond to varying application requirements of capacity, reach and delay. We envisage that machine learning (ML) will become ubiquitous in future optical networks, at all levels of design and operation, from digital coding, equalisation and impairment mitigation, through to monitoring, fault prediction and identification, and signal restoration, traffic pattern prediction and resource planning. TRANSNET will focus on the application of machine techniques to develop a new family of optical transceiver technologies, tailored to the needs of a new generation of self-x (x = configuring, monitoring, planning, learning, repairing and optimising) network architectures, capable of taking account of physical channel properties and high-level applications while optimising the use of resources. We will apply ML techniques to bring together the physical layer and the network; the nonlinearity of the fibres brings about a particularly complex challenge in the network context as it creates an interdependence between the signal quality of all transmitted wavelength channels. When optimising over tens of possible modulation formats, for hundreds of independent channels, over thousands of kilometres, a brute force optimisation becomes unfeasible. Particular challenges are the heterogeneity of large scale networks and the computational complexity of optimising network topology and resource allocation, as well as dynamical and data-driven management, monitoring and control of future networks, which requires a new way of thinking and tailored methodology.We propose to reduce the complexity of network design to allow self-learned network intelligence and adaptation through a combination of machine learning and probabilistic techniques. This will lead to the creation of computationally efficient approaches to deal with the complexity of the emerging nonlinear systems with memory and noise, for networks that operate dynamically on different time- and length-scales. This is a fundamentally new approach to optical network design and optimisation, requiring a cross-disciplinary approach to advance machine learning and heuristic algorithm design based on the understanding of nonlinear physics, signal processing and optical networking.