Next Generation Software-defined Intelligent Radio Access Network (SIRAN) - Leveraging Deep Learning for Autonomous and Intelligent Service Provisioning

下一代5G及以后的无线网络将依赖于软件定义智能无线接入网（SIRAN）设备，这些设备可以随时重新配置，以满足移动网络运营商（mno）的多样化和不断变化的需求，为各种应用提供高质量的服务，从增强型移动宽带通信到大规模机器类型通信，再到超可靠的低延迟通信。SIRAN可以提供“切片”的通信和计算资源，以保证服务质量（QoS）和用户体验质量（QoE）。此外，可扩展sians支持集成宏、微、微和/或飞蜂窝的异构网络（HetNets）。sians前所未有的灵活性和可编程性，以及网络业务需求、用户流量特征、用户位置和移动性的多变性和动态性，对sians的高效运行和管理提出了巨大挑战，需要优化利用频谱、信道带宽和传输时间调度、传输功率、计算和存储资源以及能耗等网络资源。同时满足服务和应用的QoS/QoE要求。当系统条件动态变化时，多个运行参数需要同时优化时，传统的建模和优化技术难以处理未来的多业务HetNets。我们的总体目标是通过开发实时管理SIRAN资源分配的技术（例如，通信、缓存、计算）来填补这一空白。我们将利用当代机器学习，特别是深度学习技术来优化资源利用，同时满足所需的QoS/QoE。我们提出的技术和解决方案将利用siran的可编程性实现自主和智能的网络服务供应。我们将开发无模型和组合建模/无模型技术，在深度学习引擎的驱动下，根据动态变化的网络和用户流量条件下的业务需求，快速使系统运行适应所需的最优运行区域。该项目开发的技术将成为未来与移动网络运营商合作的行业伙伴关系项目的基础，以收集网络数据，使这些技术能够根据实际网络条件进行评估，并开发测试平台，以实验验证我们的工作和概念验证技术转让。该项目将为培训下一代无线网络工程师和研究人员提供一个极好的机会，这些工程师和研究人员熟悉使用当代机器智能技术来解决下一代无线网络的复杂性和动态性。

Next generation 5G and beyond wireless networks will rely on software-defined intelligent radio access network (SIRAN) equipment that can be readily reconfigured to satisfy the diverse and changing needs of mobile network operators (MNOs) to provide high-quality services for a wide range of applications, from enhanced mobile broadband communications, to massive machine-type communications, to ultra-reliable low-latency communications. SIRAN enables provisioning of "slices" of communication and computing resources to guarantee quality of service (QoS) and users' quality of experience (QoE). Furthermore, scalable SIRANs enables heterogeneous networks (HetNets) that integrate macro-, micro-, pico- and/or femto-cells. Such unprecedented flexibility and programmability of SIRANs together with the variability and dynamicity of the network service demands, user traffic characteristics, and user location and mobility present great challenges to the efficient operation and management of SIRANs by optimizing the utilization of network resources including frequency spectrum, channel bandwidth and transmission time schedule, transmission power, computation and storage resources, and energy consumption, while satisfying the QoS/QoE requirements of services and applications. Classical modeling and optimization techniques have difficulty dealing with future multi-service HetNets when multiple operational parameters need to be simultaneously optimized while system conditions are dynamically changing. Our overall objective is to fill this gap by developing techniques to manage in real-time the allocation of SIRAN resources (e.g., communication, caching, computing). We will leverage contemporary machine learning, particularly deep learning techniques to optimize resource utilization while satisfying the required QoS/QoE. Our proposed techniques and solutions will enable autonomous and intelligent network service provisioning that takes advantage of the programmability of SIRANs. We shall develop both model-free as well as combined modeling/model-free techniques, driven by deep-learning engines to quickly adapt system operation towards the desired optimal operation region based on service requirements under dynamically varying network and user traffic conditions. The techniques developed in this project will form the basis of future industry-partnership projects in collaboration with MNOs to collect network data that enables these techniques to be evaluated based on practical network conditions, and to develop testbeds for experimental verification of our work and proof-of-concept technology transfer. This project will provide an excellent opportunity to train the next generation wireless networking engineers and researchers who are knowledgeable on the use of contemporary machine intelligence techniques to address the complexity and dynamic nature of the next generation wireless networks.