CAREER: AutoEdge: Deep Reinforcement Learning Methods and Systems for Network Automation at Wireless Edge

下一代无线技术承诺具有极高带宽和超低延迟的先进网络功能，将促进运输、媒体和制造业等垂直行业的广泛新型移动服务和客户应用。网络连接的爆炸式增长和网络服务的多样化将极大地增加网络管理的复杂性。该CAREER项目旨在开发特定领域的深度强化学习（DRL）方法和系统，以实现下一代无线边缘计算网络中网络资源和服务的自动化配置、供应和编排。本次CAREER项目的成功完成将促进对DRL、通信、计算和网络之间内在关系的理解，为研究无线边缘计算中基于学习的网络自动化算法和系统奠定坚实的基础。此外，该项目开发的技术将大大降低无线网络的运营成本，从而为包括低收入和偏远社区在内的所有社区提供负担得起的高性能无线连接。此外，该项目提供跨学科教育，通过将研究与教育、产学研和跨学科合作相结合，培养掌握先进无线和人工智能（AI）技术的下一代工程师和研究人员。该CAREER项目旨在开发深度强化学习（DRL）方法和系统，使无线边缘计算网络中的端到端资源编排自动化。为此，本文研究了两个基本问题：1)如何设计特定领域的DRL，有效解决大规模无线边缘计算网络中的端到端编排问题；2)如何在大规模网络系统中高效部署基于DRL的编排解决方案。针对第一个问题，本项目研究了特定领域DRL的状态、奖励函数、训练算法和神经网络的设计，开发了基于DRL的端到端资源编排中各种约束的处理方法，以避免约束违反，并设计了上下文感知的多智能体DRL方法，利用无线边缘计算的领域知识来提高DRL的学习效率。为了解决第二个问题，本项目开发了策略蒸馏方法来解决网络模拟与真实网络系统差异导致的DRL部署问题，设计了跨尺度知识转移方法来解决小规模试验台与大规模无线边缘计算系统尺寸不匹配导致的DRL部署问题。该项目还开发了一个增强网络模拟器和一个边缘计算系统原型，用于评估基于drl的端到端编排解决方案。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The next-generation wireless technology promises advanced network capabilities with extremely high bandwidth and ultra-low latency that will catalyze a wide range of new mobile services and customer applications in vertical sectors such as transport, media, and manufacturing. The explosion of networking connections and the diversification of network services will dramatically increase the complexity of network management. This CAREER project aims to develop domain-specific deep reinforcement learning (DRL) methods and systems to automate the configuration, provisioning, and orchestration of network resources and services in next-generation wireless edge computing networks. The successful completion of this CAREER project will advance the understanding of the inherent relationships among DRL, communications, computing, and networking and lay a solid foundation for studying learning-based algorithms and systems for network automation in wireless edge computing. Besides, the technologies developed in the project will significantly reduce the operational cost of wireless networks and thus allow affordable high-performance wireless connectivity for all communities including low-income and remote communities. Moreover, the project provides interdisciplinary education to cultivate next-generation engineers and researchers who master both advanced wireless and Artificial Intelligence (AI) technologies via the integration of research into education and industrial-academic and cross-disciplinary collaborations.This CAREER project aims to develop deep reinforcement learning (DRL) methods and systems that automate end-to-end resource orchestration in wireless edge computing networks. Toward this end, two fundamental research problems are investigated: 1) how to design domain-specific DRL that can effectively solve end-to-end orchestration problems in large-scale wireless edge computing networks and 2) how to efficiently deploy DRL-based orchestration solutions in large-scale networking systems. To solve the first problem, the project studies the design of states, reward functions, training algorithms, and neural networks of domain-specific DRL, develops methods of handling various constraints in DRL-based end-to-end resource orchestration to avoid constraint violations, and designs context-aware multi-agent DRL methods to leverage domain knowledge of wireless edge computing to improve the learning efficiency of DRL. To solve the second problem, this project develops policy distillation methods to address the DRL deployment issues caused by the divergence between network simulations and real network systems, and designs cross-scale knowledge transfer methods to address the DRL deployment issues caused by the mismatch of the dimensions of small-scale testbeds and large-scale wireless edge computing systems. The project also develops an augmented network simulator and an edge computing system prototype for evaluating DRL-based end-to-end orchestration solutions.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.