EAGER: SC2: PHY-Layer-Integrated Collaborative Learning in Spectrum Coordination

随着无线设备的爆炸性增长，频谱正成为无线系统激烈争夺的稀缺资源。为了确保未来的民用和军用系统(从物联网(IoT)设备到战场自组织网络)继续支持质量越来越高的服务，系统必须超越传统的频谱许可模式，朝着智能频谱共享模式发展，在这种模式中，网络节点协作以高效地共享频谱。这种彻底的范式转变需要将最新的机器学习进展与软件定义无线电的最新进展相结合，以便赋予无线设备必要的智能和灵活性，以实现高效无监督频谱共享的愿景。这项拟议的研究旨在解决这一根本问题，并将提供充足的机会，在机器学习和通信工程的交叉点为学生提供跨学科培训。该项目设想了无线电设计的范式转变，将灵活的通信工程技术与先进的机器学习算法交织在一起，将传统的物理层和链路层融合为一个“协作层”。具体地说，该方法包括三个关键元素：(1)物理层的多载波调制格式，其提供了对干扰信号做出反应所需的灵活性；(2)高性能的调制识别软件，其利用深度学习和卷积神经网络的最新进展来准确地对环境中的射频信号进行分类；以及(3)决策模块，其利用遗憾最小化在线算法的最新进展，以在无线环境中实现高的探测与开发性能。

With the explosion of wireless devices, spectrum is becoming a scarce resource that wireless systems fiercely compete for. To ensure that future civilian and military systems, ranging from connected Internet of Things (IoT) devices to battlefield ad-hoc networks, continue to support services with growing quality, systems must evolve beyond the traditional spectrum licensing model and towards an intelligent spectrum sharing paradigm, in which networks nodes collaborate to efficiently share the spectrum. This radical paradigm shift requires the integration of the latest machine learning advances with the more recent progress in software defined radio, in order to endow wireless devices with the intelligence and agility necessary to realize the vision of efficient unsupervised spectrum sharing. The proposed research aims at addressing this fundamental issue, and will offer ample opportunities to provide interdisciplinary training of students at the intersection of machine learning and communications engineering.The project envisions a paradigm shift in radio design, which will intertwine agile communications engineering techniques with advanced machine learning algorithms to fuse the traditional physical-layer and link layers into a "collaboration layer". Specifically, the approach comprises three key elements: (1) a multi-carrier modulation format at the physical layer that provides the required agility to react to interfering signals; (2) a high-performing modulation recognition software that exploits the latest advances in deep learning and convolutional neural networks to accurately classify the radio frequency signals in the environment; and (3) a decision module exploiting the latest advances in regret minimization online algorithms to achieve high exploration versus exploitation performance in the wireless environment.