Handling Coexistence of LoRa with Other Networks through Embedded Reinforcement Learning
Handling Coexistence of LoRa with Other Networks through Embedded Reinforcement Learning
复制标题
DOI：
10.1145/3576842.3582383
复制
发表时间：
2023-05
期刊：
影响因子：
--
通讯作者：
Sezana Fahmida;Venkata Prashant Modekurthy;Mahbubur Rahman;Abusayeed Saifullah
中科院分区：
文献类型：
--
作者：
Sezana Fahmida;Venkata Prashant Modekurthy;Mahbubur Rahman;Abusayeed Saifullah
各种低功耗广域网（LPWAN）技术在有限频谱上的快速发展带来了它们共存的挑战。今天，LPWAN还没有能力应对这一迫在眉睫的挑战。对于低功率节点，很难采用复杂的介质访问控制协议。WiFi或传统短距离无线网络的共存处理将不适用于LPWAN。由于长距离，它们的节点可能会受到前所未有数量的隐藏节点的影响，需要高能效技术来处理这种共存。在本文中，我们解决了LoRa的共存问题，这是一种领先的LPWAN技术。为了提高LoRa网络与多个独立网络共存时的性能，我们提出了一种基于LoRa节点轻量级强化学习的新型嵌入式学习代理的设计。这是通过开发一个Q学习框架来实现的，同时确保LoRa节点的内存和计算开销最小。该框架根据节点做出的传输决策，利用传输反馈作为来自网络的反馈。据我们所知，这是第一个用于处理低功耗网络共存的Q学习方法。考虑到LoRa网络的各种共存场景，我们通过室内和室外实验来评估我们的方法。室外测试结果表明，我们的Q学习方法在LoRa网络中平均实现了46%的数据包接收率提高，同时降低了66%的能耗。在室内实验中，我们观察到一些共存场景，其中当前的LoRa网络丢失了所有数据包，而我们的方法使数据包接收率达到99%，能耗提高了90%。
The rapid growth of various Low-Power Wide-Area Network (LPWAN) technologies in the limited spectrum brings forth the challenge of their coexistence. Today, LPWANs are not equipped to handle this impending challenge. It is difficult to employ sophisticated media access control protocol for low-power nodes. Coexistence handling for WiFi or traditional short-range wireless network will not work for LPWANs. Due to long range, their nodes can be subject to an unprecedented number of hidden nodes, requiring highly energy-efficient techniques to handle such coexistence. In this paper, we address the coexistence problem for LoRa, a leading LPWAN technology. To improve the performance of a LoRa network under coexistence with many independent networks, we propose the design of a novel embedded learning agent based on a lightweight reinforcement learning at LoRa nodes. This is done by developing a Q-learning framework while ensuring minimal memory and computation overhead at LoRa nodes. The framework exploits transmission acknowledgments as feedback from the network based on what a node makes transmission decisions. To our knowledge, this is the first Q-learning approach for handling coexistence of low-power networks. Considering various coexistence scenarios of a LoRa network, we evaluate our approach through experiments indoors and outdoors. The outdoor results show that our Q-learning approach on average achieves an improvement of 46% in packet reception rate while reducing energy consumption by 66% in a LoRa network. In indoor experiments, we have observed some coexistence scenarios where a current LoRa network loses all the packets while our approach enables 99% packet reception rate with up to 90% improvement in energy consumption.