| 摘要: |
| 传统的功率分配算法由于复杂的矩阵运算与迭代所造成的高时延，在实际通信中实时获取信道信息十分困难，当前重要的研究方向是在系统性能和计算复杂度之间找到有效平衡。针对终端直通（Device-to-Device，D2D）用户与蜂窝用户的联合功率分配问题，提出一种异构功率控制图神经网络（Heterogeneous Power Control Graph Neural Network，HPCGNN）算法，旨在最大化所有用户的加权和速率。首先通过构建干扰的异构图，将信道和噪声等信息嵌入到图的节点和边；再由HPCGNN完成消息传递和更新，采用无监督学习方式优化深度神经网络(Deep Neural Network,DNN)参数，最终得到最佳的功率分配。仿真结果表明，相较于其他深度学习算法，所提算法能够有效提高系统性能，且在损失5%性能下相较分式规划（Fractional Programming，FP）能降低82%~98%的时间复杂度。 |
| 关键词: D2D：功率分配：异构图神经网络 |
| DOI：10.20079/j.issn.1001-893x.240129003 |
|
| 基金项目:重庆市自然科学基金创新发展联合基金（中国星网）（CSTB2023NSCQ-LZX0114） |
|
| D2D Joint Power Allocation Based on Heterogeneous Graph Neural Network |
| CHEN Fatang,XU Xiaopeng,WANG Wenhao,LIU Ze |
| (School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China) |
| Abstract: |
| Traditional power allocation algorithms are difficult to obtain channel information in real time in real communication due to the high latency caused by complex matrix operations and iterations,and the current important research direction is to find an effective balance between system performance and computational complexity.For the joint power allocation problem between device-to-device(D2D) users and cellular users,a heterogeneous power control graph neural network(HPCGNN) algorithm is proposed,which aims to maximize the weighted sum of all the users rate.Firstly,by constructing a heterogeneous graph of interference,the information such as channel and noise is embedded into the nodes and edges of the graph.Then the HPCGNN completes the message passing and updating,and uses unsupervised learning to optimize the deep neural network(DNN) parameters,and ultimately obtains the optimal power allocation.Simulation result shows that compared with other deep learning algorithms,the proposed algorithm can effectively improve the system performance,and can reduce the time complexity by 82%~98% compared with Fractional Programming(FP) at a loss of 5% performance. |
| Key words: D2D:power allocation:heterogeneous graph neural network |
