UcnPowerNet: Residual learning based clustering and interference coordination in 5G user-centric networks
UcnPowerNet: Residual learning based clustering and interference coordination in 5G user-centric networks
复制标题
DOI：
10.1002/ett.4324
复制
发表时间：
2021-06
影响因子：
3.6
通讯作者：
Jianghui Liu;Hongtao Zhang
中科院分区：
文献类型：
--
作者：
Jianghui Liu;Hongtao Zhang
在第五代及以后（5G&B）无线网络中，可以探索人工智能来解决具有不确定性、时变和复杂特征的问题。以用户为中心的网络（UCN）可以消除小区边界并减少干扰。而在智能资源管理和干扰协调方面，现有的基于深度学习的工作的可扩展性随着网络规模的增加而大大下降，这需要堆叠更多的层并导致梯度消失问题。在这篇文章中，提出了一种新的基于深度学习的资源管理模型UcnPowerNet，以近似复杂UCN中动态聚类和干扰协调的迭代算法，其中多个剩余块通过快捷连接连接以渐进地学习输入和输出之间的映射。具体而言，在真实的域上实现UCN的协同加权最小均方误差（WMMSE）算法，生成大规模的近优训练数据集;然后通过最小化均方误差损失函数来训练UcnPowerNet逼近WMMSE算法的输出。此外，卷积层和归一化层分别用于减少权重的数量和避免梯度消失。大量的实验证明了UcnPowerNet的高近似精度，相对于传统的迭代算法，其总速率为94.90%，同时实现了超过100倍的加速。
In 5th Generation and beyond (5G&B) wireless networks, artificial intelligence can be explored to address problems with uncertain, time‐variant, and complex features. User‐centric network (UCN) can eliminate cell boundaries and reduce interference. And as for the intelligent resource management and interference coordination, the scalability of existing deep learning‐based works degrades greatly as the size of the network increases, which requires more layers to be stacked and causes gradient vanishing problem. In this article, a new deep learning‐based resource management model, UcnPowerNet, is proposed to approximate iterative algorithms for dynamic clustering and interference coordination in complicated UCN, where multiple residual blocks are concatenated with shortcut connections to asymptotically learn the mapping between input and output. Specifically, a cooperative weighted minimum mean square error (WMMSE) algorithm for UCN is carried out in the real domain to generate large near‐optimal training datasets; then we train UcnPowerNet to approximate the output of WMMSE algorithm by minimizing the loss function of mean square error. Moreover, convolutional layer and normalization layer are leveraged to decrease the number of weights and avoid gradients vanishing, respectively. Extensive experiments demonstrate the high approximation accuracy of UcnPowerNet with 94.90% sum‐rate relative to the conventional iterative algorithm while achieving more than 100 × speed up.