SCIENCE CHINA Information Sciences, Volume 63 , Issue 8 : 180303(2020) https://doi.org/10.1007/s11432-020-2937-y

Multitask deep learning-based multiuser hybrid beamforming for mm-wave orthogonal frequency division multiple access systems

More info
  • ReceivedJan 2, 2020
  • AcceptedJun 3, 2020
  • PublishedJul 15, 2020


Multiuser hybrid beamforming of a wideband millimeter-wave (mm-wave) system is a complex combinatorial optimization problem. It not only needs large training data, but also tends to overfit and incur long run-time when multiple serial deep learning network models are used to solve this problem directly. Preferably, multitask deep learning (MTDL) model could jointly learn multiple related tasks and share their knowledge among the tasks, and this has been demonstrated to improve performance, compared to learning the tasks individually. Therefore, this work presents a first attempt to exploit MTDL for multiuser hybrid beamforming for mm-wave massive multiple-input multiple-output orthogonal frequency division multiple access systems. The MTDL model includes a multitask network architecture, which consists of two tasks- user scheduling and multiuser analog beamforming. First, we use the effective channel with a low dimension as input data for the two parallel tasks to reduce the computational complexity of deep neural networks. In a shallow shared layer of the MTDL model, we utilize hard parameter sharing in which the knowledge of multiuser analog beamforming task is shared with the user scheduling task to mitigate multiuser interference. Second, in the training process of the MTDL model, we use the exhaustive search algorithm to generate training data to ensure optimal performance. Finally, we choose the weight coefficient of each task by traversing all weight coefficient combinations in the training phase. Simulation results prove that our proposed MTDL-based multiuser hybrid beamforming scheme could achieve better performance than traditional algorithms and multiple serial single tasks deep learning scheme.


This work was supported by National Natural Science Foundation of China (Grant Nos. 61871321, 61901367), National Science and Technology Major Project (Grant No. 2017ZX03001012-005), and Shaanxi STA International Cooperation and Exchanges Project (Grant No. 2017KW-011).


[1] Roh W, Seol J Y, Park J. Millimeter-wave beamforming as an enabling technology for 5G cellular communications: theoretical feasibility and prototype results. IEEE Commun Mag, 2014, 52: 106-113 CrossRef Google Scholar

[2] Bogale T E, Le L B. Massive MIMO and mmWave for 5G Wireless HetNet: Potential Benefits and Challenges. IEEE Veh Technol Mag, 2016, 11: 64-75 CrossRef Google Scholar

[3] Ayach O E, Rajagopal S, Abu-Surra S. Spatially Sparse Precoding in Millimeter Wave MIMO Systems. IEEE Trans Wireless Commun, 2014, 13: 1499-1513 CrossRef Google Scholar

[4] Sharma S K, Bogale T E, Le L B. Dynamic Spectrum Sharing in 5G Wireless Networks With Full-Duplex Technology: Recent Advances and Research Challenges. IEEE Commun Surv Tutorials, 2018, 20: 674-707 CrossRef Google Scholar

[5] Samimi M K, Rappaport T S. Ultra-wideband statistical channel model for non line of sight millimeter-wave urban channels. In: Proceedings of IEEE Global Communications Conference (Globecom), 2014. 3483--3489. Google Scholar

[6] Kwon G, Park H. A joint scheduling and millimeter wave hybrid beamforming system with partial side information. In: Proceedings of IEEE International Conference on Communications (ICC), 2016. Google Scholar

[7] He S, Wu Y, Ng D W K. Joint Optimization of Analog Beam and User Scheduling for Millimeter Wave Communications. IEEE Commun Lett, 2017, 21: 2638-2641 CrossRef Google Scholar

[8] Jiang Z, Chen S, Zhou S. Joint User Scheduling and Beam Selection Optimization for Beam-Based Massive MIMO Downlinks. IEEE Trans Wireless Commun, 2018, 17: 2190-2204 CrossRef Google Scholar

[9] Jiang J, Kong D. Joint User Scheduling and MU-MIMO Hybrid Beamforming Algorithm for mmWave FDMA Massive MIMO System. Int J Antennas Propagation, 2016, 2016: 1-10 CrossRef Google Scholar

[10] Qin Z, Ye H, Li G Y. Deep Learning in Physical Layer Communications. IEEE Wireless Commun, 2019, 26: 93-99 CrossRef Google Scholar

[11] Lu C, Xu W, Shen H. MIMO Channel Information Feedback Using Deep Recurrent Network. IEEE Commun Lett, 2019, 23: 188-191 CrossRef Google Scholar

[12] He Y F, Zhang J, Wen C K, et al. TurboNet: a model-driven DNN decoder based on max-log-map algorithm for turbo code. 2019,. arXiv Google Scholar

[13] Ruder S. An overview of multi-task learning in deep neural networks. 2017,. arXiv Google Scholar

[14] Sun Y, Wang X G, Tang X O. Deep learning face representation by joint identification-verification. 2014,. arXiv Google Scholar

[15] Alkhateeb A, El Ayach O, Leus G. Channel Estimation and Hybrid Precoding for Millimeter Wave Cellular Systems. IEEE J Sel Top Signal Process, 2014, 8: 831-846 CrossRef ADS arXiv Google Scholar

[16] Cai M M, Gao K, Nie D, et al. Effect of wideband beam squint on codebook design in phased-array wireless systems. In: Proceedings of IEEE Global Communications Conference (GLOBECOM), 2016. Google Scholar

[17] Wang B, Jian M, Gao F. Beam Squint and Channel Estimation for Wideband mmWave Massive MIMO-OFDM Systems. IEEE Trans Signal Process, 2019, 67: 5893-5908 CrossRef ADS arXiv Google Scholar

[18] Wang M, Gao F, Shlezinger N. A Block Sparsity Based Estimator for mmWave Massive MIMO Channels With Beam Squint. IEEE Trans Signal Process, 2020, 68: 49-64 CrossRef ADS arXiv Google Scholar

[19] Caruana R. Multitask learning. Machine Learning, 1997, 28: 41-75 CrossRef Google Scholar

[20] Nair V, Hinton G E. Rectified linear units improve restricted boltzmann machines. In: Proceedings of the 27th International Conference on Machine Learning (ICML-10), 2010. 807--814. Google Scholar

[21] Srivastava N, Hinton G F, Krizhevsky A, et al. Dropout: a simple way to prevent neural networks from overfitting. J Mach Learn Res, 2014, 15: 1532--4435. Google Scholar

[22] Kingma D, Ba J. Adam: a method for stochastic optimization. In: Proceedings of the 3rd International Conference on Learning Representations (ICLR), 2014. Google Scholar

  • Figure 1

    The structure of multi-user hybrid beamforming scheme.

  • Figure 2

    The structure of multitask deep learning-based multiuser hybrid beamforming scheme.

  • Figure 3

    The structure of single task deep learning-based multiuser hybrid beamforming scheme.

  • Figure 4

    The sum-rate cumulative distribution function (CDF) of different schemes with SNR = 0 dB. (a) ${{N}_{\rm~RF}}=2$, ${{N}_{\rm~RB}}=2$; (b) ${{N}_{\rm~RF}}=2$, ${{N}_{\rm~RB}}=4$.

  • Figure 5

    The sum-rate of different scheme versus SNR. (a) ${{N}_{\rm~RF}}=2$, ${{N}_{\rm~RB}}=2$; (b) ${{N}_{\rm~RF}}=2$, ${{N}_{\rm~RB}}=4$.

  • Figure 6

    The elapsed time of different schemes.


    Algorithm 1 Generation of the training data


    Output:$\mathcal{R}$, $\mathcal{B}$. Initialization: $~\mathcal{M}\in~O\left(~{{K}_{\rm~all}}\times~{{N}_{c}}~\right)$, $~\mathcal{U}\in~O\left(~{{N}_{c}}\times~{{N}_{\rm~RB}}~\right)$, $\mathcal{R}\in~O\left(~{{K}_{\rm~all}}\times~{{N}_{\rm~RB}}~\right)$, $\mathcal{B}\in~O\left(~{{K}_{\rm~all}}\times~{{N}_{c}}~\right)$, the selected MU-MIMO user set ${{\Omega~}_{S}}={{\emptyset}}$.

    According to the best beam index of each user, set the corresponding element of $\mathcal{M}$ to 1, e.g., if ${{\boldsymbol~B}_{k}}=n_c$, then $\mathcal{M}(k,n_c)=1$.

    Assume there are $Q_{{n}_{c}}$ possible schemes of resource allocation for users with the same best beam. Exploiting the exhaustive search algorithm to find the user scheduling scheme with the maximum sum rate. for ${{n}_{c}}=1:{{N}_{c}}$ for ${q}=1:{Q_{{n}_{c}}}$ $\left~\{~k^*~\right~\}=~{\mathop~{\arg~\max~}\nolimits_{q}}~{\sum_{n=1}^{N_{\rm~RB}}}~\Big(~{\frac{{{{\left\|~{{{\overline~{\boldsymbol~H}~}_{k,n}}}~\right\|}^2}}}{{{\sigma~^2}}}}~\Big)$, $\mathcal{U}({{n}_{c},n})={{k}^{*}}$ and ${\mathcal~R}({n_c},n)~=~1$, end end

    After resource allocation, each beam is regarded as a virtual OFDMA user multiplexing the whole frequency resource. Then, the integrated channel of each virtual OFDMA user is merged as follows, $\Omega~=\{~{{\tilde{\boldsymbol~H}}_{{{n}_{1}}}},{{\tilde{\boldsymbol~H}}_{{{n}_{2}}}},\ldots,{{\tilde{\boldsymbol~H}}_{{{N}_{c}}}}~\}$ and ${{\tilde~{\boldsymbol~H}}_{{n_c}}}~=~{[~{{{\boldsymbol~H}_{{\cal~U}({n_c},1)}}|{{\boldsymbol~H}_{{\cal~U}({n_c},2)}}|\cdots~|{{\boldsymbol~H}_{{\cal~U}({n_c},{N_{\rm~RB}})}}}~]_{{N_r}~\times~{N_{{t}}}}}$ is the channel matrix of a user allocated in the $n$th RB for the ${n}_{c}$th virtual OFDMA user. When a frequency resource is not been allocated to any user, ${\boldsymbol~H}_{\mathcal{U}({{n}_{c}},{n}})$ is equal to zero matrix.

    Select ${N}_{\rm~RF}$ virtual OFDMA users to maximizes sum-rate. for ${{n}_{c}}=1:{{N}_{\rm~RF}}$ if $~{n_c}~=~\mathop~{\max~}\nolimits_{{{\tilde~H}_{{n_c}}}~\in~\Omega~}~{\log~_2}\Big(~{1~+~\frac{{{{\|~{{\boldsymbol~U}_{{n_c}}^{\rm~H}{{{\tilde~{\boldsymbol~H}}}_{{n_c}}}{{\boldsymbol~V}_{{n_c}}}}~\|}^2}}}{{{\delta~^2}~+~\sum\nolimits_{j~\in~{\Omega~_s},{n_c}~\in~\Omega~}~{{{\|~{{\boldsymbol~U}_j^{\rm~H}{{{\tilde~{\boldsymbol~H}}}_{{n_c}}}{{\boldsymbol~V}_j}}~\|}^2}}~}}} +~\sum\nolimits_{j~\in~{\Omega~_s}}~{\frac{{{{\|~{{\boldsymbol~U}_j^{\rm~H}{{{\tilde~{\boldsymbol~H}}}_j}{{\boldsymbol~V}_j}}~\|}^2}}}{{{\delta~^2}~+~{{\|~{{\boldsymbol~U}_{{n_c}}^{\rm~H}{{{\tilde~{\boldsymbol~H}}}_j}{{\boldsymbol~V}_{{n_c}}}}~\|}^2}~+~\sum\nolimits_{i~\in~{\Omega~_s},i~\ne~j}~{{{\|~{{\boldsymbol~U}_i^{\rm~H}{{{\tilde~{\boldsymbol~H}}}_j}{{\boldsymbol~V}_i}}~\|}^2}}~}}}~~\Big)$, where ${\boldsymbol~U}$ and ${\boldsymbol~V}$ are the left unitary matrix and the right unitary matrix of the singular value decomposition of virtual OFDMA user ${{\tilde~{\boldsymbol~H}}}_{n_c}$, respectively. then ${\cal~B}(:,{n_c})~=~1$, $\Omega~~\leftarrow~\Omega~\backslash~\{~{{\tilde~{\boldsymbol~H}}_{{n_c}}}\}$, ${\Omega~_s}~\leftarrow~{\Omega~_s}~\cup~\{~{{\tilde~{\boldsymbol~H}}_{{n_c}}}\}~$. end

Copyright 2020  CHINA SCIENCE PUBLISHING & MEDIA LTD.  中国科技出版传媒股份有限公司  版权所有

京ICP备14028887号-23       京公网安备11010102003388号