A Convolutional Neural Network based
Resource Management Algorithm for NOMA
enhanced D2D and Cellular Hybrid Networks
1
st
Zhenfeng Zhang
School of Electronics and Information
Northwestern Polytechnical University
Xian, Shaanxi, 710072, China
2450182446@mail.nwpu.edu.cn
2
nd
Daosen Zhai
School of Electronics and Information
Northwestern Polytechnical University
Xian, Shaanxi, 710072, China
zhaidaosen@nwpu.edu.cn
3
rd
Ruonan Zhang
School of Electronics and Information
Northwestern Polytechnical University
Xian, Shaanxi, 710072, China
rzhang@nwpu.edu.cn
4
th
Xiao Tang
School of Electronics and Information
Northwestern Polytechnical University
Xian, Shaanxi, 710072, China
tangxiao@nwpu.edu.cn
5
th
Yutong Wang
School of Electronics and Information
Northwestern Polytechnical University
Xian, Shaanxi, 710072, China
wangyutong@mail.nwpu.edu.cn
Abstract—This paper mainly studies the channel and
power allocation for the device-to-device (D2D) and cel-
lular hybrid network with non-orthogonal multiple access
(NOMA) technology. We formulate the joint channel and
power allocation problem as a mixed integer programming
problem (MIP). Since the MIP is non-convex and NP-hard,
the computational complexity of the traditional optimization
method is very high. To overcome this drawback, we con-
struct a convolutional neural network (CNN) to approximate
traditional optimization methods. Specifically, the inputs of
the CNN are the channel state information of users, and the
outputs are the channel allocation and power control policies.
The relation between the inputs and the outputs is established
by a hidden layer, which consists of a convolutional layer, a
pooling layer, and a fully connected layer. The simulation
results indicate that the CNN based resource allocation
scheme can achieve a good performance with a ultra-low
computational complexity.
Index Terms—Device to device (D2D), non-orthogonal
multiple access (NOMA), heterogeneous network (HetNet),
convolutional neural network (CNN), channel allocation,
power control.
I. INTRODUCTION
In recent years, the machine learning techniques [1], [2]
has been widely applied. The use of a deep neural network
(DNN), i.e., deep learning, has gained in popularity in
fields including image classification and speech recogni-
tion. Several deep learning technologies have also been
applied in wireless communication research, e.g., transmit
power control, channel estimation, encoder/decoder de-
sign, and cognitive satellite communication in conjunction
with reinforcement learning. The use of DNN can enable
us to exploit an optimal control strategy for communica-
tion systems without the need to solve complex problems
This work was supported in part by the National Natural Science
Foundation of China under Grant 61901381 and Grant 61901379, in part
by the China Postdoctoral Science Foundation under Grant BX20180262
and Grant 2018M641019, in part by the Natural Science Basic Research
Plan in Shaanxi Province under Grant 2019JQ-631 and Grant 2019JQ-
253, and in part by the Fundamental Research Funds for the Central
Universities under Grant G2019KY05302.
explicitly. Moreover, given that a trained DNN model
with moderate number of layers can be used with a low
computation time, it is well suited to real-time operations.
The deep neural networks (DNNs) can learn rich pat-
terns and approximate arbitrary function mappings [3].
In [4], the DNN was used to treat a given signal pro-
cessing (SP) algorithm as a black box, and to learn its
input/output relation. It was shown that only a small net-
work is sufficient to obtain high approximation accuracy,
and DNNs can achieve orders of magnitude speedup in
computational time compared to the state-of-the-art inter-
ference management algorithm. In [5], it was shown that
combining DNN with ensemble learning results of a high-
performance and low-complexity power control method
outperforms the state-of-the-art methods in a variety of
system configurations. The authors in [6] used the concept
of universal function approximation of DNNs and the
theory of Lagrangian duality to show that, despite the non-
convex nature of these problems, they can be formulated as
a finite-dimensional unconstrained optimization problem
in the dual domain. For arbitrarily large DNNs, it can
obtain the precise optimality with respect to the original
problem. In [7], an ultra-low complexity algorithm was
designed which is based on the DNN. It can directly utilize
the relationship learned by the training samples to generate
the channel allocation results according to the channel
state information (CSI).
The convolutional neural network (CNN) is a feed-
forward neural network with convolutional computation
and deep structure. Specifically, the convolution kernel
parameter sharing and sparseness of the inter-layer con-
nections in the hidden layers enable CNN to learn the
features of the input data. Besides, CNN also has a lot of
advantages. For instance, it only consumes a small amount
of computation to capture the features of the original data
and performs a stable effect. In [8], the authors used the
CNN to solve the linear sum assignment problems, such as
the channel allocation in an interference-free network. The