Accepted for Presentation at 2022 IEEE International Symposium on Antennas and Propagation, Denver, CO, July 2022

Optimization of 5G Infrastructure Deployment

Abstract—The application of machine learning for optimal The apartment is assumed to contain 306 receivers evenly
deployment of 5G infrastructure, such as the position and the distributed within the room to measure the signal coverage and
orientation of the antenna that help achieve the best signal 306 transmitters to test the coverage corresponding to all
coverage, is investigated in this paper. This avoids the need to possible deployments. One receiver and one transmitter are
perform on-site measurements or extensive software simulations.
Multivariate Regression (MR) and Neural Network (NN) models
collocated as an integrated pair and the spacing between each
were applied to predict the signal coverage in an indoor transmitter-receiver pair is 0.5 m. The transmitter-receiver pair
environment. The results showed that the average prediction error is mounted at a height of 1 m. The time cost will be significantly
using NN for the case investigated is 7 dB for a 60-GHz operating enormous if running the simulation for every single possible
frequency, whereas the error using the MR technique is lower than deployment of the transmitter is required; therefore, the
6 dB. The unique aspect in our work is the integration of the machine learning algorithm is proposed.
clustering algorithm and the NN machine learning model for
predicting indoor signal coverage. II. PROPOSED APPROACHES
The merit of the signal coverage is defined as the weighted
I. INTRODUCTION
sum of the received signal strength (RSS) at all positions in the
With the large-scale commercial usage of 5G room. We assume that locations such as the living room and
communications being the goal for the next decade, the bedrooms require stronger signals than restrooms and the
reception of 5G signal is becoming an issue. This is due to the kitchen because the placement is supposed to be optimized for
unique characteristics of the 5G signals whose frequency can rooms where the user may be most often. Consider the 𝑚𝑚𝑡𝑡ℎ
be above 40 GHz in order to improve the communications transmitter at location (𝑥𝑥𝑡𝑡𝑚𝑚 , 𝑦𝑦𝑡𝑡𝑚𝑚 ) is enabled and as a result the
capacity. However, these high signal frequencies undergo RSS of the 𝑛𝑛𝑡𝑡ℎ receiver at location (𝑥𝑥𝑟𝑟𝑛𝑛 , 𝑦𝑦𝑟𝑟𝑛𝑛 ) is 𝑃𝑃(𝑚𝑚,𝑛𝑛) , the
significant attenuation and are distorted while interacting with
optimal placement �𝑥𝑥� 𝑚𝑚 � 𝑚𝑚
𝑡𝑡 , 𝑦𝑦𝑡𝑡 � is defined as:
various materials in the propagation medium. This factor makes
it more challenging to characterize the received signal pattern. �𝑥𝑥�
𝑚𝑚 �
, 𝑦𝑦 𝑚𝑚 � = arg max ∑𝑁𝑁
𝑡𝑡 𝑡𝑡 𝑛𝑛=1 𝑤𝑤(𝑥𝑥 𝑛𝑛 ,𝑦𝑦 𝑛𝑛 ) 𝑃𝑃(𝑚𝑚,𝑛𝑛) ,
𝑟𝑟 𝑟𝑟
(1)
𝑚𝑚
Some work has been undertaken to predict the signal coverage
for indoor and outdoor environments [1], [2] and to develop where 𝑤𝑤(𝑥𝑥𝑟𝑟𝑛𝑛 ,𝑦𝑦𝑟𝑟𝑛𝑛 ) is the weight for the 𝑛𝑛 receiver. The ground
𝑡𝑡ℎ

numerical models for the signal propagation of 5G networks [3]. truth of received signal power is obtained from the Wireless
This paper investigates machine learning approaches for the InSite® ray-tracing simulation software which can provide
accurate and efficient prediction of signal coverage at a accurate estimates of signal power in situ. The data collected
frequency of 60 GHz in an indoor environment and thus, to are randomly divided into a training set and a testing set to
provide a strategy for the deployment of 5G infrastructure. verify the accuracy of the algorithms.
Results are compared with traditional approaches. The A. Neural Networks (NN)
representative scene investigated here is an apartment of floor
area around 60 m2, as shown in Fig. 1. 1) Description
Neural networks are machine learning models widely used
for nonlinear regression. The prediction of signal coverage is
generalized so as to extract the relationship between the RSS
and the position of the 𝑚𝑚𝑡𝑡ℎ transmitter and the 𝑛𝑛𝑡𝑡ℎ receiver. A
neural network, shown in Fig. 2, was designed with three layers,
including one hidden layer with a specific number of hidden
units. The prediction of the signal power 𝑃𝑃� can be described as
Fig. 1. View of the indoor environment, wherein the receivers and transmitters
𝑃𝑃�(𝑚𝑚,𝑛𝑛) = 𝑊𝑊2 �tansig�𝑊𝑊1 𝑋𝑋(𝑚𝑚,𝑛𝑛) + 𝑏𝑏1 �� + 𝑏𝑏2 , (2)
are evenly distributed within the rooms.
Accepted for Presentation at 2022 IEEE International Symposium on Antennas and Propagation, Denver, CO, July 2022

where 𝑊𝑊 is the weight, 𝑏𝑏 is the bias, 𝑋𝑋(𝑚𝑚,𝑛𝑛) is the normalized in those areas will provide better signal coverage compared to
input space containing the coordinates of the transmitter and the transmitters deployed in regions with dark blue pixels.
receiver, and tansig[⋅] is the activation function. The However, the MR algorithm predicts the signal coverage at a
coefficient in the network is updated to minimize the sum of the low resolution, because only the average signal power of each
squares of the errors between the prediction of the signal power cluster is predicted and the power will be assigned to each
and the ground truth, obtained from Wireless Insite, using the receiver within that cluster, while the NN algorithm predicts the
Levenberg-Marquardt (LM) algorithm [4]. signal coverage for all the receivers.
2) K-Means Clustering
The success of the design of neural networks may depend
on reasonable values of initial weights and number of hidden
units. While the attenuation of the wave along the propagation
paths is inversely proportional to the square of the length of the
Fig. 2. NN structure.
path, the pattern of the received signal power will not strictly
follow this pattern due to the interactions of high-frequency
waves with walls and furniture. Consequently, it is appropriate
Fig. 3. RMSE comparison.
to develop neural networks applied for different groups of
receivers which are classified by K-Means Clustering based on
the distance of propagation and the signal power collected from
the simulation, rather than to design a single neural network to
be applied for all the receivers. Since the number of unknown
variables in the network is large, clustering facilitates network
convergence. The number of clusters determined by the well-
known elbow method is eight (8). The number of hidden units
in each cluster is optimized through a number of trials.
(a)
B. Multivariate Regression (MR)
The multivariate regression is to obtain the relationship
between independent variables by minimizing the sum of
squares of error of prediction. The model of regression is
expressed as
𝑌𝑌𝑖𝑖 = 𝛽𝛽0 + 𝛽𝛽1 𝑥𝑥1 + 𝛽𝛽2 𝑥𝑥2 + ⋯ + 𝛽𝛽𝑘𝑘 𝑥𝑥𝑘𝑘 , (3)
where 𝑌𝑌𝑖𝑖 denotes the output which is a function of inputs 𝑥𝑥𝑘𝑘 ,
and the features of the input space is the same as that in Neural (b) (c)
Fig. 4. Average signal power corresponding to the position of transmitter,
Networks. The 𝛽𝛽𝑘𝑘 s are parameters to be estimated by ordinary
obtained by: (a) Wireless InSite; (b) NN; (c) MR. The transmitter should be
least square regression, using placed in the bright yellow regions for best overall coverage.
� = (𝑋𝑋 𝑇𝑇 𝑋𝑋)−1 𝑋𝑋 𝑇𝑇 𝑦𝑦.
𝜷𝜷 (4)
IV. CONCLUSIONS
III. RESULTS This paper has presented a machine learning NN algorithm
A. Performance of MR and NN to efficiently predict 5G signal coverage to optimize transmitter
deployment in an indoor apartment environment with high
The performance of the MR and the NN models are resolution, and compared it to the statistical MR method. The
compared by plotting the root-mean-square error (RMSE) of NN algorithm yields accurate predictions with the knowledge
the prediction as a function of the size of the training set, as of only 20% of the data collected from the simulation software.
shown in Fig. 3. The testing error using the NN has an average This implies that the transmitter deployment strategy can be
RMSE of 7 dB while the MR gives 5.9 dB. The fluctuating plot inferred quickly and accurately.
of the error for the NN method might come from the
randomness of the initial weights and the fact that the NN REFERENCES
depends on the appropriate choice of the number of hidden units. [1] H. Huang, et al, ‘‘Deep learning for super-resolution channel estimation
and DOA estimation based massive MIMO system,’’ IEEE Trans. Veh.
B. Prediction of Signal Coverage Technol., vol. 67, no. 9, pp. 8549–8560, Sep. 2018.
The signal coverage corresponding to the transmitter placed [2] Z. Li, et al., “Impact of wall penetration loss on indoor wireless
networks,” IEEE Wirel. Propag. Lett., vol. 20, pp. 1888–1892, 2021.
at a particular position in each room can be predicted by either [3] T. Imai, et. al, “Outdoor-to-indoor path loss modeling for 0.8 to 37 GHz
method discussed in the previous sections. The best position to band,” in European Conf. Antennas Propag. (EuCAP), Davos,
locate the transmitter could then be determined by the average Switzerland, Apr. 2016, doi: 10.1109/EuCAP.2016.7481469.
of predicted signal power, as shown in Fig. 4 for the entire [4] L. Azpilicueta, et al., “A ray launching-neural network approach for
radio wave propagation analysis in complex indoor environments,”
apartment. The yellow pixels indicate that transmitters placed IEEE Trans. Antennas Propag., vol. 62, no. 5, pp. 2777–2786, 2014.