2022 IEEE Globecom Workshops (GC Wkshps): Workshop on Emerging Topics in 6G Communications

Artificial Intelligence-Based Spatial Domain Beam

Prediction for 5G Beyond
Jun Zuo∗ , Jiazhen Zhang∗ , Yuhua Cao∗ , Xinfang Chen† , Fei Wang∗ , Nan Hu∗ , Xiaodong Xu∗
∗
Institute of Wireless and Terminal Technology, China Mobile Research Institute,
Beijing, China, Email: {zuojun, zhangjiazhen, caoyuhua,wangfei, hunan,xuxiaodong}@chinamobile.com
† The Key Lab of Universal Wireless Communications, Ministry of Education,
2022 IEEE Globecom Workshops (GC Wkshps) | 978-1-6654-5975-4/22/$31.00 ©2022 IEEE | DOI: 10.1109/GCWkshps56602.2022.10008651

Beijing University of Posts and Telecommunications (BUPT), Beijing, China, Email: [email protected]

Abstract—In this paper, we investigate Artificial Intelligence allows UE to measure a small number of beam pairs and UE
(AI)-based beam management (BM) schemes to deal with the or BS predicts the best beam pair among all the beam pairs
beam selection issue with less beam measurement overhead. We with AI model, so as to reduce beam measurement overhead
design a VGG (Visual Geometry Group) based network, named
as VGG-BMNet, to predict the quality of all the beam pairs based or increase beam prediction accuracy under the same beam
on the measured quality of partial beam pairs. To improve the measurement overhead.
generalization capability of VGG-BMNet, a Unified-VGG-BMNet For AI-based spatial domain beam prediction, the beam
is proposed, which can achieve higher average beam prediction measurement configurations, including the number and the
accuracy than VGG-BMNet. To speed up the training time and pattern of measured beam pairs, can be fixed or uncertain.
reduce the size of training dataset, a transfer learning-based
VGG-BMNet is further designed. The simulation results show The beam prediction AI model trained for a particular beam
that transfer learning-based VGG-BMNet not only achieves good measurement configuration may not be suitable for other
generalization capability, but also has much faster convergence configurations. To support beam prediction under different
rate with fewer training samples. All the proposed schemes can beam measurement configurations, BS or UE side may need
reduce the beam measurement overhead by more than 80% to train and storage multiple AI models, hence the training
meanwhile achieving comparable beam selection accuracy as the
exhaustive beam measurement method. and memory storage cost problems of AI model become
Index Terms—Beam management, AI, beam prediction, gen- serious. To promote the practical application of AI-based
eralization, transfer learning spatial domain beam prediction and reduce the implementation
cost of multiple models, the generalization problem of AI
I. I NTRODUCTION model is a significant issue to be dealt with.
To overcome the generalization problem, we consider two
Beam management (BM) is an effective technique to im- different approaches. One way forward is to design a unified
prove network performance by dynamically selecting appro- AI model where the number of input neurons equals to the
priate beam pair between base station (BS) and each UE based total number of beam pairs. This unified AI model has good
on channel quality variation in 5G systems. With the widely generalization capability and can be compatible with any beam
application of massive MIMO and millimeter wave (mmWave) measurement configurations, but the scale and computational
communication techniques in satisfying boosting data rate complexity of neural networks may be large, e.g. the length
demand, the beam measurement overhead and complexity of of the input layer is proportional to the total number of beam
BM becomes even more severe due to the dramatic increasing pairs. To further reduce the scale and storage cost of neural
number of beam pairs, more complicated propagation environ- networks to favour the implementation or deployment of AI
ment, and more frequent beam pair switching [1-2]. Driven model, another way forward is to apply transfer learning to
by the powerful capability of Artificial Intelligence (AI) in make AI model trained under a certain beam measurement
making predictions and resolving complex decision-making configuration be adaptive to beam prediction under other
problems, AI-based BM emerges as a promising technology beam measurement configurations. Transfer learning is an
to address the BM issue in 5G beyond systems. efficient method to improve the performance of target AI
AI-based BM has drawn much interest in recent studies, model in unknown network environment by transferring the
which is mainly applied in beam-forming of large-scale an- structure and the parameters of pre-trained AI model and
tenna array systems for beam selection in mobility scenario fine-tuning with a small number of samples [7], which has
[3-5]. The Third Generation Partnership Project (3GPP) has been widely applied to recognition tasks [8], recommender-
recently started the standardization process of Release 18 of system applications [9], wireless-network applications [10],
New Radio (NR) access technology, where BM is one of etc. Transfer learning can solve the generalization problem
the typical use cases of AI for NR Air Interface topic [6]. of spatial domain beam prediction since that the relationship
Currently, the representative sub-use case of AI-based BM between the quality of measured beam pairs under new beam
focuses on AI-based spatial domain beam prediction, which measurement configuration and that of all beam pairs can be
978-1-6654-5975-4/22 © 2022 IEEE partially learned from the relationship between the quality of

978-1-6654-5975-4/22/$31.00 ©2022 IEEE 1460

Authorized licensed use limited to: Edinburgh Napier University. Downloaded on September 19,2023 at 13:50:27 UTC from IEEE Xplore. Restrictions apply.
 2022 IEEE Globecom Workshops (GC Wkshps): Workshop on Emerging Topics in 6G Communications

measured beam pairs under the existing beam measurement To maintain a proper beam pair adaptive to variable environ-
configuration and that of all beam pairs. ment, periodic beam measurement and reporting are required.
In this paper, we study the AI-based spatial domain beam The best beam pair can be determined by the exhaustive
prediction schemes based on VGG (Visual Geometry Group) measurement of NT X × NRX beam pairs. But the number
neural network [11]. For each beam measurement configu- of transmit beams at the BS is usually large (e.g.,NT X = 64 )
ration, a VGG based network, named as VGG-BMNet is in mmWave systems, it is unrealistic to exhaustively measure
designed. To improve the generalization capability of VGG- the quality of all the beam pairs due to the large measurement
BMNet, we further design a Unified-VGG-BMNet to be overhead. Instead, a limited number of transmit beams, i.e.,
compatible with different beam measurement configurations, KT X beams from the NT X beams, are selected for beam
meanwhile overcoming the expensive training and storage cost measurement, and the proper beam pair is selected from the
of multiple VGG-BMNets. Besides, a transfer learning-based KT X ×NRX beam pairs, but the best beam pair may be missed
VGG-BMNet model is proposed to reduce the training time since it may be not in the set of beam pairs for measurement.
of VGG-BMNet and the required number of training samples In this paper, we use an AI-based method to solve the best
to facilitate the implementation of AI model in real world. beam pair selection problem subject to the knowledge of beam
Numerical results verify the effectiveness of the proposed quality of partial beam pairs. The problem can be formulated
schemes and reveal that when achieving comparable beam as
selection accuracy as the exhaustive beam measurement, the 

proposed schemes can reduce the beam measurement overhead i = argmax fw RSRP(KT X × NRX ) (1)
by more than 80%. The Unified-VGG-BMNet has superior
generalization capability and achieves higher average predic- where i is the index of the beam pair with the largest RSRP
tion accuracy than VGG-BMNet, at the cost of larger compu- within all NT X × NRX beam pairs. The AI model with
tational complexity. When transfer learning with fine-tuning parameters fw (·) is used to predict the RSRP of all the
by a small number of training samples is used, the transfer NT X × NRX beam pairs based on the measured RSRP of
learning-based VGG-BMNet obtains higher beam prediction KT X × NRX beam pairs.
accuracy and faster convergence.
III. AI-BASED BM P ROCEDURE AND P ROPOSED S CHEMES
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
We consider a network with a BS and multiple UEs oper- In this section, we first describe the procedure of AI-based
ating in the mmWave frequency band. As shown in Fig. 1, BM, then the proposed AI-based schemes are introduced. We
both the BS and UE have multiple beams. To improve the rely on a VGG neural network to perform spatial domain
coverage and throughput, a proper beam pair for each UE and beam prediction. Different beam measurement configurations,
BS should be selected for data transmission. i.e. different values of KT X , are considered to predict the
beam quality of all the beam pairs. Firstly, we train VGG
based models for each configuration. Then we design a unified
model which can be applicable for different values of KT X
to improve the generalization capability of the model. To
reduce the training time and overhead of collecting training
samples, transfer learning based AI model is further proposed
to deal with the generalization problem under different beam
measurement configurations.

A. Procedures of AI-based BM
Fig. 1: System Model. The AI model for beam prediction can be deployed at
BS side or UE side, the BM procedure of two deployment
For traditional downlink BM procedure, the proper beam scenarios are respectively illustrated in Fig. 2(a) and Fig. 2(b)
pair from NT X beams at the BS side and NRX beams at the Considering the beam measurement overhead, BS usually
UE side is selected by the measurement of reference signals transmits a small number of beams for measurement. The
(RSs). BS transmits each RS with a specific beam direction, value of KT X configured for different UEs or for different
such that the beam quality can be derived from the received measurement time instances may be different. AI model will
power of the corresponding RS at the UE side. In 3GPP 5G use the beam quality of these measured beam pairs for spatial
standard, UE can report the quality of the measured beams to domain beam prediction. The Top-K best beam pairs from
the BS, and BS will select a proper beam from the reported all the predicted beam pairs can be selected and transmitted
beams. The reporting information includes the RS ID and the to UE to obtain the realistic beam quality. UE reports the
corresponding Reference Signal Received Power (RSRP) or realistic beam quality of Top-K beam pairs to BS, and BS
Signal to Interference plus Noise Ratio (SINR) of the reported indicates the best beam among them to UE for the subsequent
RS. data transmission.

1461
Authorized licensed use limited to: Edinburgh Napier University. Downloaded on September 19,2023 at 13:50:27 UTC from IEEE Xplore. Restrictions apply.
 2022 IEEE Globecom Workshops (GC Wkshps): Workshop on Emerging Topics in 6G Communications

According to the procedures illustrated in Fig. 2, the predicted

Top-K beam pairs will be measured by UE for best beam pair
selection, so we can define the beam prediction accuracy as
the percentage of the genie-aided best beam pair is one of the
Top-K predicted beam pairs. For different beam measurement
configurations, the size of input layer can be different, hence
multiple VGG-BMNets need to be trained for different beam
measurement configurations.
2) Unified-VGG-BMNet: The generalization capability of
VGG-BMNet is not satisfactory since each VGG-BMNet can
only be applied to a fixed beam measurement configuration,
e.g., fixed KT X and NRX . In the real network, due to the train-
ing cost and storage cost, it is not realistic to train AI model for
(a) Beam prediction at BS side each beam measurement configuration. To tackle this problem,
we design a Unified-VGG-BMNet which can be used for
all beam measurement configurations. For the Unified-VGG-
BMNet, the input data is composed of the beam pair IDs
and the corresponding RSRP values of all the NT X × NRX
beam pairs. Since only KT X ×NRX beam pairs are measured,
the RSRP value of the remaining (NT X − KT X ) × NRX
beam pairs can be set as a particular value, e.g., 0. The
output of Unified-VGG-BMNet is the predicted RSRP of all
the NT X × NRX beam pairs. In our study, we consider five
different beam measurement configurations to construct the
training dataset, and the trained Unified-VGG-BMNet can
work well for all these five configurations.
3) Transfer Learning: Although Unified-VGG-BMNet can
(b) Beam prediction at UE side be applied for all beam measurement configurations, the time
and cost for model training and dataset construction may still
Fig. 2: Procedures of AI-based beam prediction be large. Transfer learning is a technique in which a model
trained for one task with a large amount of data can be used as
the basic model for a new task. Inspired by this, we combine
B. AI-based beam prediction schemes
VGG-BMNet with transfer learning to reduce the training time
1) VGG-BMNet: VGG is a convolutional neural network and the amount of training data. We first collect the quality of
model used in image recognition. To achieve good perfor- measured beam pairs under beam measurement configuration
mance, multiple small size (3×3) convolution filters are used in A to construct the dataset, and a VGG-BMNet-A is trained
the VGG network. Inspired by this thought, we design a VGG- with the constructed dataset. Then, we train a VGG-BMNet-B
BMNet for spatial-domain beam prediction. The structure of for beam measurement configuration B. The structure of VGG-
the designed VGG-BMNet is shown in Fig. 3. The input of BMNet-B is the same as VGG-BMNet-A except for the input
layer, since the number of input neurons are related to the
number of measured beam pairs. We use the trained weights
of VGG-BMNet-A as the initial weights of VGG-BMNet-B
except for the input layer, and the weights of VGG-BMNet-
B are fine-tuned with a small dataset collected under beam
measurement configuration B. The same method can be used
for other beam measurement configurations.

IV. S IMULATION R ESULTS

In this section, the simulation results are provided to
Fig. 3: Structure of VGG-BMNet verify the superiority of the proposed Unified-VGG-BMNet
and transfer learning based method compared with VGG-
the model is the measured RSRP of the KT X × NRX beam BMNet under various beam measurement configurations. The
pairs, and the output of the model is the RSRP of all the basic parameters are listed in Table I. The indoor hotspot
NT X × NRX beam pairs. The Top-K beam pairs are selected channel model defined in 3GPP [12] is adopted to generate
according to the RSRP of all the NT X × NRX beam pairs. the dataset in our simulations. As for the training phase, the

1462
Authorized licensed use limited to: Edinburgh Napier University. Downloaded on September 19,2023 at 13:50:27 UTC from IEEE Xplore. Restrictions apply.
 2022 IEEE Globecom Workshops (GC Wkshps): Workshop on Emerging Topics in 6G Communications

TABLE I: BASIC SIMULATION PARAMETERS

Parameter Value
Scenario TR38.901: InH, 12 sites, 3 cells per site
SCS 120kHz
Bandwidth 40MHz
Carrier Frequency 30GHz
Tx power 20dBm
BS antenna config [Mg Ng M N P ] = [1 1 16 4 2]
UE antenna config [Mg Ng M N P ] = [1 3 1 4 2]
64 Tx = 16 horizontal * 4 vertical
Azimuth angle = [-16π/32, -14π/32, . . . ,
BS Tx beam 0, . . . , 12π/32, 14π/32]
Fig. 4: Beam prediction accuracy of VGG-BMNet
pattern Zenith angle = [2π/8 3π/8 4π/8 5π/8]
(azimuth, zenith) = (0, π/2) is the direction
perpendicular to the array
BS mech.
20 degree
tilting
4 Rx per panel = 4 horizontal * 1 vertical
UE panels deployment [-90, 0, 90]
UE Rx beam pattern
Azimuth angle = [-3*pi/8, 1*pi/8, pi/8, 3*pi/8]
Zenith angle = [pi/2]
UE height 1.5
Beam selection method L1-RSRP

default adaptive momentum (Adam) optimizer with learning

rate lr = 0.001 is adopted to train the proposed models.
Fig. 5: Beam prediction accuracy of Unified VGG-BMNet
A. Performance of VGG-BMNet
In our simulation, we set KT X = [4, 6, 8, 10, 12] for five
beam measurement configurations, and NRX = 4 is used in BMNets under several beam measurement configurations. For
all the configurations. For each beam measurement config- each configuration, the minimum value of K which satisfies
uration, we generate 39600 data samples. 90% of the data 95% or higher beam prediction accuracy is selected. It can
samples are used to train the model, and the rest samples are be noted that the AI-based schemes can achieve more than
used for testing. Five VGG-BMNets are respectively trained 80% beam measurement overhead reduction under most beam
with the dataset of each beam measurement configuration. measurement configurations.
For Unified-VGG-BMNet, the mixed dataset of all the five In Fig. 6 and Fig. 7, we compare the beam prediction
beam measurement configurations are used to train the model. accuracy of the five VGG-BMNets and the proposed Unified-
The beam prediction accuracy under each beam measurement VGG-BMNet when selecting Top-1 and Top-2 best beam pairs.
configuration, i.e. KT X × NRX , of the trained VGG-BMNets It can be seen that the Unified-VGG-BMNet achieves similar
and the Unified-VGG-BMNet are shown in Fig. 4 and Fig. 5, or better performance than the five VGG-BMNets under each
respectively. It can be seen that the beam prediction accuracy beam measurement configuration. The Unified-VGG-BMNet
increases with the value of K for all the scenarios. The reason
is that with larger value of K, the AI model has higher
TABLE II: BEAM MEASUREMENT OVERHEAD REDUC-
probability to choose the best beam pair from the Top-K
TION
beam pairs. Besides, when the number of measured beam
pairs KT X × NRX increases, the prediction accuracy is also Beam BM The Beam
improved.
measurement overhead value of prediction
When the beam prediction accuracy of AI model achieves
95% or higher, it can be said that the AI model has nearly the configuration reduction K accuracy
same performance as exhaustive measurement of 64 × 4 beam 12×4 80.08% 3 96.90%
pairs. The beam measurement overhead reduction of AI-based 10×4 83.20% 3 96.40%
beam measurement compared with the exhaustive measure- 8×4 85.94% 4 96.00%
ment is calculated as 1 − (KT X × NRX + K)/(NT X × NRX ).
6×4 88.67% 5 95.40%
In Table II, we show the BM overhead reduction of VGG-

1463
Authorized licensed use limited to: Edinburgh Napier University. Downloaded on September 19,2023 at 13:50:27 UTC from IEEE Xplore. Restrictions apply.
 2022 IEEE Globecom Workshops (GC Wkshps): Workshop on Emerging Topics in 6G Communications

outperforms VGG-BMNets on the average beam prediction TABLE III: FLOPS AND TRAINABLE PARAMETERS
accuracy of five beam measurement configurations, which COMPARISON
exhibits the excellent generalization performance.
Beam FLOPs Trainable Par.
Solution
configuration (×107 ) (×106 )
4×4 1.93 2.32
6×4 2.88 2.45
VGG-
8×4 3.84 2.58
BMNet
10×4 4.8 2.71
12×4 5.75 2.85
Unified-
VGG- 64×4 29.9 3.1
BMNet

respectively. For each target configuration, 10000 data samples

are used to fine-tune the basic model and train the VGG-
Fig. 6: Comparison bewteen Unified-VGG-BMNet and VGG-
BMNets without transfer learning. The simulation results
BMNet when K = 1
indicate that the transfer learning-based method outperforms
VGG-BMNet without transfer learning under all the target
beam measurement configurations.

Fig. 7: Comparison bewteen Unified-VGG-BMNet and VGG- Fig. 8: Beam prediction accuracy comparison when KT X = 4,
BMNet when K = 2 NRX = 4

The complexity of our proposed Unified-VGG-BMNet and

the five VGG-BMNets is listed in Table III from the per-
spective of floating point operations (FLOPs) and trainable
parameters. The results show that the FLOPs and trainable
parameters of Unified-VGG-BMNet are much larger than
VGG-BMNet, since the unified model need to be applicable
for various beam measurement configurations.
B. Performance of Transfer Learning
In our simulations, we use the VGG-BMNet trained with
39600 data samples of KT X = 12, NRX = 4 as the
basic model. We transfer the basic model to target beam
measurement configurations with the method described in
Section III. We set KT X = [4, 6, 8, 10] as four target beam
measurement configurations. The prediction accuracy compar-
ison between transfer learning-based VGG-BMNet and VGG- Fig. 9: Beam prediction accuracy comparison when KT X = 6,
BMNets without transfer learning under each target beam NRX = 4
measurement configuration are shown in Fig. 8, 9, 10 and 11,

1464
Authorized licensed use limited to: Edinburgh Napier University. Downloaded on September 19,2023 at 13:50:27 UTC from IEEE Xplore. Restrictions apply.
 2022 IEEE Globecom Workshops (GC Wkshps): Workshop on Emerging Topics in 6G Communications

Fig. 10: Beam prediction accuracy comparison when KT X =

8, NRX = 4
Fig. 12: Loss curves for VGG-BMNet with transfer learning
and without transfer learning when KT X = 8, NRX = 4

BMNet achieves higher beam prediction accuracy and faster

convergence.
R EFERENCES
[1] G. Eason, Y. R. Li, X. Zhang, et al. “Beam Management in Millimeter-
Wave Communications for 5G and Beyond,” IEEE Access, vol. 8, pp.
13282-13293, Jan. 2020.
[2] S. S. Kalamkar, F. Baccelli, F. M. A. Jr., et al. “Beam Management in
5G: A Stochastic Geometry Analysis,” IEEE Trans. Wireless Commun.,
vol. 21, no. 4, pp. 2275-2290, Apr. 2022.
[3] V. Va, T. Shimizu, G. Bansal, and R. W. Heath, Jr., “Online learning for
position aided millimeter wave beam training,” IEEE Access, vol. 7, pp.
30507-30526, 2019.
Fig. 11: Beam prediction accuracy comparison when KT X = [4] Y. Wang, M. Narasimha, and R.W. Heath, Jr., “MmWave beam prediction
10, NRX = 4 with situational awareness: A machine learning approach,” in Proc. IEEE
19th Int. Workshop Signal Process. Adv. Wireless Commun. (SPAWC),
Jun. 2018.
[5] Y. Lim, H. Ji, J. Park, et al. “Artificial Intelligence-Based Beam Manage-
The loss curves for transfer learning-based VGG-BMNet ment for High Speed Applications in mmWave Spectrum” in Proc. IEEE
Globecom Workshops (GC Workshops), Mar. 2021.
and VGG-BMNet without transfer learning are illustrated in [6] RP-213599, New SI: Study on Artificial Intelligence (AI)/Machine Learn-
Fig. 12. It can be noticed that the downward trend of transfer ing (ML) for NR Air Interface, RAN94-e, e-Meeting, Dec. 2021.
learning-based VGG-BMNet becomes flat much earlier than [7] F. Zhuang, Z. Qi, K. Duan, et al. “A Comprehensive Survey on Transfer
Learning,” Proceedings of the IEEE, vol. 109, no. 1, pp. 43-76, Jan. 2021.
that of VGG-BMNet without transfer learning, which means [8] U. Cote-Allard et al., “Deep learning for electromyographic hand gesture
the model trained by transfer learning can converge much signal classification using transfer learning,” IEEE Trans. Neural Syst.
faster. Thus, we conclude that transfer learning can speed Rehabil. Eng., vol. 27, no. 4, pp. 760–771, Apr. 2019.
[9] F. Zhuang, J. Zheng, J. Chen, et al. “Transfer collaborative filtering from
up the training of VGG-BMNet for target beam measurement multiple sources via consensus regularization,” Neural Netw., vol. 108,
configurations with fewer training samples. pp. 287–295, Dec. 2018.
[10] R. Li, Z. Zhao, X. Chen, et al. “TACT: A transfer actor-critic learning
framework for energy saving in cellular radio access networks,” IEEE
V. C ONCLUSION Trans. Wireless Commun., vol. 13, no. 4, pp. 2000–2011, Apr. 2014.
[11] K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks
In this paper, we propose several AI-based schemes for for Large-Scale Image Recognition ,” Computer Science, 2014.
spatial domain beam prediction. The proposed schemes can [12] “3rd generation partnership project; technical specification group radio
significantly reduce the beam measurement overhead when access network; study on channel model for frequencies from 0.5 to 100
GHz (release 16),” 3GPP, Sophia Antipolis, France, 3GPP Rep. TR 38.901
achieving comparable beam selection accuracy with the ex- V16.1.0, 2020
haustive beam measurement. The simulation results showed
that the proposed Unified-VGG-BMNet can achieve better
average beam prediction accuracy than VGG-BMNet. The
generalization capability of the Unified-VGG-BMNet is im-
proved at the cost of higher computation complexity. Besides,
the presented results verified the effectiveness of transfer
learning-based VGG-BMNet. When fine-tuning with a small
amount of data is used, the transfer learning-based VGG-

1465
Authorized licensed use limited to: Edinburgh Napier University. Downloaded on September 19,2023 at 13:50:27 UTC from IEEE Xplore. Restrictions apply.