ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) | 978-1-7281-7605-5/20/$31.

00 ©2021 IEEE | DOI: 10.1109/ICASSP39728.2021.9413777

DEEP LEARNING-BASED CROSS-LAYER RESOURCE ALLOCATION FOR WIRED

COMMUNICATION SYSTEMS

Pourya Behmandpoor Jeroen Verdyck Marc Moonen

KU Leuven, Department of Electrical Engineering (ESAT)

STADIUS Center for Dynamical Systems, Signal Processing and Data Analytics

ABSTRACT the instantaneous user demands are time-varying due to dif-

In this paper, a cross-layer resource allocation (RA) scheme ferent applications they use in different time slots, making it
based on deep learning is introduced for multi-tone multi-user necessary for the RA method to be adapted accordingly.
wired communication systems such as, for instance, digital The generic form of the optimization problem considered
subscriber line (DSL) systems under the current G.fast stan- here is written as,
X
dard. The upper layer reports time-varying user demands to max wi Ri (p)
the physical layer using proportional priority weights. Un- p (1)
like the deep neural network (DNN)-based RA schemes avail- s.t. p ∈ P
able in the literature for wireless communication, here the dy- where p is the vector of powers to be allocated, Ri is the ith
namic part is in adjusting the priority weights which are then user’s data rate, and P is the feasible set of the powers. In
fed to a DNN to map these priority weights to an appropriate the cross-layer model, this optimization problem is affected
power allocation, maximizing the weighted sum rate (WSR). by the upper layer through the priority weights wi which are
The DNN is trained employing an unsupervised strategy to dependent on the time-varying user demands. This generic
increase its generalization capabilities. Through the simula- RA problem (1) is generally non-convex. Furthermore, the
tions, we show that the proposed cross-layer scheme can scale dimension of the variable p in (1) can be very large, e.g. due
very well, even for real-world systems with thousands of RA to a large number of subcarriers or users, in the order of thou-
variables. Moreover, the proposed cross-layer scheme sig- sands in recent communication settings. In [1, 2, 3, 4] the RA
nificantly outperforms the conventional WMMSE method in problem is addressed in different wireless and wired commu-
terms of computation speed, while achieving the same data nication settings. Cross-layer RA is also studied in the litera-
rates. Hence, the proposed cross-layer scheme is more suit- ture in [5, 6, 7].
able for cross-layer RA than the conventional methods, as it Recently, efforts have been made to solve the above RA
increases the speed with which the physical layer can react to problem using machine learning approaches – specifically
time-varying upper layer data rate demands. deep neural networks (DNNs) – [8, 9, 10]. In [11, 12, 13,
Index Terms— Resource allocation, deep neural net- 14, 15] the machine learning (ML) approach has been used
work, cross-layer, wired communication for different applications in communication. These works all
consider wireless channels where it is important to solve the
1. INTRODUCTION RA problem fast due to the time-varying nature of the channel
conditions. Hence, we employ the same DNN-based algo-
Resource allocation (RA) plays a crucial role in today’s rithms for cross-layer RA in wired communication systems,
wired communication systems such as, for instance, digital hoping that the same speed improvements can be attained.
subscriber line (DSL) systems under the current G.fast stan- In the deep learning-based schemes available in the liter-
dard. Due to the rapidly increasing complexity of these wired ature, cross-layer RA is absent and hence conventional deep
communication systems, their associated RA problems are learning-based approaches need to be retrained when new
becoming more challenging to solve. Furthermore, although user demands are considered. Furthermore, the scalability of
the channels are long-term stationary in these wired settings, DNN has not been tested yet in the literature to investigate if
the performance is still acceptable when the number of RA
This research work was carried out at the ESAT Laboratory of KU Leu-
variables grows considerably. This paper, therefore, answers
ven, in the frame of Research Project FWO nr. G.0B1818N ’Real-time adap-
tive cross-layer dynamic spectrum management for fifth generation broad- the following questions:
band copper access networks’ and Fonds de la Recherche Scientifique -
FNRS and the Fonds Wetenschappelijk Onderzoek - Vlaanderen under EOS
1. Is a DNN capable of performing RA optimization
Project no 30452698 ’(MUSE-WINET) MUlti-SErvice WIreless NETwork’. when the dimensionality of the RA optimization vari-
The scientific responsibility is assumed by its authors. ables and hence the number of DNN outputs, i.e. the

978-1-7281-7605-5/21/$31.00 ©2021 IEEE 4120 ICASSP 2021

Authorized licensed use limited to: University of the West Indies (UWI). Downloaded on July 03,2023 at 20:31:45 UTC from IEEE Xplore. Restrictions apply.
 allocated power values, is vastly larger than the num-
ber of DNN inputs, i.e. the priority weights? Also,
a related question is addressed, namely if a DNN can
satisfy the constraints imposed by the vectoring. As
mentioned in [10], it is not straightforward for a DNN
to satisfy constraints, so the possibility of having new
constraints in wired communication systems should be
investigated for DNN based scenarios.
2. Can a DNN generalize a RA optimization when the ob-
jective function parameters are priority weights rather
than channel coefficients? While in wireless commu-
nication systems this generalization is not straightfor- Fig. 1. DNN-based cross-layer resource allocation scheme
ward, in wired communication systems according to
our simulations the problem can be generalized very
well for real-world systems by an applicable DNN. a group do not experience interference from other users in
the same group, but do experience interference from the other
groups.
2. SYSTEM MODEL The achieved data rate of user i on subcarrier k is then,
We consider a multi-channel wired communication system k
|hi,i |2 pki
with N users, K subcarriers, and input-output equation for Rik (pk ) = log2 (1 + ) (6)
k
σ 2 + j6=i |hi,j |2 pkj
P
subcarrier k as,

y k = H k xk + z k (2) where a user’s achieved total data rate is sum of data rates on
PK
the subcarriers i.e. Ri = k=1 Rik , and pk = [pk1 , ..., pkN ].
where the ith element of the vector y k is the received signal at Our objective is to allocate power between users and their
the receiver of user i. Also, the vector xk corresponds to the subcarriers such that the WSR is maximized as follows,
transmitted signals of the users with the power pki = E{|xki |2 }
for user i, and z k contains the noise signal at the receivers N
X K
X
with the same power σ 2 for all users. The channel gain be- max W SR(w, p) = wi Rik (pk )
[p1 ,p2 ,...,pK ]
tween ith user’s transmitter and jth user’s receiver is given by i=1 k=1
(7)
[H k ]ji = hkji . K X
X N
We also consider vectoring, in particular group vectoring, s.t. [F k ]ij pkj ≤ Pmax ∀i
such as used in DSL systems made the G.fast standard [16]. k=1 j=1
As a simple example, for instance, if there are two groups
k where wi is the priority weight corresponding to the ith user,
of
 co-located
k k
users, the channel is decomposed as H = Pmax is the maximum power constraint for each user’s total
H 11 H 12
where the channel block H kii corresponds to power, and p = [p1 , ..., pK ]. Also [F k ]ij is the element of the
H k21 H k22
vectoring matrix in ith row and jth column. In the considered
the ith group and the channel block H kij defines the inter- cross-layer RA model, the upper layer dynamically interacts
ference between the groups i and j. In this example of group with the physical layer to steer the corresponding RA prob-
vectoring, by multiplying the transmitted signal by the vector- lem, depending on the time-varying user demands. Each user
ing matrix F k , the received signal is then written as follows, is proportionally prioritized by the corresponding weight wi
in (7) according to its demand in the current time slot.
y k = H k F k xk + z k (3)
 k k
H k12
 k 
D 11 G11 F 11 0
= xk + z k (4) 3. RESOURCE ALLOCATION SCHEME WITH DNN
H k21 D k22 Gk22 0 F k22
k
= H xk + z k (5) Our scheme utilizes a fully connected DNN for the cross-
layer RA problem as in Fig.1. As in the wired communication
where D kii is a diagonal matrix with the diagonal elements setting, channels are long-term stationary and the dynamic
equal to the diagonal elements of channel block H kii . The part of the problem is the dynamic demands of the users, the
matrix Gkii is then chosen such that D kii Gkii = H kii . Also, the network is trained for a fixed channel realization. The priority
fixed matrix F kii is the vectoring matrix which can be inverse weights of the users are fed to the network as the input vec-
of the matrix Gkii to cancel out the interference. As not all tor. The output layer provides the allocated power for each
the users are co-located, with group vectoring, users within subchannel for each user.

4121

Authorized licensed use limited to: University of the West Indies (UWI). Downloaded on July 03,2023 at 20:31:45 UTC from IEEE Xplore. Restrictions apply.
 The network is trained in an unsupervised manner. We upper application layer. Some users may become idle in a pe-
perform unsupervised training as this approach does not need riod or their minimum demands may vary depending on the
labeling of the data, where labeling a large number of train- application they use.
ing samples for such a large problem is a very complex task.
If we denote the input vector consisting of the user priority
4. SIMULATION RESULTS
weights by w = [w1 , w2 , ..., wN ], and the network parame-
ters by θ, ideally we want to have a trained network with op- In this section, we present simulations to assess the proposed
timum parameters θ ∗ which is able to find the local solution cross-layer scheme. We assume N = 5 users each with
of the optimization problem (7) such that: K = 4096 subcarriers are present in a DSL system under the
p∗ = argmax W SR(w, p) = net(w; θ ∗ ) current G.fast standard [17]. The noise variable σ 2 is set to
(8) -140 dBm/Hz and the maximum power Pmax is set to -43.13
s.t. constraints of (7)
dBm/Hz. The subcarrier spacing is 51.75 kHz and also the
where WSR is in the objective function of the problem (7). symbol rate is set equal to 48 KHz. Channel measurements
We form the following loss function to be minimized during for five 104 m long TPs from the Dutch telecom operator KPN
the DNN training to address the power allocation problem, are used from [18]. Group vectoring is applied for two groups
of three and two users.
M N K
1 X X X k The number of neurons per layer in our DNN is {N, 128,
loss(w; θ) = {− wi Ri (net(w; θ))
M N K, N K}. The relu activation function is used for hidden
i=1l=1 k=1
N K X
N
(9) layers. The regularization parameter λ is set to 50. To train
X X k the network, user priority weights, as the training samples, are
+λ relu( [F ]ij net(w; θ)kj − Pmax )}
generated uniformly and independently for each of the users.
i=1 k=1 j=1
These samples are generated online so that during the itera-
where M is the batch size, and net(w; θ)kj is the output of tions a training sample is not seen by the DNN twice. Online
the network corresponding to kth subcarrier and jth user. training is beneficial for the unsupervised learning as it avoids
Moreover, λ is the penalization parameter by which the net- overfitting. For evaluation purposes, 10000 samples are gen-
work is steered to satisfy the constraints while minimizing erated to adjust the hyper-parameters of the network during
the objective function. The relu-function, which is defined as the training, and 1000 samples are also generated for testing
relu(.) = max{0, .}, penalizes the loss function when the the network. The optimizer is Adam [19] with the learning
total power of a user exceeds the maximum allowed power rate of 0.0001 and appropriate reductions accordingly as the
Pmax . training continues. The simulations are done in Python with
The same approach of penalizing the constraints is also the Pytorch library on a system with 8GB memory, 2.8GHz
applied in [10, 15], with a difficulty in satisfying the rate con- CPU, and one Tesla K40 GPU.
straints, however, the constraints here are fixed affine func- Also, a supervised training phase is used as a warm start
tions of the RA variables so can be learned by the network by for the unsupervised phase. This supervised phase is use-
the penalization method. Moreover, the proposed setting does ful as it can steer the network in the right direction from
not need the feasibility check of the training samples, due to which the unsupervised training can be effective. For this
the non-empty feasible set of (7), hence no pre-processing for phase, 500 samples are considered and the labeling is done
the samples is needed. by the WMMSE [1] method. To avoid overfitting, the dropout
To improve the generalization capability of the Pnetwork, method [20] with an appropriate dropout factor is used. The
the user priority weights are constrained to satisfy i wi = 1 learning rate is also dynamically adjusted to decrease the
without loss of generality. This constraint limits the input mean squared error (MSE) loss function appropriately. More-
space to be learned by the network so the network can gen- over, from the simulations, the average power violation is
eralize the problem with a lower number of training samples. only 0.76 percent of each user’s power budget Pmax , indicat-
Furthermore, the mentioned constraint does not affect the op- ing the network can meet the constraints very well. However,
timization solution as for arbitrary weights the constraint can for real implementations a simple scaler can be applied after
be satisfied by scaling the weights with a number, which is the network’s output. The scaler simply multiplies the net-
equivalent to the scaling of the objective function in (7). work’s output by a scalar to adjust the powers to ensure that
It should be mentioned that in the literature the power al- the power constraints are met. As the user powers are related
location is done without any feedback from the upper appli- to each other due to the vectoring matrix, a same scalar should
cation layer. The user demands are considered fixed for the be applied to all user powers.
different users and the network needs to be retrained if these In the first set of simulations, we assess the performance
user demands are changed. However, the proposed approach of the proposed method in terms of the achieved WSR. The
here provides the flexibility for the physical layer to be ad- assessment is done in Table 1 for three different configura-
justed to the real-time demands of each user according to their tions of the communication network, and for the DNN with

4122

Authorized licensed use limited to: University of the West Indies (UWI). Downloaded on July 03,2023 at 20:31:45 UTC from IEEE Xplore. Restrictions apply.
 Deep Neural Network
3000
Table 1. Performance comparisons of WMMSE and the pro- user 1

Rate (Mb/sec)
user 2
posed cross-layer scheme in resource allocation over 1000 2000 user 3
user 4
test samples. 1000
user 5
WSR
WSR*
Weighted Sum Data Rates (Mbps) 0
0 5 10 15 20 25 30 35 40 45 50
User-Tone WMMSE DNN DNN+Scaler
R DNN - R WMMSE
5 - 1024 591.14 590.86 590.60 4
user 1
5 - 2048 852.24 851.61 851.11

Rate (Mb/sec)
user 2
2 user 3
5 - 4096 1061.24 1060.74 1059.77 user 4
user 5
0
WSR

-2
0 5 10 15 20 25 30 35 40 45 50
Table 2. Computation speed comparisons of WMMSE and
sample
the proposed cross-layer scheme in resource allocation over
1000 test samples. Fig. 2. User rate comparison, for 5 users and 4096 subcarriers
Average Processing Time (Second) for each user, drawn for 50 test samples. The WSR of the
User-Tone WMMSE DNN DNN trained without warm-starting is also shown by W SR∗
5 - 1024 1.94 0.006 .
5 - 2048 2.75 0.025 20
Deep Neural Network

5 - 4096 3.11 0.032 Bitloading (bit)

user 1
15 user 2
user 3
10 user 4
user 5
5
and without the scaler. Moreover, the processing times of
0
the two methods for each sample are compared in Table 2, 0 500 1000 1500 2000 2500 3000 3500 4000
indicating that the proposed cross-layer scheme can do the
R DNN - R WMMSE
power allocation much faster than the conventional WMMSE 2
Bitloading (bit)

user 1
method while achieving the same data rates. Although it is 1 user 2
not a straightforward task to compare the computational com- user 3
0 user 4
plexity of the methods theoretically, this increase in speed can user 5
-1
be justified by the opportunity of parallel processing in DNN,
compared to iterative processing in WMMSE. For the results -2
0 500 1000 1500 2000 2500 3000 3500 4000
reported in Table 2, the DNN is run on a CPU. Using a GPU Frequency Channel
will even further increase the speed of the network.
In the second set of simulations, we compare the proposed Fig. 3. Bitloading comparison, for 5 users and 4096 subcarri-
cross-layer scheme with the WMMSE method. It should be ers for each user.
mentioned that we do not expect the solutions to be exactly
equal as, with the unsupervised training phase, the network 5. CONCLUSIONS
may end up in a different local optimum. Nonetheless, it is
worth investing to what extent the solutions are similar. We have studied a deep learning-based cross-layer RA in a
In Fig.2, we can see while the WSR achieved by the DNN multi-tone multi-user wired communication system under the
is close to that of WMMSE method, the achieved individual current DSL G.fast standard. A DNN considers real-time dy-
user rates are larger for the proposed cross-layer scheme. The namic user demands as proportional priority weights reported
zero data rates in the figure correspond to users which are by the upper layer. Similar performance with WMMSE
inactive or idle in that specific time slot so other users can method can be achieved in a much faster way. Hence, the
benefit from this by having increased data rates. This figure upper layer can dynamically report the user demands to the
shows the proposed DNN-based cross-layer RA method can physical layer, guaranteeing more efficient deployment of the
have good generalization capability for unseen samples. limited resources amongst the users and maximized WSR.
In Fig.3, we plot the data rate on each subcarrier called The DNN speed gain over the conventional methods can al-
bitloading of the users for one sample of the testing set to low the time slots to be reduced by a factor of 100, so that
compare with WMMSE method. As shown in the Fig. 3, studying the throughput stability and exploring the dynamics
the data rates are close to those obtained with the WMMSE of the arriving traffic and packet queues in the context of
method, and also correspond to their user priority weights as cross-layer resource allocation is made possible and hence
user 5 is idle in this sample. will be a direction for future studies.

4123

Authorized licensed use limited to: University of the West Indies (UWI). Downloaded on July 03,2023 at 20:31:45 UTC from IEEE Xplore. Restrictions apply.
 6. REFERENCES [11] Hao Huang, Wenchao Xia, Jian Xiong, Jie Yang, Gan
Zheng, and Xiaomei Zhu, “Unsupervised learning-
[1] Qingjiang Shi, Meisam Razaviyayn, Zhi-Quan Luo, and based fast beamforming design for downlink mimo,”
Chen He, “An iteratively weighted mmse approach to IEEE Access, vol. 7, pp. 7599–7605, 2018.
distributed sum-utility maximization for a mimo inter-
fering broadcast channel,” IEEE Transactions on Signal [12] Zhiyuan Xu, Yanzhi Wang, Jian Tang, Jing Wang, and
Processing, vol. 59, no. 9, pp. 4331–4340, 2011. Mustafa Cenk Gursoy, “A deep reinforcement learning
based framework for power-efficient resource allocation
[2] Mung Chiang, Chee Wei Tan, Daniel P Palomar, Daniel in cloud rans,” in 2017 IEEE International Conference
O’neill, and David Julian, “Power control by geometric on Communications (ICC). IEEE, 2017, pp. 1–6.
programming,” IEEE transactions on wireless commu-
nications, vol. 6, no. 7, pp. 2640–2651, 2007. [13] Woongsup Lee, “Resource allocation for multi-channel
underlay cognitive radio network based on deep neural
[3] Wei Yu, George Ginis, and John M Cioffi, “Distributed
network,” IEEE Communications Letters, vol. 22, no. 9,
multiuser power control for digital subscriber lines,”
pp. 1942–1945, 2018.
IEEE Journal on Selected areas in Communications,
vol. 20, no. 5, pp. 1105–1115, 2002. [14] Yasar Sinan Nasir and Dongning Guo, “Multi-agent
[4] Jeroen Verdyck, Wouter Lanneer, Paschalis Tsiaflakis, deep reinforcement learning for dynamic power alloca-
Werner Coomans, Panagiotis Patrinos, and Marc Moo- tion in wireless networks,” IEEE Journal on Selected
nen, “Optimal dynamic spectrum management algo- Areas in Communications, vol. 37, no. 10, pp. 2239–
rithms for multi-user full-duplex dsl,” IEEE Access, vol. 2250, 2019.
7, pp. 106600–106616, 2019. [15] Hao Ye, Geoffrey Ye Li, and Biing-Hwang Fred Juang,
[5] Jeremy Van den Eynde, Jeroen Verdyck, Marc Moonen, “Deep reinforcement learning based resource allocation
and Chris Blondia, “A delay-based cross-layer sched- for v2v communications,” IEEE Transactions on Vehic-
uler for adaptive dsl,” in 2017 IEEE International Con- ular Technology, vol. 68, no. 4, pp. 3163–3173, 2019.
ference on Communications (ICC). IEEE, 2017, pp. 1–6.
[16] Jeroen Verdyck, Chris Blondia, and Marc Moonen,
[6] Paschalis Tsiaflakis, Yung Yi, Mung Chiang, and Marc “Network utility maximization for adaptive resource al-
Moonen, “Throughput and delay performance of dsl location in dsl systems,” in 2018 26th European Signal
broadband access with cross-layer dynamic spectrum Processing Conference (EUSIPCO). IEEE, 2018, pp.
management,” IEEE transactions on communications, 787–791.
vol. 60, no. 9, pp. 2700–2711, 2012.
[17] ITU-T Study Group 15, “Fast access to subscriber ter-
[7] Mahdi Ben Ghorbel, Bechir Hamdaoui, Mohsen minals (G.fast) – Physical layer specification,” Tech.
Guizani, and Bassem Khalfi, “Distributed learning- Rep. ITU-T G.9701, International Telecommunication
based cross-layer technique for energy-efficient multi- Union, 2015.
carrier dynamic spectrum access with adaptive power
allocation,” IEEE Transactions on Wireless Communi- [18] TNO, “G.fast: Release of measured transfer
cations, vol. 15, no. 3, pp. 1665–1674, 2015. characteristics of the 104m KPN access cable,”
https://www.joepeesoft.com/Public/
[8] Fuhui Zhou, Guanyue Lu, Miaowen Wen, Ying-Chang DSL_Corner/_Index.html, 2013.
Liang, Zheng Chu, and Yuhao Wang, “Dynamic spec-
trum management via machine learning: state of the art, [19] Diederik P Kingma and Jimmy Ba, “Adam: A
taxonomy, challenges, and open research issues,” IEEE method for stochastic optimization,” arXiv preprint
Network, vol. 33, no. 4, pp. 54–62, 2019. arXiv:1412.6980, 2014.
[9] Haoran Sun, Xiangyi Chen, Qingjiang Shi, Mingyi [20] Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky,
Hong, Xiao Fu, and Nikos D Sidiropoulos, “Learning to Ilya Sutskever, and Ruslan R Salakhutdinov, “Improv-
optimize: Training deep neural networks for wireless re- ing neural networks by preventing co-adaptation of fea-
source management,” in 2017 IEEE 18th International ture detectors,” arXiv preprint arXiv:1207.0580, 2012.
Workshop on Signal Processing Advances in Wireless
Communications (SPAWC). IEEE, 2017, pp. 1–6.
[10] Fei Liang, Cong Shen, Wei Yu, and Feng Wu, “Towards
optimal power control via ensembling deep neural net-
works,” IEEE Transactions on Communications, vol.
68, no. 3, pp. 1760–1776, 2019.

4124

Authorized licensed use limited to: University of the West Indies (UWI). Downloaded on July 03,2023 at 20:31:45 UTC from IEEE Xplore. Restrictions apply.