Centralized and Distributed Power Allocation for

Max-Min Fairness in Cell-Free Massive MIMO

Sucharita Chakraborty∗ , Emil Björnson∗ , and Luca Sanguinetti†
∗ Department
of Electrical Engineering (ISY), Linköping University,
Linköping, Sweden ({sucharita.chakraborty, emil.bjornson}@liu.se),
† Dipartimento dellâInformazione, University of Pisa, 56122 Pisa, Italy ([email protected])

Abstract—Cell-free Massive MIMO systems consist of a large dense networks, where the number of APs grows large,
number of geographically distributed access points (APs) that significant communication overhead will be incurred in
serve users by coherent joint transmission. Downlink power acquiring and disseminating CSI to all of the cooperative
allocation is important in these systems, to determine which
APs should transmit to which users and with what power. If the APs or to an edge cloud computer.
system is implemented correctly, it can deliver a more uniform 2) An enormous amount of information regarding the payload
user performance than conventional cellular networks. To this data (e.g., coding/modulation scheme, decoding errors)
end, previous works have shown how to perform system-wide must be monitored to guarantee user satisfaction. Immense
max-min fairness power allocation when using maximum ratio buffer sizes and fronthaul signaling capacity are required to
precoding. In this paper, we first generalize this method to
arbitrary precoding, and then train a neural network to perform store and share such global information over the network.
approximately the same power allocation but with reduced 3) The time delay and complexity incurred during channel
computational complexity. Finally, we train one neural network estimation, precoding/combining, and fronthaul signaling
per AP to mimic system-wide max-min fairness power allocation, increase at least linearly with the number of APs.
but using only local information. By learning the structure of
Most of these issues were overlooked in the early works
the local propagation environment, this method outperforms the
state-of-the-art distributed power allocation method from the on Cell-free mMIMO [2], [4]. However, the recent works
Cell-free Massive MIMO literature. [6], [7] have indicated that a scalable implementation might
Index Terms—Cell-free Massive MIMO, Power allocation, be realizable in practice. This paper focuses on the scalable
Max-min fairness, Deep learning, Scalability. implementation of downlink power allocation algorithms.
I. I NTRODUCTION Network-wide downlink power allocation algorithms for
Cell-free mMIMO systems were developed in [4], [8] for the
Coordinated distributed wireless systems liberate the con-
purpose of achieving max-min fairness; that is, all user equip-
ventional co-located multiple-input multiple-output (MIMO)
ments (UEs) get the same spectral efficiency (SE) and that
from its shackles of inherent form-factor constraint [1]. If
common value is maximized. The results apply to maximum
an arbitrarily large number of collaborative access points
ratio (MR) precoding with long-term power constraints, which
(APs) jointly serve the users in a wide area, it constitutes
are undesirable assumptions since the use of regularized zero-
a dense large network without any cell boundaries. This type
forcing (RZF) precoding at every AP gives higher SE [6] and
of systems is gaining popularity with the name of Cell-free
real systems are subject to short-term power constraints [9].
massive MIMO (mMIMO) systems [2]–[4]. It reaps many Even in the case when these algorithms are applicable, the
of the advantages of two cornerstone technologies: mMIMO deployment feasibility is limited since global CSI must be
(e.g., favorable propagation) and Network MIMO (e.g., more available at a central processing unit (CPU) and the compu-
uniform user performance) by exploiting coherent signal co- tational complexity grows polynomially with the number of
processing among multiple distributed APs. APs and UEs. The first of these issues can be addressed by
The main challenge in bringing such a network to reality utilizing the dynamic cooperation cluster (DCC) concept [5],
is scalability, in terms of computational complexity for signal [6], in which each user is assumed to be served by a user-
processing and resource allocation, fronthaul requirements, centrically selected subset of the APs with the best channel
etc. General guidelines for Network MIMO were provided conditions. In this paper, we first develop an optimal power
in [5] and later particularized for Cell-free mMIMO in [6], allocation algorithm for max-min fairness in DCC-based Cell-
[7]. Existing algorithms in Network MIMO, or coordinated free mMIMO systems, but the complexity is unscalable. We
multipoint (CoMP), are mainly limited by three factors [6]: then utilize the learn to optimize framework [10], [11] for
1) Dependency on the availability of full channel state in- offline training of deep neural networks (DNNs) that perform
formation (CSI) in the network, or at least partial CSI approximated max-min fairness power allocation.
that is shared between neighboring APs. In future ultra More precisely, we train one DNN to perform centralized
power allocation with reduced computational complexity. We
This work was partially supported by the Wallenberg AI, Autonomous
Systems and Software Program (WASP) funded by the Knut and Alice also train one DNN per AP to perform distributed power
Wallenberg Foundation. allocation using only locally available information as input,

978-1-7281-4300-2/19/$31.00 ©2019 IEEE 576 Asilomar 2019

Authorized licensed use limited to: ULAKBIM UASL - MIDDLE EAST TECHNICAL UNIVERSITY. Downloaded on May 28,2023 at 18:47:14 UTC from IEEE Xplore. Restrictions apply.
 where Φkl = E{yptl (yptl ) } =
H P 2
but training it using the globally optimal max-min fairness i∈Pt τp pi Ril + σ IN
solution. Hence, different from the heuristic power allocation denotes the correlation matrix of the pilot signal.
method that was recently proposed in [7], each AP is utilizing
B. DCC Framework Based Cell-free mMIMO
a unique algorithm that can take the actual network structure
and propagation environment into account. We assume that each AP serves a subset of the UEs and
Notations: x, x, X denote a scalar, vector, and matrix, we use the DCC framework [5], [6]. We let Dl ⊂ {1, . . . , K}
respectively. (·)T , (·)H , | · |, k · k, IX stand for transpose, denote the UEs served by the lth AP. In accordance to [5],
Hermitian transpose, absolute value, the L2 vector norm, and we then define the matrices Dkl ∈ CN ×N , for l = 1, . . . , L
X × X identity matrix, respectively. The multivariate circular and k = 1, . . . , K, as
symmetric complex Gaussian distribution with correlation
(
IN for k ∈ Dl ,
matrix R is denoted NC (0, R). Dkl = (3)
0N for k ∈
/ Dl .
II. S YSTEM M ODEL Notice that this matrix is IN if the kth UE is served by the
We consider a Cell-free mMIMO system with K single- lth AP and 0N otherwise. The received downlink signal at
antenna UEs and L APs, each equipped with N antennas. We the kth UE reads as
assume a block fading channel model where the channels are L
X X√
static within time-frequency coherence blocks with τc channel ykdl = hHkl ρil wil si + nk (4)
uses, and independent random channel realizations appear in l=1 i∈Dl
each block. The channel between the kth UE and the lth AP L K
X X √
is denoted as hkl ∈ CN ×1 and is modeled by correlated = hHkl ρil Dil wil si + nk (5)
Rayleigh fading hkl ∼ NC (0, Rkl ), where Rkl ∈ CN ×N l=1 i=1
is the spatial correlation matrix. The normalized trace βkl = where ρil is the downlink power allocated to UE i by AP l
1/N tr(Rkl ) accounts for the average channel gain from an and wil ∈ CN ×1 is the corresponding normalized precoding
antenna at AP l to UE k. vector with kwil k2 = 1. Moreover, si denotes the signal
The APs are connected to a CPU via fronthaul connections, transmitted to UE i and nk ∼ NC (0, σ 2 ) is the receiver noise.
which are used to convey uplink and downlink data between The benefit of using the matrix notation in (5), instead of
the APs and the CPU. The connections are assumed to be the set notation in (4), is that sizes of all matrices and vectors
error-free but the capacity is limited, thus each UE can only become independent of which APs serve which UEs. This will
be served by a subset of the APs. No instantaneous CSI is be convenient in Section III. The original Cell-free mMIMO
conveyed over the fronthaul. model in [2], [4] is obtained by setting Dl = {1, . . . , K} and
thus Dkl = IN for all l and k. Practical methods to select the
A. Channel Estimation
sets D1 , . . . , DL are found in [5], [6].
We consider a time division duplex (TDD) protocol having
a pilot transmission phase for channel estimation and a data C. Downlink SE
transmission phase. Following the standard TDD protocol [1], The downlink SE of Cell-free mMIMO was characterized
the coherence block is divided in three parts: τp for uplink for MR precoding with long-term power constraints in [2], [4].
pilot transmission, τu for uplink data transmission, and τd for This enabled the development of max-min power allocation
downlink data. It thus follows that τc = τp + τu + τd . optimization. The SE with arbitrary precoding schemes was
In the uplink pilot phase, a set of τp mutually orthogonal τp - considered in [6], but the expression was not amendable for
length pilots are utilized. Each UE is assigned to one of these power optimization. The following lemma provides a new
pilots. Let us denote the subset of UEs that are assigned to simplified expression for the general case, which will enable
pilot t as Pt ⊂ {1, . . . , K}. The received signal yptl ∈ CN ×1 the power optimization in Section III.
at AP l when the UEs in Pt transmit is defined as
X√ Lemma 1. An achievable downlink SE for UE k is
yptl = τp pi hil + ntl (1) τd  
SEdl
k = log2 1 + SINRdlk (6)
i∈Pt τc
where pi is the transmit power from the ith UE and ntl ∼ where
NC (0, σ 2 IN ) is the additive white Gaussian noise (AWGN)
SINRdl
k =
with variance σ 2 . By utilizing the standard minimum mean
L √ 2
square error (MMSE) estimator at the lth AP, the estimate of
ρkl E{hHkl Dkl wkl }
P
the channel hkl from UE k ∈ Pt is [3] l=1
!−1 L P
K n L
o P
2 2
ρil E |hHkl Dil wil | − ρkl |E{hHkl Dkl wkl }| +σ 2
P
√ X
ĥkl = τp pk Rkl 2
τp pi Ril + σ IN yptl l=1 i=1 l=1
i∈Pt (7)
NC 0, τp pk Rkl Φ−1


∼ kl Rkl (2) is the effective signal-to-interference-and-noise ratio (SINR).

577

Authorized licensed use limited to: ULAKBIM UASL - MIDDLE EAST TECHNICAL UNIVERSITY. Downloaded on May 28,2023 at 18:47:14 UTC from IEEE Xplore. Restrictions apply.
 The SE expression in Lemma 1 can be utilized along with The first constraint in (12) can be rewritten as
any precoding scheme and correlated Rayleigh fading model.1 L 2 L X
K
We consider precoding vectors {wil } satisfying short-term 1 X X
µkl akl ≥ µ2il bkil + σ 2 . (13)
power constraints, which means that kwil k2 = 1 must be s
l=1 l=1 i=1
satisfied in every coherence block and not on average (i.e.,
E{kwil k2 } = 1) as in [2], [4]. This is the conventional After taking the square root on both sides and noting that one
approach to precoding normalization [5]. We notice that can always rotate the phase of precoders to make alk ≥ 0,
the relaxed long-term average power constraints are popular (13) is rewritten as
v
in Massive MIMO because they lead to closed-form SINR r L u L K
1 X uX X
expressions. The relaxation is rather tight in Massive MIMO µkl |akl | ≥ t µ2il bkil + σ 2 (14)
since the channel hardening makes kwil k2 ≈ E{kwil k2 }. s i=1
l=1 l=1
However, the same relaxation should not be used in Cell-free
and, equivalently, in vector form as
mMIMO since the channel between an AP with few antennas r  
and a UE does not harden (although the joint channel from 1 T µ
c µ ≥ Bk (15)
all APs might harden in some cases). s k k σ
An arbitrary normalized precoding vector is defined as T
wkl = w̄kl /kw̄kl k where w̄kl can be arbitrarily selected. where ck = [|ak1 | . . . |akL |] ∈√ CL×1 , √and µk
=
T L×1
In this paper, we consider MR and RZF precoding, which are [µk1 . . . µkL ] ∈ C . Bk = diag bk11 . . . bkKL 1 ∈
T
defined as C(KL+1)×(KL+1) and µ = [µT1 . . . µTK ] ∈ CKL×1 .
 Hence, (12) can be equivalently written as
ĥkl

!−1
for MR,
w̄kl = (8) maximize s (16)
H {µkl :∀k,l},s
pi ĥil ĥil + σ 2 IN
P

 pk ĥkl for RZF. r  
i∈Dl 1 T µ
subject to ck µk ≥ Bk √ 2 , k = 1, . . . , K,
III. D OWNLINK M AX -M IN P OWER A LLOCATION s σ
K
In this section, we generalize the max-min fairness algo- X
rithm from [2], [4] to general precoding schemes and corre- µ2il ≤ Pmax
dl
, l = 1, . . . , L.
lated Rayleigh fading. Hence, the goal is to find the optimal i=1

power allocation coefficients {ρkl : ∀k, l} that maximize the This is still a non-convex problem but we notice that if s is
lowest SE among all UEs in the network. Hence, we select constant, the SINR constraint in (15) becomes a second-order
the coefficients to give all UEs the same effective SINR and cone (SOC) constraint. Hence, for a given s, the problem in
this value is to be maximized, under the constraint that each (16) becomes a second-order cone program (SOCP), which
dl can be solved through the bisection method [5] by considering
AP has the same maximum power PK Pmax . Thisdl
means that the
power constraint at AP l is k=1 ρkl ≤ Pmax . a sequence of s that converges to the global optimum. In each
Before formulating the max-min fairness optimization subproblem, the following problem must be solved:
problem, we first rewrite (7) as
maximize c (17)
L 2 {µkl :∀k,l},c
P
µkl akl
r  
1 T õ
SINRdl = l=1
(9) subject to c µ ≥ Bk , k = 1, . . . , K
k L P
K S k k σ2
µ2il bkil + σ 2
P
K
X
l=1 i=1
√ µ2il ≤ cPmax
dl
, l = 1, . . . , L
by introducing the new variables µkl = ρkl and i=1

akl = E{hHkl Dkl wkl } (10) and verified to be feasible and have c ≤ 1 at the solution. Note
that the power scaling variable c is introduced to improve the
(
n
H 2
o 0 for i 6= k,
bkil = E |hkl Dil wil | − 2 (11) algorithm convergence, as suggested in [5]. If not included,
|akl | for i = k. the bisection algorithm requires an extremely high number of
The max-min fairness optimization problem is then expressed iterations (accuracy) to find the max-min solution.
in epi-graph form as
IV. N EURAL N ETWORK BASED P OWER A LLOCATION
maximize s (12) The central goal of this paper is to demonstrate that large-
{µkl :∀k,l},s
scale fading information is sufficient for computing the opti-
subject to SINRdl k ≥ s, k = 1, . . . , K,
mal powers. This is in contrast to the traditional optimization
K
X approach for solving (17) that requires knowledge of {akl }
µ2il ≤ Pmax
dl
, l = 1, . . . , L. and {bkil }. We advocate using the UEs’ large-scale fading
i=1
coefficients {βkl } to perform power allocation because they
1 Actually, it holds for any arbitrary independent fading distribution. already capture the main feature of propagation channels and

578

Authorized licensed use limited to: ULAKBIM UASL - MIDDLE EAST TECHNICAL UNIVERSITY. Downloaded on May 28,2023 at 18:47:14 UTC from IEEE Xplore. Restrictions apply.
 TABLE I: Simulation parameters of the Cell-free mMIMO network

Cell area (wrap around) 150 m2

140 APs Locations Bandwidth 20 MHz
120 UEs Locations Number of APs L=9
Number of UEs K=5
Number of antennas per AP M =2
100 Pathloss exponent (α) α = 3.76
Maximum downlink transmit power per AP dl
Pmax =1W
80
Y

UL noise power −94 dBm

UL transmit power pi = 100 mW
60 Length of coherence block τc = 200
40
TABLE II: Layout of centralized DNN for whole network with L = 9 and
20 K = 5. Number of trainable parameters: 241965.
0 Size Parameters Activation Function
0 20 40 60 80 100 120 140
X
Input KL -
Layer 1 (Dense) 128 5888 elu
Fig. 1: In the numerical evaluation, L = 9 APs with fixed locations are Layer 2 (Dense) 512 66048 elu
considered as shown in this figure. In every user drop, K = 5 UEs are Layer 3 (Dense) 256 131328 sigmoid
dropped randomly in the area. Layer 4 (Dense) 128 32896 sigmoid
Layer 5 (Dense) KL 5805 relu
interference in the network, and can be easily measured in
practice based on the received signal strength. Therefore, for TABLE III: Layout of decentralized DNN for an AP with L = 9 and K = 5.
a given location of APs, the problem is to learn the unknown Number of trainable parameters: 3877.
mapping between {βkl } and the optimal square-roots of the Size Parameters Activation Function
transmit powers {µ∗kl }. This is achieved by leveraging that Input K -
DNNs are universal function approximators [10], [12]. Layer 1 (Dense) 16 96 elu
Layer 2 (Dense) 64 1088 elu
We use a fully connected feed-forward DNN with N hidden Layer 3 (Dense) 32 2080 sigmoid
layers. The output layer provides an estimate {µ̂kl } of {µ∗kl }. Layer 4 (Dense) 16 528 sigmoid
The problem is thus to train effectively the weights and Layer 5 (Dense) K 85 relu
biases of the DNN so that it can learn {µ∗kl }. We consider
two different DNNs with both MR and RZF precoding.
The first is the so-called centralized DNN that receives as consider a Cell-free mMIMO network with L = 9 APs at
input the entire large-scale fading coefficients {βkl : ∀k, l} fixed locations in a square of 150 m ×150 m, as illustrated in
and provides as output the network-wide square-roots of the Fig. 1. The large-scale fading coeffcients are generated as [3]
powers {µ̂kl : ∀k, l}. The second DNN is called decentralized
 
dkl
because it operates on a per-AP basis. Specifically, the DNN βkl = −30.5 − 36.7 log10 dB (18)
1m
of AP l receives as input only the locally available coefficients
{βkl : ∀k} and aims to learn the local estimate {µ̂kl : ∀k} where dkl is the distance of UE k from AP l. In each
of optimal powers. The advantage of the decentralized DNN realization of the network, K = 5 UEs are independently
is that no exchange of large-scale fading coefficients among and random uniformly distributed in the area. The APs are
APs is required, which is important for a scalable network deployed 10m above the UEs. All other simulation parameters
operation [6]. Moreover, the number of trainable parameters are reported in Table I. The length of the pilot sequences is
per AP largely reduces. We stress that each AP has a unique τp = K, so orthogonal pilots are allocated to the UEs. We
DNN that captures features of the local propagation environ- consider the downlink with MR or RZF precoding.
ment and where the AP is located compared to other APs. The centralized and decentralized DNNs used with both
The complexity of the approach above is mainly the gener- precoding schemes are reported in Tables II and III, whose
ation of the training dataset. Suppose each layer of a DNN has trainable parameters are 241965 and 3877, respectively. The
Ni neurons. The number of multiplications and addition for DNNs were trained based on a dataset of Ns = 249900
each layer is Ni Ni−1 and 2Ni , i = 1, . . . , N , respectively. samples of independent realizations of the user locations,
Each layer needs to evaluate Ni activation functions. Once corresponding to large-scale fading coefficients {βkl (n) : n =
trained, the DNN allocates the transmit power very rapidly 1, . . . , Ns } and the corresponding max-min power allocation
without actually solving (17). In practice, such decisions need variables {µ?kl (n) : n = 1, . . . , Ns } for any given pair k and l.
to be made when the large-scale fading coefficients change Particularly, 90% percent of the samples was used for training
due to large-scale movements, the addition of new UEs, or and 10% for validation. The remaining 100 samples formed
when UEs go from active to inactive mode. the test dataset, which is independent from the training dataset
and is used to generate the simulation results presented in this
V. N UMERICAL E VALUATION section. We used the Adam optimizer and categorical cross-
To demonstrate the ability to learn how to perform power entropy as loss function. The number of epochs, batch size
allocation based on only large-scale fading coefficients, we and learning rate are optimized by a trial-and-error method.

579

Authorized licensed use limited to: ULAKBIM UASL - MIDDLE EAST TECHNICAL UNIVERSITY. Downloaded on May 28,2023 at 18:47:14 UTC from IEEE Xplore. Restrictions apply.
 1.0 1.0
RZF (equal power)
RZF (max-min)
0.8 0.8 RZF [7]
RZF (cent. DNN)
RZF (dist. DNN)
0.6 0.6
CDF

CDF
0.4 0.4
MR (equal power)
MR (max-min)
0.2 MR [7] 0.2
MR (cent. DNN)
MR (dist. DNN)
0.0 0.0
0 1 2 3 4 5 0 1 2 3 4 5
SE per UE [bit/s/Hz] SE per UE [bit/s/Hz]
Fig. 2: CDF of DL SE per UE with MR. Fig. 3: CDF of DL SE per UE with RZF.

The performance of both DNNs is evaluated by computing cess. Also, we showed that a decentralized DNN at each AP
the cumulative distribution function (CDF) of the downlink can allocate power more effectively as compared to previous
SE per user, where the randomness is due to the UE locations. heuristic methods. However, there is still improvements to be
Comparisons are made with the performance achieved by the made since the gap between the decentralized and centralized
max-min optimization algorithm developed herein and the methods is rather large.
heuristic power allocation recently proposed in [7], where
R EFERENCES
each AP uses full power and allocates power to UE propor-
tionally to the square root of the channel estimation variance. [1] G. Interdonato, E. Björnson, H. Q. Ngo, P. Frenger, and E. G. Lars-
son, “Ubiquitous cell-free massive MIMO communications,” EURASIP
Fig. 2 shows the CDF with MR precoding. The centralized Journal on Wireless Communications and Networking, no. 197, 2019.
DNN closely follows the performance achieved with the max- [2] H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta,
min algorithm. The distributed DNN outperforms [7] for 80% “Cell-free massive MIMO versus small cells,” IEEE Trans. on Wireless
Communications, vol. 16, no. 3, pp. 1834–1850, 2017.
of the UEs, but lower SE for the 20% most unfortunate users. [3] E. Björnson and L. Sanguinetti, “Making cell-free massive MIMO
A possible explanation for the performance improvement competitive with MMSE processing and centralized implementation,”
using the DNN based power allocation is that it learns the IEEE Trans. on Wireless Commun., 2019.
[4] E. Nayebi, A. Ashikhmin, T. L. Marzetta, H. Yang, and B. D. Rao, “Pre-
propagation environment of the network so that every AP can coding and power optimization in cell-free massive MIMO systems,”
apply a locally optimized power allocation policy. However, IEEE Trans. on Wireless Commun., vol. 16, no. 7, pp. 4445–4459, 2017.
there is still a substantial gap between the distributed and [5] E. Björnson and E. Jorswieck, “Optimal resource allocation in coordi-
nated multi-cell systems,” Foundations and Trends R in Communica-
centralized methods. tions and Information Theory, vol. 9, no. 2–3, pp. 113–381, 2013.
Fig. 3 shows the CDF of downlink SE per user with RZF [6] E. Björnson and L. Sanguinetti, “Scalable cell-free massive MIMO
precoding. For 40% users, the centralized DNN power allo- systems,” arXiv preprint arXiv:1908.03119, 2019.
[7] G. Interdonato, P. Frenger, and E. G. Larsson, “Scalability Aspects of
cation closely approximate the conventional max-min fairness Cell-Free Massive MIMO,” in IEEE ICC, 2019, pp. 1–6.
algorithm without having to actually solve (17). It achieves [8] E. Nayebi, A. Ashikhmin, T. L. Marzetta, and B. D. Rao, “Performance
42% higher average SE than that of [7]. In addition, local of cell-free massive MIMO systems with MMSE and LSFD receivers,”
in IEEE Asilomar Conf. on Signals, Systems and Computers, 2016, pp.
DNN at each AP achieves 4% higher SE compared to [7]. 203–207.
[9] G. Interdonato, H. Q. Ngo, E. G. Larsson, and P. Frenger, “On the per-
VI. C ONCLUSIONS formance of cell-free massive mimo with short-term power constraints,”
in IEEE International Workshop on Computer Aided Modelling and
In this paper, we proposed a deep learning framework for Design of Communication Links and Networks (CAMAD), Oct 2016,
downlink power allocation in a Cell-free mMIMO network pp. 225–230.
[10] H. Sun, X. Chen, Q. Shi, M. Hong, X. Fu, and N. D. Sidiropoulos,
with MR and RZF precoding and short-term power con- “Learning to optimize: Training deep neural networks for interference
straints. We developed the optimal power allocation strategy management,” IEEE Transactions on Signal Processing, vol. 66, no. 20,
using the max-min fairness approach (which was previously pp. 5438–5453, 2018.
[11] L. Sanguinetti, A. Zappone, and M. Debbah, “Deep learning power al-
only known for MR precoding with long-term power con- location in Massive MIMO,” in IEEE Asilomar Conference on Signals,
straints) and used it to generate the training dataset for DNNs. Systems, and Computers, 2018, pp. 1257–1261.
We showed that a properly trained feed-forward DNN is able [12] I. Goodfellow, Y. Bengio, and A. Courville, Deep learning. MIT press,
2016.
to learn how to allocate powers. This is achieved by using only
large-scale fading information, thereby substantially reducing
the complexity and processing time of the optimization pro-

580

Authorized licensed use limited to: ULAKBIM UASL - MIDDLE EAST TECHNICAL UNIVERSITY. Downloaded on May 28,2023 at 18:47:14 UTC from IEEE Xplore. Restrictions apply.