Physical Communication 49 (2021) 101464

Contents lists available at ScienceDirect

Physical Communication
journal homepage: www.elsevier.com/locate/phycom

Full length article

User association for load balancing in coordinated multipoint green

HetNets: A Quasi-Newton-based approach
∗
Mohamad Khattar Awad a , , Ali A.M.R. Behiry b , Mohammed W. Baidas c
a
Department of Computer Engineering, College of Engineering and Petroleum, Kuwait University, Kuwait
b
College of Engineering and Applied Sciences, American University of Kuwait, Kuwait
c
Department of Electrical Engineering, College of Engineering and Petroleum, Kuwait University, Kuwait

article info a b s t r a c t

Article history: The demand for high capacity network services with stringent quality of service requirements is at a
Received 25 February 2021 rapidly accelerating rate due to the exponential rise in the numbers of mobile-connected devices. This
Received in revised form 1 August 2021 demand has motivated the use of the heterogeneous network (HetNet) architectures. However, even
Accepted 11 September 2021
though small-cell base-stations have relatively low power consumption, the overall aggregate power
Available online 24 September 2021
consumption of a dense HetNet is significant. Due to high inter-cell-interference and imbalanced loads
Keywords: in dense HetNets with conventional user association techniques, cell-edge users perceive dramatically
Heterogeneous networks (HetNets) less quality of service than their cell-center counterparts. The use of a Coordinated Multipoint (CoMP)
Green energy association can augment the service perceived by cell-edge users by allowing a single user to be jointly
Coordinated multipoint (CoMP) served by two base-stations. In this work, we propose a load balancing scheme for CoMP-enabled
transmission HetNets with hybrid energy supplies that jointly optimizes user latency and green energy utilization.
Broyden–Fletcher–Goldfarb–Shanno (BFGS) The proposed scheme employs a fractional solution to the user association problem to decide CoMP
Load balancing
transmission for cell-edge users, ultimately improving their data rates. Performance evaluations of the
User association
proposed scheme show a reduction in latency of 79% and on-grid power consumption by 99% compared
to conventional user association schemes that associate users based on the maximum received signal
strength. Furthermore, an improvement in the network sum-rate for cell-edge users by 24% has been
achieved compared to the traditional association scheme and as much as 40% over other existing
schemes.
© 2021 Elsevier B.V. All rights reserved.

1. Introduction in the network, co- and cross-tier1 interferences increase, lead-

ing to a drastic degradation in the users’ throughput [3,6]. This
5G is envisioned to introduce new use cases, such as the degradation in throughput is more substantial for cell-edge users.
Internet-of-Things (IoT), mission-critical services, ultra-reliable Coordinated Multipoint (CoMP) transmission is a technique
communications, and real-time control applications [1]. Such use where users are served by more than one base-station simulta-
cases impose extreme quality-of-service (QoS) and seamless con- neously. It is considered as a promising solution for mitigating
nectivity requirements, and hence entail massive network capac- inter-cell interference (ICI) for cell-edge users [7], thereby im-
ity and improved coverage, which are already on the rise due proving their signal quality. Particularly, multiple neighboring BSs
to the exponential increase in usage of smart mobile devices [2, can coordinate and jointly transmit signals to users in order to
3]. The deployment of heterogeneous networks (HetNets) fulfills
mitigate such interference. Therefore, in HetNets, coordination
these demands, where small-cell base-stations (SBSs) are densely
between the MBS and SBSs and/or among the SBSs could help
deployed underlaying the macro base-station (MBS) [4]. This de-
reduce the cross- and co-tier interferences, respectively, for cell-
ployment not only improves the network capacity and coverage,
edge users and consequently improve their data rates [6,8]. An
but also offloads the traffic from the MBS to SBSs [5]. Moreover,
due to the close proximity of the users and their respective illustration of the CoMP HetNets powered by grid (in gray) and
SBSs, their received signal strength is significantly improved [3]. green energy sources is shown in Fig. 1.
However, due to the dense deployment of base-stations (BSs)
1 Co-tier interference is the interference among BSs that belong to the
∗ Corresponding author. same tier, e.g., among SBSs. On the other hand, cross-tier interference is the
E-mail addresses: mohamad@ieee.org (M.K. Awad), abehiry@auk.edu.kw interference caused between entities of different tiers, i.e., between MBSs and
(A.A.M.R. Behiry), baidas@ieee.org (M.W. Baidas). SBSs.

https://doi.org/10.1016/j.phycom.2021.101464
1874-4907/© 2021 Elsevier B.V. All rights reserved.
M.K. Awad, A.A.M.R. Behiry and M.W. Baidas Physical Communication 49 (2021) 101464

rates of users located at the edge of the cell. However, the

question of how to load-balance traffic while providing QoS guar-
antees in CoMP-enabled hybrid energy powered HetNets remains
unanswered.
In this work, we tackle the problem of balancing the loads
among BSs of CoMP-enabled and SoftRAN-based HetNets pow-
ered by both on-grid and renewable-energy supplies. To this
end, we propose a centralized load balancing user association
scheme, which jointly optimizes latency and power consumption,
given the network-wide information on average traffic loads and
available green energy. This joint optimization is modeled as a
cost function and minimized using the approximated Newton
direction, which is evaluated using the method jointly proposed
by Broyden, Fletcher, Goldfarb and Shanno (BFGS) [19]. A trade-
off coefficient is implemented to balance power consumption and
latency, while leveraging CoMP transmissions in order to improve
QoS perceived by cell-edge users. Specifically, the objectives of
our proposed scheme are as follows: (1) balance the distribution
of downlink traffic loads among the BSs in such a way that the
latency across BSs is optimized, (2) utilize the available green
Fig. 1. Illustration of a CoMP-enabled HetNets powered by grid and green energy sources to minimize the on-grid power consumption,
energy sources. and (3) improve users’ rates located at the cell-edge via CoMP-
enabled transmissions. The main contributions of this work are
summarized as follows:
Two significant challenges emerge due to the deployment of
CoMP-enabled dense HetNets, namely increased power consump- • We develop a user association optimization model for
tion [9,10] and imbalanced traffic load distribution among cells [4, software-defined networks. The model handles the trade-off
5,8,10]. The utilization of renewable energy sources could aid between latency minimization and on-grid power consump-
in reducing the on-grid power consumption, and thus minimiz- tion conservation, while accounting for CoMP transmissions.
ing CO2 emissions [11]. To support the utilization of renewable • Accordingly, we develop a centralized user association
energy sources and flexible control of the network traffic load dis- scheme that optimizes this trade-off, while enabling CoMP
tribution, softwarization of radio access networks (RAN) becomes transmissions for cell-edge users.
essential [12]. The software-defined RAN (SoftRAN) is an emerg- • We adopt a Quasi-Newton method leveraging the BFGS
ing centralized architecture, which abstracts BSs in a geographical method of approximating the inverse Hessian of an objec-
area as one virtual BS and considers physical BSs as remote tive function. This allows second-order information to be
radio heads (RRHs) [13,14]. SoftRAN facilitates unprecedented employed without any storage or computational constraints.
programmability of network control functions, (e.g., power man-
agement and load balancing), based on a broader view of the The rest of this paper is organized as follows. We review
network status, i.e., BSs’ stored renewable energy, users distribu- related works in Section 2. In Section 3, we present the system
tion, and traffic demand; hence, overcoming challenges posed by model and related network assumptions. In Section 4, we de-
enabling CoMP transmissions in HetNets. vise the proposed user association scheme. Section 5 presents
For renewable energy-powered SoftRAN-based and CoMP- the simulation results of the proposed scheme as well as other
enabled HetNets, optimization of user association, i.e., assigning benchmark schemes. In Section 6, conclusions are drawn from
users to BSs, is critical to exploit network capacity and energy simulation results.
resources. However, associating users with MBSs or SBSs storing
significant renewable energy overloads these BSs with network 2. Literature review
traffic, resulting in a substantial increase in traffic latency. On
the other hand, associating users with BSs in such a way traffic There are numerous efforts dedicated to traffic load balancing
latency is minimized may shift the network traffic load to BSs and distribution for cellular HetNets. One approach is completely
with limited access to green energy; consequently, increasing turning OFF a BS after offloading its traffic to neighboring BSs. For
the on-grid power consumption. Hence, this trade-off between instance, the authors of [20] proposed a low-complexity ON/OFF
renewable energy utilization and network capacity exploitation switching and radio resource management algorithm to optimize
must be warily managed through user association-driven traffic the energy efficiency in dense HetNets. Likewise, the authors
load balancing. However, conventional user association schemes, of [21] proposed an energy-saving algorithm employing joint user
(e.g. [15–18] reviewed in Section 2), aim to maximize perceived association, clustering, and ON/OFF strategies. However, switch-
user signal quality. Such schemes are optimized for homogeneous ing BSs ON/OFF merely based on traffic demands without consid-
and grid-powered traditional networks and therefore, may fail eration of available green energy may lead to under-utilization
in dense HetNets, or at best, yield performance that is severely of renewable energy resources. The authors of [22] switch BSs
far from optimal [15]. In particular, such schemes may result ON/OFF, and optimize subcarrier assignment and energy alloca-
in extremely imbalanced cells’ traffic loads [8], and thus the tion to minimize the average network power consumption. The
degradation of user rates. The latter problem is particularly severe authors of [23] implement ON/OFF BS switching as well as power
at cell-edges, where users suffer from poor reception. Further- control to satisfy QoS requirements, while utilizing green energy.
more, conventional schemes overlook the impact of balancing However, under heavy traffic conditions – where all BSs are ON to
traffic load distribution on the network power efficiency. More- support traffic loads – such ON/OFF switching-based algorithms
over, the previous works reviewed in Section 2 employ CoMP- lose their advantage and struggle with imbalanced traffic load
enabled techniques to enhance network throughput and data distributions on BSs.
2
M.K. Awad, A.A.M.R. Behiry and M.W. Baidas Physical Communication 49 (2021) 101464

An alternative approach to ON/OFF switching is redistributing BSs. Therefore, the control plane is decoupled from the user plane,
traffic to active BSs in order to maintain a specific performance allowing for an intuitive transition to programmable networks
measure while optimizing resource allocation. In [24], a joint and innovative development of novel services [33]. The RANC
BS activation and user association scheme is proposed to load- receives periodic updates from the physical BSs on the interfer-
balance backhaul in dense HetNets. The authors of [25] pro- ence map, traffic flows data, and user’s channel state information,
posed a distributed load-balancing scheme for renewable energy- based on which control decisions are made. These decisions are
powered HetNets, which redistributes traffic loads to optimize reported back to physical BSs as commands and configurations
users’ throughput. In [26], traffic loads were redistributed to for implementation.
achieve proportional-fairness of energy and load-based logarith- Several recent efforts have been devoted to the realization of
mic utility functions. The authors of [16] formulated the users the SoftRAN architecture in evolving wireless networks. For ex-
association problem as a control problem, where the controller ample, in [33], a flexible and programmable platform for SoftRAN
aims to guarantee the required QoS given the available renewable consisting of custom-tailored southbound application-
energy. programming-interfaces (APIs) was developed. Similarly, the au-
The aforementioned works have mainly focused on conserving thors in [34] developed a flexible and programmable platform
the on-grid consumed energy while maintaining proportional specifically for heterogeneous 5G RANs. Furthermore, recently
fairness of traffic loads on BSs, or optimizing user rates. However, the authors in [12] presented a benchmarking tool for the Sof-
the impact of high traffic loads on traffic delivery latency is not tRAN architecture and its controllers. In addition, in recent 3rd
captured by users’ achievable throughput. In [17], latency-aware Generation Partnership Project (3GPP) standards on 5G and next-
load balancing is considered. A distributed latency and energy- generation RAN (NG-RAN) architectures [35], the separation be-
aware user association scheme for 3-tier HetNets is proposed tween the control plane and user plane is strongly adopted.
in [27], while a software-defined centralized latency and energy- In this paper, we adopt a SoftRAN architecture under which
aware load balancing scheme is devised for HetNets in [28]. the physical MBSs and SBSs provide the RANC with a full view of
Despite the significant energy savings and overall network QoS the network status; e.g. estimates of harvested green energy and
enhancements achieved by the aforementioned schemes, they average traffic loads. Based on this status, the proposed scheme
provide no guarantees on the QoS performance of cell-edge users. associates users with BSs to balance the traffic-load, such that
In HetNets, CoMP transmissions can be used to improve the a well-balanced trade-off between traffic latency and on-grid
QoS perceived by cell-edge users. The authors of [18] proposed power consumption is maintained. Then, the RANC reports the
a greedy algorithm combined with a neural network-based al- user association decisions to the physical BSs, which routes flows
gorithm to balance network throughput and traffic loads for and transmits signals accordingly.
CoMP-enabled HetNets. Results demonstrate significant improve-
ments in the QoS perceived by cell-edge users. A load-aware 3.2. Network traffic model
CoMP-enabled HetNet with an arbitrary number of BS tiers is
considered in [3]. The approach aims to decrease the probability Let the set of BSs serving a geographical area A be denoted
of void BSs, while improving downlink data rates and coverage by B. Furthermore, let x ∈ A refer to a specific location on a two-
area of the network. The authors of [29] proposed a resource dimensional grid in the same geographical area2 Therefore, a user
allocation algorithm to maximize energy efficiency for both dense is uniquely identifiable by x and individual users can be referred
green HetNets and network backhaul. Cell range expansion (CRE) to by x. Then, the average signal-to-interference-plus-noise ratio
is a popular technique to load-balance users in HetNets. However, (SINR) measured at location x from BS j can be expressed as
biasing different BSs to achieve specific performance measures
can be a challenging problem. The authors of [30] propose a Q- P g (x)
SINRj (x) = ∑j j , (1)
learning based selection strategy (QSS) to bias user rates and σ +
2
k∈Ij (x) Ik (x)
achieve load-balancing. In [31], particle swarm optimization is
used to optimize the biases for the CRE approach and achieve where Pj denotes the transmission power of BS j to all users.
load-balancing. In [32], the authors evaluate the effectiveness Since the focus of this work is on user association rather than
of load-balancing using CRE in CoMP-enabled HetNets under power allocation and resource allocation, it is common to assume
different biases. a fixed equal transmission power to all users, as in [17,28,36,37].
Then, users channel gains are mainly differentiated by the large-
3. Network model and problem formulation scale fading and shadowing they experience. In addition, user
association is performed at a large time-scale, whereas power
A downlink HetNet consisting of two types of BSs, MBSs and allocation is performed at a smaller time-scale; therefore, by
SBSs, is considered. Both types of BSs are connected to local optimizing one, the other can be considered constant. Moreover,
renewable energy sources as well as the power grid. Particularly, gj (x) is the average channel gain at location x measured from BS
the grid complements the network energy needs when renewable j, which captures the quasi-static effects of path-loss and log-
energy is insufficient. In the following subsections we present our normal shadowing. The channel gain is assumed to be measured
adopted network architectural model, traffic model, and energy at a large time scale [28]. The variable σ 2 models the power
model, followed by problem. level of the noise. The set Ij (x) denotes the subset of BSs that
introduce interference to user’s reception at location x from BS j;
3.1. Network architectural model whereas, Ik (x) is the average interference power introduced by
BS k ∈ Ij . Due to the fractional frequency reuse plan of the
Enabling CoMP transmissions in HetNets requires a high level network, the interference is modeled as static noise [17]. It is
of coordination among MBSs and SBSs, which the SoftRAN ar- important to mention that our proposed scheme is applicable
chitecture can naturally facilitate [33]. SoftRAN first appeared in CoMP-enabled HetNets adopting either fractional frequency
in the pioneering work of Gudipati et al. [14]; when they pro- reuse or interference randomization, in which interference can
posed softwarization of the RAN via abstraction of the BSs as
a single programmable virtual-BS consisting of a radio access 2 The traffic modeling in this section follows the model used in similar works
network controller (RANC) and RRHs representing the physical in the literature, such as in [17,28].

3
M.K. Awad, A.A.M.R. Behiry and M.W. Baidas Physical Communication 49 (2021) 101464

be considered as static noise [17]. A user’s downlink rate rj (x) at that can be served only by utilizing the available green energy,
x is given by as [28]

rj (x) = Wj log2 1 + SINRj (x) , ej − psj

( ) [ [ ]]
(2)
ρjg = max ϵ, min ,1 − ϵ . (11)
where Wj is the total bandwidth of the jth BS. The traffic load at βj
g
the jth BS that is designated for a user located at x is expressed Here, ϵ is a positive small constant that ensure 0 < ρj < 1.
as Therefore, minimization of on-grid power consumption, requires
λ(x)ν (x)ηj (x) g
the traffic load assigned to a given BS be less than or equal to ρj .
ϱj (x) = , (3)
It should be noted that renewable energy does not incur any cost
rj (x)
to our system, and therefore it should not be penalized in the
where λ(x) denotes the number of traffic arrivals (i.e. number of objective function. To reflect this, we use the Courant–Beltrami
packets) at x. The traffic arrivals follow a Poisson distribution, function, given by [39]
while the size of traffic arrivals, denoted by ν (x), is exponen- ]2
tially distributed. The variable ηj (x) denotes a binary association φj (ρj ) = [ρj − ρjg ]+ ,
[
(12)
variable, defined as
which penalizes only on-grid power consumption. The function
1, if user at x is associated with BS j,
{
ηj (x) = (4) is differentiable, continuous and smooth, making it suitable for
0, otherwise. iterative optimization [39]. Minimizing the objective function in
Hence, the traffic load at BS j can be expressed as (12) achieves best green energy utilization, and therefore mini-
∫ mizes the on-grid energy consumption. Notice that L̂j (ρj ) in (8)
ρj = ϱj (x)dx. (5) and φj (ρj ) in (12) are functions of the BSs load ρj ∀j ∈ B, which is
x∈A determined by the network parameters and association variables
The traffic load 0 ≤ ρj < 1, known in queuing theory as ηj (x) ∀j ∈ B, ∀x ∈ A in (4). In the following subsection, we present
utilization, refers to the fraction of time the BS is busy serving our formulation of the users association problem, in which we
ν (x)
users. Let us define the following parameter γ = r (x) , where optimize ηj (x) to minimize the network traffic latency and on-grid
j power consumption.
γ can be considered constant during a single user association
period, such that [28]
3.4. User association problem
E[γ 2 ]
ϑj = , (6)
2 The minimization of the latency indicator function in (8) en-
where E[·] is the expectation operator. Now, the average latency tails offloading users from a heavily loaded BS to a lightly loaded
experienced by a traffic arrival can be expressed as [28] BS. However, it may be the case that a lightly loaded BS does
not have sufficient green energy, ultimately raising the cost in
ϑj ρj (12) and consuming more on-grid power. In addition, the lightly
Lj (ρj ) = . (7)
1 − ρj loaded BS may be far from a user, thereby lowering its data rate
Note that Lj (ρj ) is a function of the jth BS traffic density, and since and increasing traffic delivery latency. Thus, latency minimization
ϑj is a constant, minimizing the latency at the jth BS is equivalent and power conservation often conflict with each other. In turn,
to minimizing the approximation of Lj (ρj ), given by the following define a cost function f that combines both latency minimization
latency indicator function and power conservation, such that
∑ ∑
1 f = L̂j (ρj ) + kj φj (ρj ), (13)
L̂j (ρj ) = . (8)
1 − ρj ∀j ∀j

where kj sets the trade-off between the on-grid power con-

3.3. Energy model
sumption and traffic delivery latency. Then, the user association
problem can be posed as
Both types of BSs are powered by renewable energy sources ∑ ∑
(e.g. solar panels and/or wind turbines), and are also connected minimize f = L̂j (ρj ) + kj φj (ρj ) (14)
to the grid. Although the priority is always given to utilization ηj (x)∀j,∀x
∀j ∀j
of green energy, whenever green energy is fully depleted, a BS
s.t. ϵ ≤ ρj ≤ 1 − ϵ ∀j ∈ B, (15)
complements its excess power requirements from the grid.3 Let ∑
the power consumption of BS j be modeled by [28] ηj (x) ≥ 1, ∀x ∈ A , (16)
pj = βj ρj + psj , (9) ∀j

ηj (x) ∈ {0, 1}, ∀j ∈ B, ∀x ∈ A, (17)

where βj is a constant variable that translates the traffic load to
λ(x)ν (x)ηj (x)
∫
its corresponding load-dependent power consumption, and psj is ρj = dx. (18)
the static load-independent power consumption. Furthermore, let x∈ A rj (x)
poj denote the on-grid power consumption of the jth BS, which is For a sufficiently small ϵ , Constraint (15) ensures BS queue
defined as stability, while Constraint (16) guarantees that all users are as-
poj = [pj − ej ]+ , (10) sociated with at least one BS, and enables CoMP transmission
by allowing users to be associated with more than one BS. It is
where ej is the amount of green energy stored at the jth BS, and worth noting that our problem formulation does not differenti-
[·]+ = max(0, ·). Then, ρjg is defined as the amount of traffic load ate between cell-center users and cell-edge users. All users are
considered CoMP candidates; however naturally, the users most
3 Battery management and energy redistribution are beyond the scope of benefiting from such a CoMP system are those with comparable
this work. For reference, a battery management system for HetNets powered by perceived signal strength from multiple BSs (i.e. cell-edge users).
hybrid energy sources can be found in [38]. The binary association variable ηj (x) is defined in Constraint (17).
4
M.K. Awad, A.A.M.R. Behiry and M.W. Baidas Physical Communication 49 (2021) 101464

The Eq. (18) defines the traffic load of a BS in terms users traffic Negative logarithmic functions are used for inequalities, as they
arrivals, traffic size, data rates and binary association variable. heavily penalize approaching the constraint from feasibility,
The association problem is a mixed-integer nonlinear pro- reaching up to infinity at the constraint. Similarly, the cost of
gramming (MINLP) problem, which is known to be NP-hard [40]. deviating from the equality Constraint (22) is represented by a
Classical MINLP solving techniques, such as branch-and-bound, quadratic penalty function given by [39]
branch-and-cut or even heuristic solutions are too computation- [∑ ]2
ally expensive to the point where their solution would be redun- Q (x) = η̂j (x) − 1 . (26)
dant in the scale of a single user association period. Therefore, ∀j
solving this problem efficiently necessitates relaxation of the
A quadratic function is used to penalize any deviation (positive or
binary association variable ηj (x) into a new association variable
negative) from the equality line, while having a penalty of zero
η̂j (x), where 0 ≤ η̂j (x) ≤ 1. Furthermore, in order to guarantee
when the equality constraint is met. Given these functions, the
that all users are served and associated with at least one BS, we
unconstrained UA problem can be written as
replace Constraint (16) with
∑ ∑ ω∑
B1 (ρj ) −µ2 B2 η̂j (x) + Q (x),
∑ ( )
η̂j (x) = 1, ∀x ∈ A. minimize fuc = f −µ1
η̂j (x)∀j∈B,∀x∈A 2
∀j ∀j ∀j,x ∀x

Thus, a sub-space of the original problem feasible space is de- (27)

fined. In Section 4, this constraint is relaxed in the proposed where µ1 , µ2 and ω are scaling multipliers used to model the
algorithm for users that can be served by more than one BS, and severity of violating or approaching the constraints. Although µ1 ,
thus allowing for CoMP transmissions. Hence, the user association µ2 and ω can be set to arbitrarily large numbers, this would in-
(UA) problem is reformulated as troduce poor convergence properties [41]. Therefore, the problem
∑ ∑ is solved iteratively, while decreasing the values of µ1 , µ2 and
minimize f = L̂j (ρj ) + kj φj (ρj ) (19)
η̂j (x)∀j,∀x increasing the values of ω in each iteration. The values of µ1
∀j ∀j and µ2 are initially set high and decrease in order to iteratively
s.t. ϵ ≤ ρj ≤ 1 − ϵ, ∀j ∈ B , (20) decrease the numerical significance of the barrier functions on the
objective function, and speed up the convergence. On the other
0 ≤ η̂j (x) ≤ 1, ∀j ∈ B, x ∈ A, (21)
∑ hand, the value of ω is initially set to a small value and increases
η̂j (x) = 1, ∀x ∈ A, (22) in order to prevent the violation of the equality constraint. How-
∀j ever, the cost of meeting the equality constraint approaches zero,
∫
λ(x)ν (x)η̂j (x) resulting in the original constrained objective function when the
ρj = dx. (23) constraints are met.
x∈A rj (x) Unlike first order methods, second order methods – mainly
Unlike the original problem, the relaxed problem is convex. Here based on deriving the Hessian – allow optimization algorithms
η̂j (x) denotes the probability that the jth BS serve a user located to enjoy faster convergence. On the other hand, the computa-
at x. This sub-space is later mapped back to the original feasi- tional time and memory required to evaluate the Hessian of
bility space using the fractional association solution; therefore, multivariate functions may hinder the applicability of second
achieving CoMP associations. order methods in dense HetNets. The proposed scheme applies
a limited-memory-Broyden–Fletcher–Goldfarb–Shanno BFGS (L-
BFGS) algorithm to approximate the Hessian, requiring shorter
4. Proposed scheme
time to compute and less memory to store. The algorithm uses
the L-BFGS two loop method from [42] to estimate the inverse of
The RANC gathers information on traffic profiles of users,
the Hessian H.
channel measurements, and green energy availability at BSs. This
For notational brevity, we define η̂ to be a matrix of η̂j (x), ∀j ∈
information is employed by the proposed scheme in order to
B, ∀x ∈ A. Then, the proposed algorithm performs the following
solve the relaxed problem and compute a fractional user as- iterative update, η̂ = η̂ + α d, where d denotes the search
+
sociation. A local search algorithm is developed to project the direction, and α denotes the search step size. The search direction
generated fractional user association solution onto the feasible is computed based on Newton’s direction, which is defined by the
space of the original problem; thereby, associating users to a negative Hessian inverse of the objective function multiplied by
single or multiple BSs. The CoMP transmissions are enabled for its gradient [42]. This search direction is in a descent direction
users associated with multiple BSs. of the objective function in (27). However, in order to avoid
The proposed UA scheme uses augmented Lagrangian and the overhead associated with using the fully formed Hessian,
barrier function methods in transforming the relaxed constrained we use a quasi-newton method that implicitly approximates it.
UA problem into a series of relaxed unconstrained UA problems The objective function in (27) has a gradient with respect to the
that can be solved iteratively using a second order method. More- relaxed association variable, given by
over, the constrained relaxed UA problem is transformed into
∂ fuc cj (x)
an unconstrained relaxed UA problem, via the use of an interior = + 2cj (x)kj [ρj − ρjg ]+
barrier function, which increases exponentially when inequality ∂ η̂j (x) (1 − ρj )2
[ ]
constraints are nearly violated [41]. A logarithmic barrier function cj (x) cj (x)
− µ1 −
B1 is used to model the cost of closely satisfying Constraint (20) ρ −ϵ ((1 − ϵ ) − ρj )
as [ j ] [∑ ]
1 1
− µ2 − +ω η̂j (x) − 1 , (28)
B1 (ρj ) = log (1 − ϵ ) − ρj + log(ρj − ϵ ), η̂j (x) 1 − η̂j (x)
( )
(24)
∀j
and a logarithmic barrier function B2 to model the cost of closely λ(x)ν (x)
where cj (x) = rj (x)
. The gradient vector ∇ fuc is computed
satisfying Constraint (21), as
based on stacking the derivatives with respect to all association
B2 (η̂j (x)) = log η̂j (x) + log 1 − η̂j (x) . variables η̂j (x) for all j and x into a single matrix.
( ) ( )
(25)
5
M.K. Awad, A.A.M.R. Behiry and M.W. Baidas Physical Communication 49 (2021) 101464

The proposed scheme iteratively solves an approximated ver- In Line 15, a line search is performed to determine the step
sion of the original constrained problem as outlined in Algorithm length to move the association variables η̂ along the previously
1. The scheme approximates the Hessian inverse of the objective calculated search direction d. Firstly, we note that the association
function in (27) by storing 2 × m vectors. The size of each vector values η̂ and the search direction d do not change during the
is n, where n is the number of optimization variables and m ≪ n. line search phase of the algorithm. Therefore, for convenience, we
Practically speaking, m is in the order of tens, whereas n is in define
the problem is in the order of thousands. This actively alleviates ψ (α ) = fuc (η̂ + α d). (29)
the need for computing and storing n2 operations every time the
Hessian is computed [42]. Furthermore, the directional derivative of fuc in d (i.e. the deriva-
tive of (29)) is given by
Algorithm 1 L-BFGS User Association ∇ψ (α ) = ∇ fuc (η̂ + αd)d′ , (30)
Input: Initial values of η̂j (x) and trade-off coefficients kj . where ∇ is the gradient operator. In order to ensure the change
Output: User association η̂j∗ (x)∀j ∈ B, ∀x ∈ A to the association variables achieves a significant improvement
1: Set initial µ1 , µ2 , ω and t. in minimization of the objective function, the step length is
2: while termination condition of barrier method not satisfied required – by the line search – to meet a pair of inequalities,
do known as the strong Wolfe conditions [19,42]. The first condition
3: while termination condition of L-BFGS method not satis- is
fied do
4: q := ∇ fuc ▷ Find a search direction d ψ (α ) ≤ ψ (0) + β1 α∇ψ (0), (31)
5: for i = 0 to m − 1 do ▷ L-BFGS two-loop recursion which is referred to as ‘‘the sufficient decrease condition’’. The
6: ai := y′1s si ′ q
i i second condition is
7: q := q − ai yi
8: end for′ |∇ψ (α )| ≤ β2 |∇ψ (0)|, (32)
y s
9: r := y 0′ y0 q which is known as ‘‘the curvature condition’’ [42]. Both conditions
0 0
10: for i = m − 1 to 0 do guarantee that the step length taken achieves sufficient progress
11: b = y′1s y′i r in minimizing the objective function in (27). This is necessary
i i
12: r := r + si (ai − b) because it avoids making unnecessary steps of small length with
13: end for negligible impact on the minimization of the objective function.
14: Compute search direction d := −H∇ fuc = −r Once a step length that satisfies the strong Wolfe conditions is
15: α := Line Search(η̂,d) ▷ Find a step length α that found, the association variables η̂ are updated, as per Line 16.
satisfies the strong Wolfe conditions The vectors si and yi are stored in a queue structure of fixed
+
16: η̂ = η̂ + α d ▷ Update association variables η̂ length. The updates of s0 and y0 are made in each iteration using
17: Discard sm−1 and ym−1 ▷ Update vector lists with new local information, and the queue is pushed forward with the least
local information recent two vectors being discarded, as can be seen in Lines 17 to
18: si+1 := si , ∀i ∈ [1, m − 2] 22.
19: yi+1 := yi , ∀i ∈ [1, m − 2] After the L-BFGS method converges to a solution of the ap-
20:
+
s0 := η̂ − η̂ proximate unconstrained problem, the multipliers ω, µ1 and µ2
21: Generate the updated gradient vector ∇ f+ uc
are adjusted to better model the original constrained problem, as
22: y0 := ∇ f+ uc − ∇ fuc
in Lines 24 to 26. Adjusting the multipliers produces a new un-
23: end while constrained problem and the process is repeated until a solution
24: ω := t ω, µ1 := µ1 /t, µ2 := µ2 /t ▷ Update multipliers of for the original constrained problem is found.
penalty functions The fractional association values η̂j (x) ∀x ∈ A ∀j, ∈ B obtained
25: end while from the barrier method in Lines 2 to 25 represent the probability
26: for all fractional association variables η̂j (x), ∀x, ∀j do ▷ of a user located at x receiving traffic from the jth BS. In order
Associate users with BSs having fractional association higher to obtain a CoMP-enabled association, the algorithm associates
than threshold ϵη users with BSs having a fractional association value larger than a
27: if η̂j (x) ≤ ϵη then threshold ϵη , as seen in Lines 27 to 32.
28: ηj (x) := 0 The line search algorithm in Algorithm 2 iteratively finds a
29: else step length αi that meets the strong Wolfe conditions (Line 12)
30: ηj (x) := 1 or a range of values, i.e., bracket, that contains such a step length
31: end if (Lines 6, 9 and 15). If neither is found during an iteration of the
32: end for line search algorithm, a new trial step length αi+1 is interpolated,
as in Lines 18 and 22. Here, interpolation refers to modeling
The solution to the approximated unconstrained problem is
the objective function as a cubic function that is a function of
obtained using the L-BFGS method. Firstly, in lines 4 to 14, the the step length α and solving for that model’s minimum. The
two-loop recursion method uses two types of vectors in order to function references ‘‘SECTION(·, ·)’’ in Lines 6, 9, and 15 call a
implicitly approximate −H∇ fuc . These vectors are denoted by si sectioning method implemented in Algorithm 3 to find a viable
and yi , where i ∈ [0, m). The vectors si measure local changes α within the provided bracket. The pseudo-code of Algorithm 3
in the optimization variables, while the vectors yi measure local is presented in Appendix A. The interpolation steps mentioned in
changes in the gradients of the objective function in (27). Because Algorithms 2 and 3 (given in Lines 22 and 5, respectively) are
of the implicit use of the approximation, the need to store or used to minimize ψ (α ) within the determined bracket, where
arithmetically use the n2 sized matrix H is eliminated. Instead, a cubic interpolation method is used [19]. Therefore, the step
the two-loop recursion method requires 4 × m × n iterations in length α found by Algorithms 2 and 3 fulfills the strong Wolfe
order to calculate the search direction d [42]. conditions in (31) and (32).
6
M.K. Awad, A.A.M.R. Behiry and M.W. Baidas Physical Communication 49 (2021) 101464

Algorithm 2 Line Search Algorithm

1: function Line Search(η̂,d)
ψ −ψ (0)
2: Initialize α0 := 0, αmax = βmin∇ψ (0) , α1 ∈ (0, αmax )
1
3: i := 1
4: loop
5: if [ψ (αi ) > ψ (0) + β1 αi ∇ψ (0)] then
6: return Section(αi−1 , αi ) ▷ Bracket containing
viable α found
7: end if
8: if [ψ (αi ) > ψ (αi−1 )] and i > 1 then
9: return Section(αi−1 , αi ) ▷ Bracket containing
viable α found
10: end if
11: if |∇ψ (αi )| ≤ −β2 ∇ψ (0) or ψ (αi ) < ψmin then
12: return αi ▷ Viable step length α found
13: end if
14: if ∇ψ (αi ) ≥ 0 then Fig. 2. Base-station deployment scenario with uniformly distributed users.
15: return Section(αi , αi−1 ) ▷ Bracket containing
viable α found Table 1
16: end if Proposed scheme implementation parameters used.
17: if 2αi − αi−1 < αmax then Parameter kj β1 β2 τ1 τ2 τ3 ϵη
18: Interpolate for αi+1 ∈ [αlow , αhigh ] ▷ Determine Value 700 0.01 0.9 3 0.1 0.5 0.01
next trial step length α
19: end if
20: αlb := 2αi − αi−1
Fig. 2 shows the positioning of BSs and users, where circles rep-
21: αub := min(αmax , αi + τ1 (αi − αi−1 )
resent MBSs, triangles represent SBSs and dots represent users.
22: Interpolate for αi+1 ∈ [αi , αmax ] ▷ Determine next trial
Assigned CoMP users are indicated by a star, and are connected
step length α
with a line to each BS they are associated with. Users associated
23: i := i + 1
with only a single BS are colored with the respective BS’s color.
24: end loop
On average, there are 5 traffic arrivals per user with each request
25: end function
averaging 250 Kbs in size. In order to estimate user data rates,
The line search guaranteeing the strong Wolfe conditions does the radio propagation modeling follows baseline test conditions
not only improve the algorithm’s performance, but also pro- from [43], where the path-loss for an MBS is given by
vides useful theoretical properties on convergence. The L-BFGS
method is proven to converge at a linear rate, as well as a local PLMBS = 130.19 + 37.6 log(rp ), (33)
super-linear convergence rate near the solution. This convergence while for a SBS, it is
behavior is similar to the one observed in Newton’s method
and is guaranteed, given that the strong Wolfe conditions are PLSBS = 37 + 30 log(rp ). (34)
maintained [19]. The complexity of the proposed L-BFGS User As- The variable rp denotes the distance in kilometers between the
sociation in Algorithm 1 is analyzed in Appendix B. Furthermore, serving BS and the user location. Furthermore, the radio channel
a proof of convergence of Algorithm 1 is given in Appendix C. is affected by log-normal shadowing with standard deviation of
In summary, given the network parameters reported to the 8 dB. The noise power level is −174 dBm/Hz [43]. The antenna
RANC and the trade-off coefficients, the proposed scheme com- used has a receiver sensitivity of −123 dBm and an antenna
putes the cost of an initial association, which can be set as the gain of 15 dB [43]. The system bandwidth is 20 MHz. Each BS
last used association or according to maximum received signal is equipped with a solar panel of sizes 4 and 1 m2 , for MBSs
strength. The proposed scheme then minimizes the objective and SBSs, respectively. We assume standard test conditions such
function in (19) by iteratively solving unconstrained versions of that the temperature is 25 ◦ C, an air mass of 1.5 spectra and a
the problem in (27). The resulting solution is a fractional user solar irradiation incidence angle of 45 degrees. Under these con-
association that minimizes the objective function in (19). The ditions, the highest achievable direct irradiance is 1 kW/m2 [44].
association values are either 0 or 1 for most cell-center users, The available solar irradiance is bound by this maximum and is
and hence are associated with a single BS. However, some users, assumed to follow a Beta distribution [45,46]. MBSs and SBSs
particularly cell-edge users, experience a split association value. transmit at powers of 43 and 33 dBm, and statically consume
The algorithm simultaneously associates cell-edge users with all 750 and 40 W of power, respectively. The load to power relation
BSs that have non-trivial fractional association values, i.e., above coefficient β is 500 for a MBS and 4 for a SBS [47].
a predefined threshold ϵη . The specific parameters used for simulations are summarized
in Table 1. The proposed scheme is compared with existing as-
5. Simulation results sociation schemes. The scheme named MAXSINR associates each
user with the BS from which the highest maximum SINR is
5.1. Uniformly distributed user locations measured. Two other load balancing user association schemes
are considered as well. The α -optimal scheme aims to minimize
In order to evaluate the proposed scheme, we consider a latency by minimizing a latency indicator function, with α =
deployment scenario in a 4 km2 square geographical region. 2 [17]. Also, the vGALA scheme – presented in [28] – aims to asso-
The area is covered by 3 MBSs and 5 SBSs. These BSs serve ciate users in order to minimize both latency and on-grid power
350 users, with uniform and independently distributed locations. consumption. The vGALA scheme is configured with the default
7
M.K. Awad, A.A.M.R. Behiry and M.W. Baidas Physical Communication 49 (2021) 101464

Fig. 3. Performance comparison of the considered schemes in minimizing: (a) latency and (b) power consumption.

Table 2
Distribution of user-to-BS association observed in the compared schemes with
various storage of Green Energy (G.E.).
BS G.E. (W) MAXSINR α -optimal vGALA Proposed
MBS 1 150 41% 35.4% 32.5% 33%
MBS 2 200 23% 22.5% 21% 23%
MBS 3 335 21.7% 21.1% 21% 25%
SBS 1 5 1.4% 2.5% 2.5% 2.5%
SBS 2 5 1.4% 3.7% 3.7% 4%
SBS 3 4 1.7% 4% 4% 4%
SBS 4 3 4.5% 8% 8% 8%
SBS 5 4 4.8% 7.4% 7.4% 8%

parameters used in [28]. All results are obtained by averaging

over 500 independent network instances.
Table 2 illustrates the distribution of users among the available
MBSs and SBSs as percentages of the total number of users in the
system. It is also clear from Table 2 that users in the MAXSINR Fig. 4. A scenario of BSs deployment with larger density of users at the edges
scheme are more heavily biased towards the MBSs. The α -optimal of cells.

scheme considers latency as its major and sole performance mea-

sure, and thus offloads users towards SBSs to minimize latency.
Furthermore, because in the randomly generated scenario, MBS 1 consumes 49% less power than the MAXSINR scheme. Although,
and MBS 2 have less available green energy than MBS 3; it can the α -optimal scheme does not actively optimize power con-
be seen that the proposed scheme offloads more users towards sumption, the load balancing aspect of the system leads to the
MBS 3 to reduce the amount of power supplied from the grid. Our observed power saving results. It can be observed that the pro-
proposed scheme distributes as much load as possible over the posed scheme yields the best trade-off, as it only has 10% and 5%
SBS and decreasing the most load off MBS 2, which was heavily increase in latency relative to α -optimal and vGALA, respectively,
loaded. Our proposed scheme’s percentages add up to over 100% while consuming 33% and 12% less power compared to these
due to the CoMP associations given to users located at the cell- schemes. Therefore, the performance of the proposed scheme and
edge to improve their assigned rates. This is a part of our scheme’s vGALA are generally comparable in this scenario, where users are
trade-off, where some extra load is carefully imposed on some uniformly distributed.
BSs to improve cell-edge user rates.
Fig. 3 illustrates the performance of all considered schemes in 5.2. Increased user density at cell edges
comparison to the conventional MAXSINR scheme. Fig. 3(a) shows
that the MAXSINR scheme demonstrates the highest latency, We simulate a scenario with higher user density at cell edge
because it heavily associates users with MBSs, as can be seen than the scenario considered in the previous sub-section. Fig. 4
from Table 2. The other schemes have comparable performance in depicts a simulation scenario similar to the previous scenario, in
terms of latency optimization. The proposed scheme reduces the which 250 users are uniformly distributed; however, 100 users
latency indicator in comparison to MAXSINR by 58%. The vGALA out of the 250 are placed along the edges of coverage areas.
scheme reduces the latency by 63% compared to MAXSINR, while Along the edges of cells, signals received from multiple BSs are
the α -optimal scheme reduces the average latency by 68%, which comparable in strength. Users that have been associated with a
is the greatest among all compared schemes. This is expected, as single BS by our scheme are represented by a black dot, users
the sole objective of the α -optimal scheme is to optimize latency. associated with two BSs are represented by a red star, and those
In comparison, the proposed scheme has the least on-grid associated with three BSs are represented by a slightly larger blue
power consumption, consuming 82% less on-grid power than star. Except for user distribution, all other simulation parameters
MAXSINR, as demonstrated in Fig. 3(b). The vGALA scheme con- remain the same as the previous scenario. This leads to further
sumes 70% less power than the MAXSINR. The α -optimal scheme degradation of rates received by cell-edge users.
8
M.K. Awad, A.A.M.R. Behiry and M.W. Baidas Physical Communication 49 (2021) 101464

Fig. 7(a), the MAXSINR scheme demonstrates the worst latency

response among all considered schemes. The other compared
schemes demonstrate comparable latency responses. The α -
optimal scheme showed the most significant reduction in la-
tency equivalent to 80% in comparison to MAXSINR, followed by
the proposed scheme which demonstrates a decrease in latency
equivalent to 79% and lastly vGALA shows a drop that amounts
to 78%. It is clear from Fig. 7(b) that the MAXSINR scheme also
entails the largest power consumption. The α -optimal scheme
consumes 73% less power than the MAXSINR scheme. The pro-
posed scheme and the vGALA scheme consume approximately
no power, saving 99% and 100%, respectively. A 100% savings of
energy consumption implies that the proposed scheme did not
consume any on-grid energy and satisfied its energy needs only
Fig. 5. Percentages of users associated to a single BS or employing CoMP
from green energy. It can clearly be observed that the scheme
transmission in the proposed scheme. has comparable performance to the other schemes when consid-
ering both power consumption and latency, while improving the
Table 3
cell-edge user rates as seen in Fig. 6(b).
Distribution of users (Us.) by BS with various storage of Green Energy (G.E.) Fig. 8 displays the impact of increasing arrival rate on latency
under increased number of cell-edge users. and power consumption. It can be clearly seen from Fig. 8(a)
BS G.E. (W) Total Us. Cell-edge Us. Cell-center Us. CoMP Us. (%) that MAXSINR’s traffic latency increases exponentially as the
Cell-edge Us.
arrival rate increases. The other schemes enjoy an almost linear
MBS 1 150 67 34% 64% 74%
MBS 2 200 110 24% 76% 77%
increase in latency in response to the increase in traffic arrival.
MBS 3 335 81 43% 57% 51% Fig. 8(b) shows that the proposed scheme consumes the least
SBS 1 4 32 84% 16% 22% on-grid power as the arrival rate increases. It is note-worthy
SBS 2 5 32 81% 19% 19% that these effects are achieved while maintaining the sum-rate
SBS 3 4 24 75% 25% 33%
improvements displayed in Fig. 6.
SBS 4 4 14 50% 50% 57%
SBS 5 5 33 79% 21% 31% Fig. 9 shows the trade-off between latency and power con-
sumption ∑ for all compared schemes as total arrival rate over the
region ∀x λ(x) changes from 700 to 1100. The higher the slope
of a curve in Fig. 9, the more latency-aware the scheme is, while
Fig. 5 displays the distribution of users associated with a single the lower the slope, the more power-efficient the scheme is. It can
BS, two or three BSs in the scenario shown in Fig. 4. It can be seen that the MAXSINR scheme scales poorly with increased
be seen that the proposed scheme allows 10% of users to be arrivals, rising rapidly to high latency and high power consump-
associated with two BSs in order to improve their data rates. A tion. It is worth-mentioning that the α -optimal and the vGALA
very small percentage of users were associated with three BSs, as schemes at their given parameters are more latency inclined,
it is unlikely for a user to have comparable signal strengths from while the proposed scheme enjoys a fair balance between latency
3 or more BSs. and power consumption.
Table 3 shows the distribution of users in the proposed scheme In Fig. 10, we explore the effect of changing the trade-off
under the increased cell-edge density scenario. It can be noted
coefficient kj on both on-grid power consumption and average
that cell-center users are more heavily associated with MBSs than
latency. For simplicity, we assume that the trade-off for all BSs
SBSs, as the user rates provided to them by MBSs are higher than
is the same. Practically, the trade-off coefficients can be managed
those provided by SBSs. Therefore, offloading them to SBSs incurs
by network operators independently. It can be noted from Fig. 10
a much higher penalty to the total network latency indicator, in
that when kj = 0, the problem becomes focused solely on latency,
comparison to cell-edge users. On the other hand, cell-edge users
and thus achieving the best latency results and the highest power
are more likely to be distributed among SBSs, as they have similar
consumption. As kj increases, the scheme becomes more sensitive
data rates from several BSs. It is worth noting that the proposed
to power consumption; therefore, the on-grid power consump-
scheme does not overload the SBSs with CoMP users, even if they
tion decreases; whereas, the traffic delivery latency increases, as
are more likely to associate with cell-edge users. In this way, the
CoMP association does not severely impact the other performance can be seen observed Fig. 10. Eventually, as k → ∞, the system
measures. reduces power consumption as feasibly possible without violating
Fig. 6 shows the sum-rate for users located at the center of the problem’s constraints.
the cell and the edge of the cell, under the compared schemes. Fig. 11 illustrates the convergence behavior of the barrier
Fig. 6(a) illustrates that the sum-rate of cell-center users re- method in the proposed scheme. The larger the value of t and
mains comparable across the different schemes, with negligible the higher the initial values of the multipliers, the larger the
differences. However, it can be seen in Fig. 6(b) that the pro- progress made by each unconstrained sub-problem in the barrier
posed scheme achieves a significant sum-rate enhancement for method. However, as the number of sub-problems decreases,
cell-edge users in comparison to other schemes. Specifically, the their difficulty increases, leading the number of L-BFGS iterations
proposed scheme enhances the sum-rate of cell-edge users by required to solve them to increase. For the simulated scenario, it
41%, 48% and 62% over MAXSINR, α -optimal and vGALA, re- was noted that the values of µ1 = µ2 = 0.02, ω = 50, t = 5
spectively. Such enhancements are due to the CoMP-enabled achieves the best performance among all simulation runs, as it
transmission by associating cell-edge users to more than one balances the number of barrier sub-problems and their difficulty.
BS. However, there is a minimal trade-off in terms of slightly
decreased overall latency and power consumption performance 6. Conclusions
in return for the enhancement cell-edge users’ rates.
Fig. 7 shows a performance comparison of the considered In this paper, we investigated the user association problem in
schemes in terms of power consumption and latency. From CoMP-enabled and SoftRAN-based HetNets with hybrid energy
9
M.K. Awad, A.A.M.R. Behiry and M.W. Baidas Physical Communication 49 (2021) 101464

Fig. 6. Sum-rate performance of: (a) cell-center and (b) cell-edge users.

Fig. 7. Comparison between different schemes with increased density of cell-edge users in terms of: (a) average latency Indicator, and (b) on-grid power consumption.

Fig. 8. Effect of increased total arrival rates on the performance of considered scheme in the uniform user distribution scenario.

supplies. The user association problem is modeled as MINLP, algorithm that does not require explicit storage, and formation
which is known to be NP hard. Therefore, a sub-optimal scheme and computation of the Hessian. The proposed scheme jointly
is developed to optimize the trade-off between the on-grid power reduces the average latency and on-grid power consumption
consumption and traffic delivery latency, while ensuring BS using a per-BS trade-off coefficient to emphasize importance of
queue stability and users connectivity. The scheme has been the two objectives conflict. Furthermore, the proposed scheme
designed to be scalable to dense networks by using the L-BFGS enabled CoMP transmissions by using fractional user association
method, a limited memory version of the well-studied BFGS to select users eligible to be jointly serviced by multiple BSs to
10
M.K. Awad, A.A.M.R. Behiry and M.W. Baidas Physical Communication 49 (2021) 101464

The centralized control enabled by the SoftRAN architecture in

the considered network facilitates application of machine learn-
ing (ML) techniques. In our future work, we plan to explore appli-
cability of machine learning techniques towards load balancing.
Particularly, we plan to compare and contrast the performance of
several reinforcement learning techniques in handling the trade-
off between latency and energy efficiency in HetNets powered by
hybrid energy sources.

Declaration of competing interest

The authors declare that they have no known competing finan-

cial interests or personal relationships that could have appeared
to influence the work reported in this paper.

Acknowledgments

The authors would like to thank Dr. Tao Han at the University
Fig. 9. Trade-off between on-grid power consumption and latency for the
proposed schemes, by varying total arrival rate from 700 to 1100. of North Carolina, Charlotte, USA for the insight given on the
methods in [28]. This work was supported and funded by Kuwait
University Research Grant No. EO-08/18. This work was also par-
tially supported by the Kuwait Foundation for the Advancement
of Sciences (KFAS), under project code PN17-15EE-02.

Appendix A

The Sectioning phase in Algorithm 3 iteratively reduces the

width of the bracket that was found in Lines 6, 9 and 15 until
a viable step length is found [19]. The value αlb is updated,
such that it is guaranteed to meet the first Wolfe condition. The
bracketed interval is iteratively reduced (Lines 7, 13 and 15) in
order to converge on a step length α that meets the second
Wolfe condition (Line 10). Here, τ1 , τ2 , τ3 , β1 and β2 are constant
parameters. The algorithm is insensitive to small changes in these
parameters; thus, they do not have to be re-tuned for different
network scenarios. Algorithm 3 sets a value for α within the
identified bracket, and then iteratively reduces that bracket width
Fig. 10. Effect of increasing the trade-off coefficient kj on the system until a viable step length is identified.
performance in the uniform user distribution scenario.

Algorithm 3 Sectioning Phase

1: function Section(αlow , αhigh )
2: loop
3: αlb := αlow + τ2 (αhigh − αlow )
4: αub := αhigh − τ3 (αhigh − αlow )
5: Interpolate for α ∈ [αlb , αub ] ▷ Generate trial α
6: if ψ (α ) > ψ (0) + β1 α∇ψ (0) or ψ (α ) ≥ ψ (αlow ) then
7: αhigh := α ▷ Shrink bracket
8: else
9: if |∇ψ (α )| ≤ −β2 ∇ψ (0) then
10: return α ▷ Found step length α satisfying
strong Wolfe conditions
11: end if
12: if ∇ψ (α )(αhigh − αlow ) ≥ 0 then
13: αhigh := αlow ▷ Shrink bracket
14: end if
15: αlow := α ▷ Shrink bracket
Fig. 11. Convergence behavior under different initial values of barrier method 16: end if
parameters. 17: end loop
18: end function

improve cell-edge users’ data rates. Simulations showed that the Appendix B
proposed scheme achieves significant improvements in reducing
The total number of iterations performed by Algorithm 1 in
latency, power consumption and improving cell-edge user rates the outermost loop is κ , where κ is a constant, determined by the
over existing schemes. update factor t. Let the total number of all decision variables η̂j (x)
11
M.K. Awad, A.A.M.R. Behiry and M.W. Baidas Physical Communication 49 (2021) 101464

in the system be n. Also, the maximum number of saved vector Let z = Ḡ1/2 s, then
pairs used to approximate the Hessian is m, where m is usually
y′ y s′ Ḡ2 s z ′ Ḡz
between 10 and 15 [42]. The value of m is comparatively small = = ≤ M. (42)
to the problem size n, and is not scenario specific. The two-loop y′ s s′ Ḡs z′z
recursion method used to approximate the Hessian has a worst- The subsequent proof relies on the following theorem, called
case complexity of O(mn) [42]. Experimentally, the line search Zoutendijk’s theorem [42].
algorithm does not contribute significantly to the computational
load in comparison to the two-loop recursion. Therefore, for Theorem C.1. Consider an algorithm driven by the iterative update
brevity, we assume that line search has worst-case complexity in (36), where d is a descent direction and α satisfies the Wolfe
O(ln). Therefore, the worst-case complexity of the algorithm is conditions in (31) and (32). Furthermore, the function fuc is bounded
O(κ mn + κ ln). It can be shown that the convergence rate of the below and is continuously differentiable on an open set N containing
algorithm is linear and locally superlinear near the solution of the level set L := {η̂ : fuc (η̂) ≤ fuc (ηˆ0 )}. Also, the gradient ∇ fuc is
each sub-problem, when meeting the strong Wolfe conditions.
Lipschitz continuous on N . Then
This is similar to the convergence behavior achieved when using ∑[
cos2 (θk ) ∥∇ fuc ∥2 < ∞
]
Newton’s direction [42]. (43)
∀k
Appendix C
where θk is the angle between the search direction d and the steepest
The proof in this section follows the approach and employs descent search direction −∇ fuc .
properties from [42]. Throughout this section, we refer to the
exact Hessian of the unconstrained objective function ∇ 2 fuc as G. Theorem C.2. Let B0 be any symmetric positive definite initial Hes-
The L-BFGS approach used closely approximates the BFGS update. sian approximation. Then the sequence generated by (36) converges
to η̂ of the objective function fuc .
∗
In this section, we prove the global convergence of our uncon-
strained optimization subproblem using a BFGS update. The un-
constrained optimization subproblem depends on the following Proof. We define the following
iterative step
y′ s y′ y
η̂+ = η̂ + α d, (35)
mk = , Mk = . (44)
s′ s y′ s
where the step size α is determined by a line search that satisfies Note that from (41) and (42) that
the Wolfe conditions in (31) and (32), and the search direction is
determined by mk ≥ m, Mk ≤ M . (45)

d = −B−1 ∇ fuc , (36) It can be shown that the trace of the BFGS update in (37) is

where B is the BFGS approximation to the Hessian updated iter- ∥Bk s∥2 ∥y∥2
trace(Bk+1 ) = trace(Bk ) − + . (46)
atively such that s′ Bs y′ s
Bss′ B yy′ It can also be shown that the determinant of the BFGS update is
B+ = B − + . (37)
s′ Bs y′ s y′ s
det(Bk+1 ) = det(Bk ) . (47)
Proving that the algorithm converges to the optimal value ηˆ∗ s′ Bs
is equivalent to proving that We can now define the angle θk as
lim inf ∥∇ fuc ∥ = 0. (38) s′ Bs
k→∞ θk = , (48)
∥s∥ ∥Bs∥
The following proof relies on properties of the objective func-
tion, which are as follows: and
s′ Bs
(i) The objective function fuc is twice differentiable. q= . (49)
(ii) The objective function fuc is continuous on the open set N s′ s
containing the level set L := {η̂ : fuc (η̂) ≤ fuc (ηˆ0 )}. We can then rewrite
(iii) The gradient ∇ fuc is Lipschitz continuous on N . ∥Bs∥2 ∥Bs∥2 s′ Bs q
(iv) The level set L is convex and there exists positive constants = = , (50)
s′ Bs ′ 2
(s Bs) ∥s∥ 2
cos (θk )
2
such that
as well as rewrite
m ∥z ∥2 ≤ z ′ Gz ≤ M ∥z ∥2 (39)
y′ s s′ s mk
for all z ∈ Rn . This assumption is preserved by the L- det(Bk+1 ) = det(Bk ) = det(Bk ) . (51)
s′ s s′ Bs q
BFGS two-loop method from [42]. The set convexity, along
with the lower bound on the objective function implies Let us now define the following function of the approximate
the existence of a unique minimizer η̂∗ to the objective Hessian
function. ψ (B) = trace(B) − log(det(B)), (52)
We define the average Hessian Ḡ as where ψ (B) is always positive. We can now write
∫ 1
Ḡ = ∇ 2 fuc (η̂ + τ α d)dτ (40) q
ψ (Bk+1 ) = trace(Bk ) + Mk − − log(det(Bk )) − log(mk ) + log(qk ),
0 cos2 (θk )
and the property following from the Taylor series representation (53)
of the BFGS update y = Ḡα d = Ḡs. Using this property and (40),
we arrive at which in turn can be rewritten as
[
′
ys ′
s Ḡs q
= ≥ m. (41) ψ (Bk+1 ) = ψ (Bk ) + uk + 1 −
s′ s s′ s cos2 ( θk )
12
M.K. Awad, A.A.M.R. Behiry and M.W. Baidas Physical Communication 49 (2021) 101464
( )]
q [13] A. Papa, R. Durner, L. Goratti, T. Rasheed, W. Kellerer, Controlling next-
+ log + log(cos2 (θk )), (54)
cos2 ( θk ) generation software-defined RANs, IEEE Commun. Mag. 58 (7) (2020)
58–64.
where uk = Mk − log(mk ) − 1. The term inside the square bracket [14] A. Gudipati, D. Perry, L.E. Li, S. Katti, SoftRAN: Software defined radio access
is non-positive, and therefore from (45) and (54), we can state network, in: Proc. of the Second ACM SIGCOMM Workshop on Hot Topics
the following on the initial Hessian approximation, in Networking, HotSDN, 2013, pp. 25–30.
[15] J.G. Andrews, S. Singh, Q. Ye, X. Lin, H.S. Dhillon, An overview of load
k balancing in hetnets: old myths and open problems, IEEE Wirel. Commun.
∑
0 < ψ (Bk+1 ) ≤ ψ (B0 ) + u(k + 1) + log(cos2 (θi )), (55) 21 (2) (2014) 18–25.
[16] L.A. Fletscher, J. Barreiro-Gomez, C. Ocampo-Martinez, C.V. Peroni, J.M.
i=0
Maestre, Atomicity and non-anonymity in population-like games for the
where 0 < u = M − log(m) − 1. From (43), we know that the energy efficiency of hybrid-power HetNets, IEEE Trans. Netw. Serv. Manag.
gradient does not converge to zero if and only if cos θi goes to 15 (4) (2018) 1600–1614.
[17] H. Kim, G. de Veciana, X. Yang, M. Venkatachalam, Distributed α -optimal
zero. Let us assume that cos θi does go to zero. Therefore, there
user association and cell load balancing in wireless networks, IEEE/ACM
will exist an iteration k1 such that for all i > k, we have Trans. Netw. 20 (1) (2012) 177–190.
[18] J. Yu, Y. Wang, Q. Zhang, Q. Jiang, S. Yue, A greedy algorithm-based
log(cos2 (θi )) < −2c . (56) BP neural network for user association in HetNets, in: Proc. of 2nd
IEEE Advanced Information Management, Communicates, Electronic and
Then we can rewrite (55) as
Automation Control Conference, IMCEC, 2018, pp. 340–345.
k1
∑ k
∑ [19] R. Fletcher, Practical Methods of Optimization, second ed., Wiley-
0 < ψ (B0 ) + u(k + 1) + log(cos2 (θi )) + (−2c), (57) Interscience, New York, NY, USA, 1987.
[20] H. Ghazzai, M.J. Farooq, A. Alsharoa, E. Yaacoub, A. Kadri, M. Alouini, Green
i=0 i=k1 +1 networking in cellular HetNets: A unified radio resource management
framework with base station ON/OFF switching, IEEE Trans. Veh. Technol.
or equivalently,
66 (7) (2017) 5879–5893.
k1 [21] L. Tang, W. Wang, Y. Wang, Q. Chen, An energy-saving algorithm with joint
∑
0 < ψ (B0 ) + u(k + 1) + log(cos2 (θi )) + 2ck1 + c − ck, (58) user association, clustering and On/Off strategies in dense heterogeneous
networks, IEEE Access 5 (2017) 12988–13000.
i=0
[22] Y. Li, M. Sheng, Y. Sun, Y. Shi, Joint optimization of BS operation, user as-
The right hand side of the previous equation is negative for large sociation, subcarrier assignment and power allocation for energy-efficient
k, therefore cos(θi ) does not go to zero, and therefore due to HetNets, IEEE J. Sel. Areas Commun. 34 (12) (2016) 3339–3353.
[23] S. Zhang, N. Zhang, S. Zhou, J. Gong, Z. Niu, X. Shen, Energy-aware traffic
Zoutendijk’s theorem, ∥∇ fuc ∥ → 0 and the problem converges. offloading for green heterogeneous networks, IEEE J. Sel. Areas Commun.
Due to the convexity of the problem, the converged solution is 34 (5) (2016) 1116–1129.
also the optimal solution η̂∗ . □ [24] Y.L. Lee, W.L. Tan, S.B.Y. Lau, T.C. Chuah, A.A. El-Saleh, D. Qin, Joint
cell activation and user association for backhaul load balancing in green
HetNets, IEEE Wirel. Commun. Lett. 9 (9) (2020) 1486–1490.
References
[25] Q. Han, B. Yang, C. Chen, X. Guan, Energy-aware and QoS-aware load
balancing for HetNets powered by renewable energy, Comput. Netw. 94
[1] European Comission, 5G Vision: The 5G Infrastructure Public Private
(2016) 250–262.
Partnership: The Next Generation of Communication Networks and
[26] T. Zhang, H. Xu, Y. Chen, User association for energy balancing in HetNets
Services, Tech. Rep, European Comission, European Union, 2015,
with hybrid energy sources, in: Proc. of IEEE 85th Vehicular Technology
Feb., (Online), Available: https://espas.secure.europarl.europa.eu/orbis/
Conference, VTC Spring, IEEE Vehicular Technology Conference, 2017, pp.
document/5g-vision-5g-infrastructure-public-private-partnership-next-
1–5.
generation-communication.
[27] D. Liu, Y. Chen, K.K. Chai, T. Zhang, Distributed delay-energy aware user
[2] P. Agyapong, M. Iwamura, D. Staehle, W. Kiess, A. Benjebbour, Design
association in 3-tier HetNets with hybrid energy sources, in: 2014 IEEE
considerations for a 5G network architecture, IEEE Commun. Mag. 52 (11)
Globecom Workshops, GC Wkshps, 2014, pp. 1109–1114.
(2014) 65–75.
[28] T. Han, N. Ansari, A traffic load balancing framework for software-defined
[3] C. Liu, P. Chen, Load-aware coordinated multipoint joint transmission in
radio access networks powered by hybrid energy sources, IEEE/ACM Trans.
dense heterogeneous networks: Downlink coverage and throughput limits,
Netw. 24 (2) (2016) 1038–1051.
in: Proc. of IEEE International Conference on Communications, ICC, IEEE
International Conference on Communications, 2017, pp. 1–7. [29] K.M.S. Huq, S. Mumtaz, J. Bachmatiuk, J. Rodriguez, X. Wang, R.L. Aguiar,
[4] I. Siomina, D. Yuan, Load balancing in heterogeneous LTE: Range opti- Green HetNet CoMP: Energy efficiency analysis and optimization, IEEE
mization via cell offset and load-coupling characterization, in: Proc. of Trans. Veh. Technol. 64 (10) (2015) 4670–4683.
IEEE International Conference on Communications, ICC, IEEE International [30] C. Jialing, Y. Mingxi, D. Xiaohui, J. Bingli, Q -learning based selection
Conference on Communications, 2012, pp. 1357–1361. strategies for load balance and energy balance in heterogeneous networks,
[5] L. You, L. Lei, D. Yuan, Load balancing via joint transmission in het- in: 2020 5th International Conference on Computer and Communication
erogeneous LTE: Modeling and computation, in: Proc. of IEEE 26th Systems, ICCCS, 2020, pp. 728–732.
Annual International Symposium on Personal, Indoor, and Mobile Radio [31] H.P. Kuribayashi, M.A. De Souza, D. De Azevedo Gomes, K. Da Costa Silva,
Communications, PIMRC, Aug., 2015, pp. 1173–1177. M.S. Da Silva, J.C. Weyl Albuquerque Costa, C.R. Lisbôa Francês, Particle
[6] A.H. Sakr, E. Hossain, Location-aware cross-tier coordinated multipoint swarm-based cell range expansion for heterogeneous mobile networks,
transmission in two-tier cellular networks, IEEE Trans. Wireless Commun. IEEE Access 8 (2020) 37021–37034.
13 (11) (2014) 6311–6325. [32] Y. Cao, H. Xia, C. Feng, Evaluation of diverse cell range expansion
[7] G.T. V11.0.0, coordinated multi-point operation for LTE, 2011, 3GPP TSG strategies applying CoMP in heterogeneous network, in:Proc of. IEEE 24th
RAN WG1, Sept.. Annual International Symposium on Personal, Indoor and Mobile Radio
[8] L. Liu, Y. Zhou, V. Garcia, L. Tian, J. Shi, Load aware joint CoMP clustering Communications, PIMRC, 2013, pp. 1962–1966.
and inter-cell resource scheduling in heterogeneous ultra dense cellular [33] X. Foukas, N. Nikaein, M.M. Kassem, M.K. Marina, K. Kontovasilis, FlexRAN:
networks, IEEE Trans. Veh. Technol. 67 (3) (2018) 2741–2755. A flexible and programmable platform for software-defined radio access
[9] G. Cili, H. Yanikomeroglu, F.R. Yu, Cell switch off technique combined networks, in: Proceedings of the 12th International on Conference on
with coordinated multi-point (CoMP) transmission for energy efficiency in Emerging Networking EXperiments and Technologies, 2016, pp. 427–441.
beyond-LTE cellular networks, in: Proc of. IEEE International Conference [34] E. Coronado, S.N. Khan, R. Riggio, 5G-EmPOWER: A software-defined
on Communications, ICC, 2012, pp. 5931–5935. networking platform for 5G radio access networks, IEEE Trans. Netw. Serv.
[10] N. Saxena, A. Roy, H. Kim, Traffic-aware cloud RAN: A key for green 5G Manag. 16 (2) (2019) 715–728.
networks, IEEE J. Sel. Areas Commun. 34 (4) (2016) 1010–1021. [35] 3GPP, TS 138 401 V15.9.0 (2020-11) 5G and NG-RAN Architecture De-
[11] Vodafone, Sustainable Business Report 2018, Tech. Rep., Vodafone Group scription (3GPP TS 38.401 Version 15.9.0 Release 15), ETSI, 2020, Standard,
Plc., 2018. 11.
[12] A. Papa, R. Durner, E. Goshi, L. Gorattiy, T. Rasheedy, A. Blenk, W. Kellerer, [36] T. Han, N. Ansari, Green-energy aware and latency aware user associations
MARC: On modeling and analysis of software-defined radio access network in heterogeneous cellular networks, in:2013 IEEE Global Communications
controllers, IEEE Trans. Netw. Serv. Manag. (2021) 1–1. Conference, GLOBECOM, 2013, pp. 4946–4951.

13
M.K. Awad, A.A.M.R. Behiry and M.W. Baidas Physical Communication 49 (2021) 101464

[37] D. Liu, Y. Chen, K.K. Chai, T. Zhang, M. Elkashlan, Two-dimensional opti- He received the Ontario Research & Development Challenge Fund Bell
mization on user association and green energy allocation for HetNets with Scholarship in 2008 and 2009, the University of Waterloo Graduate Scholarship
hybrid energy sources, IEEE Trans. Commun. 63 (11) (2015) 4111–4124. in 2009, and a fellowship award from the Dartmouth College, Hanover, NH in
[38] X. Liu, T. Han, N. Ansari, Intelligent battery management for cellu- 2011. In 2015 and 2017, he received the Kuwait University Teaching Excellence
lar networks with hybrid energy supplies, in: Proc. of IEEE Wireless Award and Best Young Researcher Award, respectively.
Communications and Networking Conference, 2016, pp. 1–6.
[39] E. Chong, S. Zak, An Introduction To Optimization, in: Wiley Series in
Discrete Mathematics and Optimization, Wiley, 2013. Ali A. M. R. Behiry (S’18) earned his B.Eng. in computer
[40] P. Belotti, C. Kirches, S. Leyffer, J. Linderoth, J. Luedtke, A. Mahajan, engineering from the American University of Kuwait,
Mixed-integer nonlinear optimization, Acta Numer. 22 (2013). Kuwait, in 2015 and the M.Sc. in computer engineering
[41] S. Boyd, L. Vandenberghe, Convex Optimization, Cambridge University from Kuwait University, Kuwait, in 2019.
Press, New York, NY, USA, 2004. He has served as a part-time lab assistant and tutor
[42] J. Nocedal, S.J. Wright, Numerical Optimization, second ed., Springer, New during his undergraduate years from 2012 to 2015 at
York, NY, USA, 2006. the American University of Kuwait, as well as a lab
[43] IEEE, IEEE 802.16m Evaluation Methodology Document (EMD), Tech. Rep., coordinator and instructor at Kuwait University from
IEEE, 2009. 2016 to 2019, during his graduate studies. Currently,
[44] IEEE, IEEE Recommended Practice for Testing the Performance of Ali is an undergraduate lab instructor in the Depart-
Stand-Alone Photovoltaic Systems, 2004, pp. 1–18, IEEE STD 1526-2003. ment of Engineering at the American University of
[45] S.H. Karaki, R.B. Chedid, R. Ramadan, Probabilistic performance assessment Kuwait.
of autonomous solar-wind energy conversion systems, IEEE Trans. Energy His recent research interests include wireless mobile networks, wireless
Convers. 14 (3) (1999) 766–772. sensor networks, optimization and machine learning. Ali received the highest
[46] Y.M. Atwa, E.F. El-Saadany, M.M.A. Salama, R. Seethapathy, Optimal renew- academic merit scholarship from the American University of Kuwait in 2011, as
able resources mix for distribution system energy loss minimization, IEEE well as the full Excellence Scholarship from Kuwait University in 2016.
Trans. Power Syst. 25 (1) (2010) 360–370.
[47] G. Auer, V. Giannini, C. Desset, I. Godor, P. Skillermark, M. Olsson, M.A.
Mohammed W. Baidas received the B.Eng. (Hons.)
Imran, D. Sabella, M.J. Gonzalez, O. Blume, A. Fehske, How much energy
degree in communication systems engineering from the
is needed to run a wireless network? IEEE Wirel. Commun. 18 (5) (2011)
University of Manchester, Manchester, U.K., in 2005,
40–49.
the M.Sc. degree (with distinction) in wireless com-
munications engineering from the University of Leeds,
Leeds, U.K., in 2006, the M.S. degree in electrical engi-
Mohamad Khattar Awad (S’02, M’09, SM’17), earned neering from the University of Maryland, College Park,
the B.A.Sc. in electrical and computer engineering MD, USA, in 2009, and the Ph.D. degree in electrical
(communications option) from the University of Wind- engineering from Virginia Tech, Blacksburg, VA, USA, in
sor, Ontario, Canada, in 2004 and the M.A.Sc. and Ph.D. 2012. He was a Visiting Researcher with the University
in electrical and computer engineering from the Uni- of Manchester in the Academic Years of 2015/2016
versity of Waterloo, Ontario, Canada, in 2006 and 2009, and 2018/2019. He is currently an Associate Professor with the Department
respectively. From 2004 to 2009 he was a research of Electrical Engineering, Kuwait University, Kuwait, where he has been on
assistant in the Broadband Communications Research the faculty since May 2012. He is also a frequent reviewer for several IEEE
Group (BBCR), University of Waterloo. From 2009 to journals and international journals and conferences, with over 80 publications.
2012, he was an Assistant Professor of Electrical and His research interests include resource allocation and management in cognitive
Computer Engineering at the American University of radio systems, game theory, cooperative communications and networking, and
Kuwait. Since 2012, he has been with Kuwait University, where currently he is green and energy-harvesting networks. He also serves as a technical program
an Associate Professor of Computer Engineering. committee member for various IEEE and international conferences. He was
Dr. Awad’s research interest includes wireless and wired communications, a recipient of the Outstanding Teaching Award of Kuwait University for the
software-defined networks resource allocation, wireless networks resource allo- academic year of 2017/2018.
cation, and acoustic vector-sensor signal processing. He is a frequent reviewer for
several journals and conferences. Dr. Awad served on the editorial board of the
IEEE Transactions on Green Communications and Networking (TGCN) between
October 2016 and May 2021.

14