UtilitasMathematica

ISSN 0315-3681 Volume 120, 2023

Smart Dynamic Resource Allocation Controller in Bandwidth Slicing Based

on Spiking Neural Networks

Mohammed Mousa Rashid Al-Yasari1, Nadia Adnan Shiltagh Al-Jamali2

1
Information Institute for Postgraduate Studies (IIPS), Iraqi Commission for Computers and
Informatics (ICCI), Baghdad, Iraq, phd202020555@iips.icci.edu.iq
2
University of Baghdad, College of Engineering, Computer Engineering Department, Baghdad,
Iraq

Abstract

Software Defined Networking (SDN) is a modern network architectural model that manages
network traffic using software. SDN is a networking scenario that modifies the conventional
network design by combining all control features into a single place and making all choices
centrally. Controllers are the "brains" of SDN architecture since they are responsible for making
control decisions and routing packets at the same time. The capacity for centralized decision-
making on routing improves the performance of the network. SDN's growing functionality and
uses have led to the development of many controller systems. Every SDN controller idea or
design must prioritize the control plane since it is the most crucial part of the SDN architecture.
This paper, present Dynamic Resource Allocation based on the Spiking Neural Network
(DRASR) approach, which is responsible to distribute available resources among all slices justly
according to the type of demand. The simulation results show that the benefit of the network
throughput is about 98.8%.

1. Introduction
Because it is more controllable, dynamic, and cost-effective than traditional architecture, SDN is a
strong network architecture that is best for high-bandwidth applications that change quickly [1]. The
idea that a network's control operations should be kept distinct from its forwarding functions became
the foundation of the SDN design. This would make it simpler to directly program the network control
and abstract forwarding devices for services and network applications [2].
• Easy to program: Since the control function of the forwarding device has been taken away, the
network's control operation can be directly programmed.
• Agile: The network administrator can change network traffic on the fly to meet different
management needs since control is separated from the infrastructure underneath.
• Centralized management: Because the network's brain (controller) is logically centralized and
appears as a single switch to the application and policy engines, it has a global view of the entire
network.

368
 UtilitasMathematica
ISSN 0315-3681 Volume 120, 2023

• Employing dynamic SDN programs, network administrators may rapidly manage, modify, protect,
and enhance network resources with SDN.
SDN is an emerging network paradigm that enables current network architecture constraints to be
overcome; it is characterized as a foursome-pillar network architecture [3]. In a distinct data and
control plane, routing decisions are flow-based rather than destination-based, the control logic is
handled by an external entity known as the SDN Controller, and the network may be programmable
through software applications running on top of the SDN Controller. According to the examination of
the relevant literature, an attempt was made to compare the existing SDN controllers. The SDN
concept and its technical components were thoroughly analyzed in [4]. However, this assessment was
not intended to be a comparative analysis of controllers from a commercial standpoint. Referring to
aspects such as the difficulty of getting started, the Application Program Interface (APIs) that are
supported, the accessibility of documentation, and the version of Openflow, among other factors,
conducted a comparative analysis of SDN controllers based on a systematic study. Unfortunately, this
work does not do a comparative categorization of the offered controllers, nor does it emphasize the
market viewpoint, which is crucial to the acceptability of any technology [5]. Give a comparison of
SDN controllers based on characteristics however, they do not perform a market-oriented comparison,
such as programming language, documentation, modularity, and performance. This paper's primary
contribution is a comparative analysis of the existing SDN controllers and their primary
characteristics, taking into account not only functional and technical aspects, nevertheless, but market
adoption, documentation availability, and OpenFlow support are also all-important factors. To
evaluate the study, we made a comparison between the method based on Spike Response and Integrate
and fire then compared it with other techniques such as traditional neural networks.

2. SDN Architecture
Based on the prior SDN description, SDN components may be characterized as a collection of the
separate data layers, control layers, and application layers that reflect the SDN architecture, as shown
in 'figure 1', each of which has its own functionality and can interact through open standard interfaces.
These layers were then depicted using a bottom-up approach [6].

369
 UtilitasMathematica
ISSN 0315-3681 Volume 120, 2023

Figure 1. Fundamental SDN components

2.1 Data Plane: This plane provides a description of the forwarding devices, which include switches
and routers, in addition to a set of instructions that may be given via an application program interface
(API). SDN network devices function similarly to traditional network devices, except those packets
are forwarded based on a higher plane decision. This signifies that control is no longer delegated to an
external party and is now logically centralized. The data plane and the control plane are connected
through a standard interface (OpenFlow). In other words, open and standard interfaces are used to
build the network brain (control) and applications (conceptually). The controller may use this interface
to dynamically setup various forwarding devices. For traffic processing logic in SDN data plane
forwarding devices, an API for interacting with the controller, an abstraction layer, and a traffic
(packet) processing function will be implemented as software in virtual switches and as hardware in
physical switches [7]. The abstraction layer is made up of one or more flow tables, and its main
function is to enable the device to decide what to do with the next packet based on its contents. The
packet may be routed to a particular switch port, flooded to all ports, or dropped entirely [8]. A flow
table in an open flow switch is a data structure placed in a high-speed data plane data structure It
provides information about the forwarding and packet handling behavior of the open flow switch.
There are one or more flow entries in an open flow table, each with a number of components. A flow
table with three entries (match fields, action, and priority), as well as a counter and timer [9].
2.2 Control Plane: SDN controller, also known as Network Operating System (NOS), is the name
given to the control plane in SDN architecture. Due to the fact that the controller is connected to all
devices that perform forwarding in the bottom plane, management of the network exchange moves
from distributed to centralized [2]. The controller's primary functions are as follows:
370
 UtilitasMathematica
ISSN 0315-3681 Volume 120, 2023

• Provide the applications plane with an abstraction view of the underlying infrastructure so they
can link with devices that use the SDN (switches, routers).
• Execute the directives of the administration (load balancing, forwarding, and routing).
• Command and control all the devices that make up the network's data plane [10].
Because malfunctioning nodes are linked to the controller, the controller's logical location assists in the
resolution of many distributed issues, such as quicker reactions to node or link failures. Because the
controller has a full picture of the whole network, loop avoidance is substantially easier. Depending on
the programming language used for implementation, there are several kinds of SDN controllers, such
as the pox controller, which is implemented in the Python language, the flooding light controller,
which is developed in the Java programming language, and even the NOX controller, which could use
the C programming language. They're all open-source controllers, however, there are also commercial
ones like HP and NEC [11].
2.3 SDN Application Layer is a programmable platform provided by SDN technology that enables
users to build SDN applications for routing management and resolving critical network issues.
Network applications communicate with the controller using an API known as the northbound
interface in SDN architecture. These applications' primary function is to manage traffic within network
devices by modifying flow entries via the southbound interface [12].

3. Related Work
For dispersed 5G-based SDN/NFV networks, the best workload distribution has been presented [13].
A network slicing architecture from end to end is made to enable a variety of services, including
URLLC and eMBB. Additionally, by splitting client requests in the integrated environment of SDN,
NFV, and edge computing, the network operating cost may be decreased. Network slicing was carried
out by the authors in a large-scale Internet of Things (IoT) context (long-range wide area network) in a
prior research [14]. The network is divided into three slices: the best-effort slice, the reliability-aware
slice, and the urgency and reliability-aware slice. First, one-to-many matching is used to execute
cooperative slicing, with the number of IoT devices allocated to virtual slices serving as the
determining factor. Then, using a one-to-one matching game, resources are assigned for each slice
(inter-slice resource allocation). The coalitional multigame theory generates significant computational
complexity and requires a lot of processing time.
Resource sharing and allocation in 5G slice networks are thought to be improved by packet-based data
traffic scheduling [15]. Static sharing resources (SSR) and dynamic sharing resources (DSR) are the
two operational modes that are utilised. The given capacity weight is calculated for each slice and
distributed in accordance with resource allocation. In the end, it is calculated how fairly resources are
distributed across slices. A global network controller is necessary to manage vast varieties of slices,
including popular, heavy, and sensitive slices. A slice management plan that allocates resources
according to priority has also been presented by researchers [16]. Lower priority slices are sent
through different pathways, whereas requests for high priority slices are sent along the shortest paths.
200 nodes are used for the experiments, and they are set up in a grid network architecture. In the end,
371
 UtilitasMathematica
ISSN 0315-3681 Volume 120, 2023

the delays of slices are reduced by up to 11–14% while the average throughput slices up to 6, 13, and
7%. The shortest route is used for flow (high priority) forwarding in the data plane. The quickest route
is not always accessible, however. Therefore, the relevant flow must be put in the controller when the
flow does not match a flow table. This causes customers with static resource allocation to have a
bandwidth shortage issue. Network slicing has been accomplished via service function chaining [17].
Every slice has its own set of service function chains that handle various types of traffic. The trade-offs
between slicing and execution runtime are then examined using a greedy-based heuristic approach.
The requisite bandwidth and latency are then achieved using an optimisation model. Network slice
mobility is not taken into account, which lowers QoS. To assign resources for each slice in a network,
a network slicing resource management (NSRM) system has been created [18]. In this research, an
LTE network is taken into account for varying slice assignments and equitable bandwidth allocation
across slices. The LTE slice controller manages all slices and deploys the controller for each slice. The
radio network resources are distributed through a virtual eNodeB via the LTE slice controller. Slice
request dynamic provisioning is really challenging. For instance, the Industry 4.0 application has to
handle huge slice requests.

4. Machine Learning and Spiking Neural Network

Much effort is put into optimizing and effectively scheduling radio and network resources, but in 5G
networks, resource allocation on the basis of the service being provided is a must-have. The increasing
number of devices and new services offered by 5G networks will add to the already massive quantity
of data traffic that operators now deal with [19]. This stream of traffic may be divided into smaller
chunks and handled separately. Since the service provider may now charge differently for each sliced
piece and even alter the pricing for each slice, the provider can strike a good balance between
company profitability and customer happiness. As a bonus, 5G network slicing enables service
providers to construct not just established but also new applications and services. It will be a "one size
fits all" solution. Each network section may operate autonomously, with its controls and policy
administrators [20].
With ML in place, we can examine any gaps in our knowledge and make any required adjustments.
Machine learning will offer a network analysis of the massive data set, which may be researched
further to adjust any given slice swiftly and cost-effectively. As shown in Figures 1-a and 1-b, SNN
may dynamically activate network automation to adjust resource allocation. Without human
involvement, SNN will be accountable for delivering and processing, and making an intelligent choice
for network resource adaption. To make the best judgments, it will also weigh some different
variables, maybe more than a single person could take into account in a short time [21].

372
 UtilitasMathematica
ISSN 0315-3681 Volume 120, 2023

Figure 1. Spiking Neural Network Architecture

Whenever a new slice is added to the network, SNN will analyze it in real-time to determine how well
it is doing, where it stands in comparison to other networks, what problems could arise, what sections
of the network are functioning normally, and if anything seems out of the ordinary [22]. An easy way
to see this is on a slice for a fixed wireless corporate network, where capacity may be dynamically
added in response to a sudden increase in demand thanks to automation. We could then communicate
more efficiently. This will make it easier to create new services or network slices if they become
required. With automation, we can do all of this in less time with no degradation in the performance of
an ongoing session [20]. Organizational challenges now prevent the widespread use of network
slicing. This is due to the fact that several pieces of hardware and organizations inside a service
provider's network must be interacted with in order for a single modification to be implemented [23].
5G's programmability features will make it possible to create a customized end-to-end solution for
every use case. The data rate, latency, mobility, isolation, power constraints, etc., are typical
parameters that a typical consumer might seek. If the current instance of the network slice does not
have enough resources, then a different network slice type will be made available, and the associated
network services will be activated [24].
Network slicing takes into account a wide variety of factors, such as the kind of slice, bandwidth,
throughput, latency, equipment type, portability, reliability, isolation, power, and many more [25].
Because 5G enables the collection of such large datasets, big data analytics need the use of machine
learning. One of the most important and useful ML-based applications in the wireless sector is the
detection and revival of dormant cellular cells. Other relevant and useful ML-based applications in the
wireless industry include optimizing mobile tower operations, accelerating wireless channel adoption,
facilitating targeted marketing, autonomous decision-making in IoT networks, real-time data analysis,
predictive maintenance, customer churn, sentiment analysis by social networking, fraud detection, e-
commerce, and many more [26]. Since Uber employs real-time differential pricing depending on
demand, the number of available vehicles, the weather, the time of day, and other factors, using ML in
apps that are comparable to Uber will have several advantages. Better accuracy and prediction in the
future may be achieved by using a platform built on machine learning to analyze and process massive
373
 UtilitasMathematica
ISSN 0315-3681 Volume 120, 2023

amounts of historical and real-time data. This is because automatically adjusts to price differences in
real-time.

5. Proposed Controller
using an SNN to calculate an unknown value for the traffic data that changes over time based on the
number of network slices. The variance (Err) between actual and desired buffer occupancy is used to
update the SNN's weights using the Back Propagation (BP) training method. Because the online
training is used to manage the estimating traffic to accommodate it, the SNN must learn the network's
behaviour and control it.
SNN has eight input nodes, a hidden number of layers, a number of neurons in each layer, and a
number of synapses (sub-connections) that are typically determined experimentally. A large number
of hidden layers slows and decreased the training process and increases network complexity. The
improvement of the non-linearity of the solution requires the adoption of a three-layered feed-forward
neural network to achieve the accomplishment of a traffic signal controller.
The structure’s sketch of SNN is shown in Fig (2) and Fig (3), the activation function of the neurons in
the hidden layer is tanh. The number of neurons in each of the three layers’ input, the hidden, and the
output layer are 8,8, and 4 respectively.

Figure 2. Structure of SNN

374
 UtilitasMathematica
ISSN 0315-3681 Volume 120, 2023

Input neurons of the SNN are assigned to network TR features TR(t); the number of packets arriving
at the global slice buffer.

Figure 3. Structure of SNN in detail

The number of sub-connections or synapses k in the interaction between the input and hidden layers,
as illustrated in Fig. (4), is five delayed sub-connections. To determine the number of synapses k, the
trial-and-error approach is applied.

Figure 4. Sub-connection consists of five synapses

As shown in Fig. (4), the weight of each synapse effects the spike-response function ε; denotes the
neuron's activation function. In the encoding process, the actual information TR(t) is encoded
information t_h^act computed using Eq. (1).
(𝑇𝑅(𝑡)− 𝑇𝑅min )(𝑡𝑚𝑎𝑥 − 𝑡𝑚𝑖𝑛 )
𝑡ℎ𝑎𝑐𝑡 = 𝑡𝑚𝑎𝑥 − 𝑟𝑜𝑢𝑛𝑑 (𝑡𝑚𝑖𝑛 + (𝑇𝑅𝑚𝑎𝑥 − 𝑇𝑅𝑚𝑖𝑛 )
) (1)

375
 UtilitasMathematica
ISSN 0315-3681 Volume 120, 2023

𝑇𝑅𝑚𝑎𝑥 and 𝑇𝑅𝑚𝑖𝑛 that are represent the maximum and minimum values of the real input
information. The 𝑡𝑚𝑎𝑥 and 𝑡𝑚𝑖𝑛 represent the largest and minimum interval T.

The maximum and lowest values of the actual information for input are represented by the
variables 𝑇𝑅𝑚𝑎𝑥 and 𝑇𝑅𝑚𝑖𝑛 . Maximum and minimum intervals T are represented by the 𝑡𝑚𝑎𝑥
and 𝑡𝑚𝑖𝑛 .

The decoding equation derived from Eq. (1) by Eq. (2)

(𝑡𝑚𝑎𝑥 −𝑡𝑗 − 𝑡𝑚𝑖𝑛 )(𝑇𝑅𝑚𝑎𝑥 − 𝑇𝑅𝑚𝑖𝑛 )
𝑇𝑅(𝑡𝑗 ) = (𝑡𝑚𝑎𝑥 − 𝑡𝑚𝑖𝑛 )
+ 𝑇𝑅𝑚𝑖𝑛 (2)

The SNN algorithm operates in two modes. The first is known as feed-forward mode, in which each
neuron spikes just once at most during each time interval T when membrane potential m exceeds the
value. Starting from hidden layer I, the feed-forward mode constantly examine neuron i to see whether
it has spiked or not. The algorithm uses the next neuron i+1 when the neuron i is spiked. The
membrane potential m_i (t) is calculated by the SNN algorithm according to Eq. (3) based on input
spikes t_h^f of neuron h at input layer H .
𝑘 𝑓
𝑚𝑖 (𝑡) = ∑𝑁𝐻 𝐷 𝑘
ℎ=1 ∑𝑘=1 𝑤ℎ𝑖 ɛ(𝑡 − 𝑡ℎ − 𝑑 ) (3)

The neuron i is not allowed to spike anymore through the remaining period of time interval T, when
the threshold is exceeded at a particular instant t. In the next instant t+1, the neuron i will be reset.
When the neurons in the second layer have completed, the algorithm will repeat the same procedure in
the output layer J, and the back-propagation phase will then start.
The connection weights of synapses are updated when the feed-forward mode finishes. In contrast to
feed-forward, back-propagation starts at the output layer and returns to the hidden layer. The synapses
of the hidden layer will be modified in accordance with Eq. (4), Eq.(5), and Eq.(6).

𝑤𝑖𝑗𝑘 (𝑡 + 1) = 𝑤𝑖𝑗𝑘 (𝑅) − △ 𝑤𝑖𝑗𝑘 (𝑅) (4)

Eq.(2.14) , Eq. (3.8) and Eq. (3.9) show how the updating of synapses input layer.
𝑘
△ 𝑤ℎ𝑖 (𝑅) = 𝜂. 𝛿𝑖 . 𝑦ℎ𝑘 (5)
𝑘 𝑘 𝑘
𝑤ℎ𝑖 (𝑡 + 1) = 𝑤ℎ𝑖 (𝑅) − △ 𝑤ℎ𝑖 (𝑅) (6)

If the Root Mean Square Error (RMSE) exceeds the allowable level of error, the two phases are
repeated. The output of the SNN will be used to estimate network traffic for the next time.
the error which can be used to adjust the weights in SNN can be described in Eq. (7).
𝐸𝑟𝑟(𝑡) = 𝑇𝑅(𝑡) − 𝐵𝐷 (7)

376
 UtilitasMathematica
ISSN 0315-3681 Volume 120, 2023

Figure 5. Flowchart of the proposed controller

In order guarantee balance amongst slices, the Estimated Traffic Rate 〖𝑇𝑅𝑒𝑠𝑡𝑖𝑚𝑎𝑡𝑒 . Controller is then
distributed among the active slices uniformly to rate in the next time. Eq. (8) explains how the rate
adjustment control operates.
𝑇𝑅𝑒𝑠𝑡𝑖𝑚𝑎𝑡𝑒
𝑇𝑅𝑛𝑒𝑤 (𝑆𝑙𝑖𝑐𝑒𝑖 (𝑡 + 1)) = 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑎𝑐𝑡𝑖𝑣𝑒 𝑠𝑢𝑏−𝑠𝑙𝑖𝑐𝑒𝑠
(8)

Where the 𝑇𝑅𝑛𝑒𝑤 (𝑆𝑙𝑖𝑐𝑒𝑖 (𝑡 + 1)) is the new TR for each slice, which depends on it, slice determines
the number of active sub slices in the global slice to avoid congestion in next time

377
 UtilitasMathematica
ISSN 0315-3681 Volume 120, 2023

6. Simulation Setup and Results

The simulation is carried out using the simulation settings specified in Table (1). To evaluate
performance in a high-traffic situation, 100 sub-slice nodes are randomly set up in simulation with
four global slices in the initial position in the coverage region. The amount of active sub slices in each
global slice varies depending on the protocol used, which creates different Traffic (TR) based on the
number of active global slices and active sub slices. The three proposed controllers are designed to
compare the simulation (DRASR which is depend on SNN spike and response, DRAIF which is
depend on SNN integrate and fire, and DRAN that depend on traditional neural network)
Table (1) Simulation parameter

Number of global slices 4 global slices

Number of sub slice 100

Number of sub slices for each
25
global slice
Initial global Buffer size 1500 packets
Initial Buffer size of each sub
150 packets
slice
Data packet size 800 bytes
Simulation time 100 msec.
6.1 Packet Loss Ratio (PLR)
Fig. (5) illustrates the PLR for the network when three approaches based on the neural network are
implemented. These approaches are compared using the PLR parameter of the same network; the
comparison also includes a network without the controller.

Figure 5. Comparison packet loss ratio of the network when three proposed approaches are used and
without
378
 UtilitasMathematica
ISSN 0315-3681 Volume 120, 2023

It is clarified in Fig. (5) that the PLR of the proposed approaches is better than without a controller
because the buffer controller decreases sending rate of the active sub-slice during the transmission
process in high network traffic.
It is also obvious that DRAIF controller which is depend on (SNN Integrate and Fire) performs well
but it hasn’t accuracy as DRASR. The performance DRASR is very good.
6.2 Network Throughput Ratio (NTR)
Fig. (6) represents the proportion of the received packets by the global slice over the simulation in time
in the form of a comparison between networks with proposed approaches controller and networks
without controller.

Figure 6. Comparison of NTR of the network when three proposed approaches are used and without
The network with the proposed approaches controller performs better than the network without a
controller in increasing the proportion of packets received at the global slice for all active sub-slices.
For more explanation, WRA (Without Controller), which means that the network traffic is high due to
the number of active sub slice is high too. Moreover, DRASR performs better along the simulation
time because it is designed to work with any changeable traffic rate to avoid overflow on the buffer of
the global slice.

379
 UtilitasMathematica
ISSN 0315-3681 Volume 120, 2023

7. Conclusion
The SDN is responsible and monitored for the entire network, and the resources are not sufficient for
all slices, a controller must be available on the network to be responsible for a fair and efficient
distribution of resources according to the requirements of each slice based on traffic features. The
results presented the efficiency of the proposed controller spike response (DRASR) compared with the
controller with the Integrate and Fire and Artificial Neural are better in network PLR and throughput.

References
[1] Li, H., Li, P., Guo, S., & Yu, S. (2014, June). Byzantine-resilient secure software-defined
networks with multiple controllers. In 2014 IEEE International Conference on
Communications (ICC) (pp. 695-700). IEEE.
[2] Bannour, F., Souihi, S., & Mellouk, A. (2017). Distributed SDN control: Survey, taxonomy,
and challenges. IEEE Communications Surveys & Tutorials, 20(1), 333-354.
[3] Kreutz, D., Ramos, F. M., Verissimo, P. E., Rothenberg, C. E., Azodolmolky, S., & Uhlig, S.
(2014). Software-defined networking: A comprehensive survey. Proceedings of the IEEE,
103(1), 14-76.
[4] Dixit, A., Hao, F., Mukherjee, S., Lakshman, T. V., & Kompella, R. (2013). Towards an
elastic distributed SDN controller. ACM SIGCOMM computer communication review,
43(4), 7-12.
[5] Khattak, Z. K., Awais, M., & Iqbal, A. (2014, December). Performance evaluation of
OpenDaylight SDN controller. In 2014 20th IEEE international conference on parallel and
distributed systems (ICPADS) (pp. 671-676). IEEE.
[6] Chekired, D. A., Khoukhi, L., & Mouftah, H. T. (2017). Decentralized cloud-SDN
architecture in smart grid: A dynamic pricing model. IEEE Transactions on Industrial
Informatics, 14(3), 1220-1231.
[7] Perrot, N., & Reynaud, T. (2016, March). Optimal placement of controllers in a resilient SDN
architecture. In 2016 12th International Conference on the Design of Reliable
Communication Networks (DRCN) (pp. 145-151). IEEE.
[8] Raghunath, K., & Krishnan, P. (2018, July). Towards a secure SDN architecture. In 2018 9th
International Conference on Computing, Communication and Networking Technologies
(ICCCNT) (pp. 1-7). IEEE.
[9] Kleinrouweler, J. W., Cabrero, S., & Cesar, P. (2016, May). Delivering stable high-quality
video: An SDN architecture with DASH assisting network elements. In Proceedings of the
7th International Conference on Multimedia Systems (pp. 1-10).
[10] Bakshi, K. (2013, March). Considerations for software defined networking (SDN):
Approaches and use cases. In 2013 IEEE Aerospace Conference (pp. 1-9). IEEE.
[11] Benzekki, K., El Fergougui, A., & Elbelrhiti Elalaoui, A. (2016). Software‐defined
networking (SDN): a survey. Security and communication networks, 9(18), 5803-5833.
[12] Li, Y., & Li, J. (2014, November). MultiClassifier: A combination of DPI and ML for
application-layer classification in SDN. In The 2014 2nd International Conference on
Systems and Informatics (ICSAI 2014) (pp. 682-686). IEEE.

380
 UtilitasMathematica
ISSN 0315-3681 Volume 120, 2023

[13] L. Ma, X. Wen, L. Wang, Z. Lu and R. Knopp., An SDN/NFV based framework for
management and deployment of service based 5G core network, China Communications,
15(10) (2018) 86-98. https://doi.org/10.1109/CC.2018.8485472.
[14] S. Dawaliby, A. Bradai and Y. Pousset., Distributed network slicing in large scale IoT based
on coalitional multi-game theory, IEEE Transactions on Network and Service Management,
16(4) (2019) 1567-1580. https://doi.org/10.1109/TNSM.2019.2945254.
[15] S. A. AlQahtani., An efficient resource allocation to improve QoS of 5G slicing networks
using general processor sharing-based scheduling algorithm, International Journal of
Communication Systems, 33(4) (2020) e4250.https://doi.org/10.1002/dac.4250.
[16] K. Koutlia, R. Ferrús, E. Coronado, R. Riggio, F. Casadevall, A. Umbert and J. Pérez-
Romero., Design and Experimental Validation of a Software-Defined Radio Access Network
Testbed with Slicing Support, Wireless Communications and Mobile Computing, (2019) 1-
17. https://doi.org/10.1155/2019/2361352.
[17] A.J. Ramadha., Overview and implementation of the two most important candidate 5G
waveforms, Journal of Theoretical and Applied Information Technology, 97(9) (2019)
2551–2560.
[18] R. A. Addad, M. Bagaa, T. Taleb, D. L. C. Dutra and H. Flinck., Optimization model for
cross-domain network slices in 5G networks, IEEE Transactions on Mobile Computing,
19(5) (2019) 1156-1169. https://doi.org/10.1109/TMC.2019.2905599.
[19] A. S. D. Alfoudi, S. H. S. Newaz, A. Otebolaku, G. M. Lee and R. Pereira., An efficient
resource management mechanism for network slicing in a LTE network, IEEE Access, 7
(2019) 89441-89457. https://doi.org/10.1109/ACCESS.2019.2926446.
[20] Tavanaei, A., Ghodrati, M., Kheradpisheh, S. R., Masquelier, T., & Maida, A. (2019). Deep
learning in spiking neural networks. Neural networks, 111, 47-63.
[21] Wang, X., Lin, X., & Dang, X. (2020). Supervised learning in spiking neural networks: A
review of algorithms and evaluations. Neural Networks, 125, 258-280.
[22] Lobo, J. L., Del Ser, J., Bifet, A., & Kasabov, N. (2020). Spiking neural networks and online
learning: An overview and perspectives. Neural Networks, 121, 88-100.
[23] Lee, J. H., Delbruck, T., & Pfeiffer, M. (2016). Training deep spiking neural networks using
backpropagation. Frontiers in neuroscience, 10, 508.
[24] Zenke, F., & Ganguli, S. (2018). Superspike: Supervised learning in multilayer spiking
neural networks. Neural computation, 30(6), 1514-1541.
[25] Neftci, E. O., Mostafa, H., & Zenke, F. (2019). Surrogate gradient learning in spiking neural
networks: Bringing the power of gradient-based optimization to spiking neural networks.
IEEE Signal Processing Magazine, 36(6), 51-63.
[26] Bouvier, M., Valentian, A., Mesquida, T., Rummens, F., Reyboz, M., Vianello, E., &
Beigne, E. (2019). Spiking neural networks hardware implementations and challenges: A
survey. ACM Journal on Emerging Technologies in Computing Systems (JETC), 15(2), 1-
35.

381