OPEN ACCESS | Research Article

Dynamic bandwidth allocation with machine learning in

dense WiFi network
a,b
Ricardo Alvarado , Bayron Opinab,c , Johan Telleza , and Vivian Trianaa
a
Grupo de Nuevas Tecnologías (GNET), Unidades Tecnológicas de Santander (UTS), 680005 Santander, Colombia; b Departamento de
Ingeniería Telemática (DIT), Universidad de Las Palmas de Gran Canaria (ULPGC), 3507 Las Palmas de Gran Canaria, Spain;
c
Departamento de Electrónica y Telecomunicaciones (DE&T), Instituto Tecnológico Metropolitano (ITM), 050001 Antioquia, Colombia

Corresponding author: Ricardo Alvarado (email: ralvarado@correo.uts.edu.co)

Abstract
Efficient load balancing is a fundamental aspect within a densely populated Wireless Fidelity (WiFi) network, particularly
when the goal is to evenly distribute bandwidth and regulate Internet usage. The distribution of network traffic for load
balancing can be achieved by utilizing quality of service protocols and access control lists or filters. This document introduces
the application of a machine learning-based prediction model to outline time intervals of congestion in a densely populated
WiFi network employing dynamic load balancing. Historical data related to associated users and WiFi network bandwidth
serve as key indicators to assess the network’s congestion level. Dynamic load balancing is executed by allocating bandwidth to
priority traffic based on observed congestion levels. To implement load balancing, access control lists based on time and quality
of service parameters are configured within the network devices.The experimental findings reveal a notable improvement in
the performance of a WiFi network through the implementation of our dynamic load balancing approach, resulting in a 50%
reduction in lost packets. This innovative method allows for the definition of priority traffic types during different congestion
periods.
Key words: load balancing, bandwidth efficiency, traffic redistribution, WiFi networks

1. Introduction
network, the decision tree technique in machine learning
The problem of optimal bandwidth management in net- (Syrigos et al. 2019) was used. A decision tree is a supervised
works and communications is crucial. Some alternatives to machine learning technique that allows the construction of
carry out an efficient distribution of traffic and improve band- predictive models from data analysis (Farooq et al. 2021). In
width management by performing load balancing in a net- our proposal, we assign higher bandwidths to priority traf-
work are the access control list (ACL), Quality of Service (QoS) fic from the network router in periods of time where there
management policies, or use applications of third parties. is congestion, as shown in Fig. 1. In time period 1, a certain
The use of third-party applications in the restriction and man- traffic (traffic type in red) is predominant. In time period 2,
agement of bandwidth in networks worries researchers and another type of traffic (traffic type in blue) becomes the dom-
companies due to the high costs and data privacy. Due to inant traffic.
these factors and the increasing popularity in academics and Our article contributes by presenting a method that utilizes
industry of new data analysis methods, a novel method is pro- machine learning to dynamically allocate bandwidth for pri-
posed to dynamically allocate bandwidth for a dense Wireless ority traffic during defined congestion times. The degree of
Fidelity (WiFi) network using machine learning (Awad and congestion in the dense WiFi network is determined by the
Abuhasan 2016). bandwidth usage and the number of associated users. The
To carry out this research, we used the campus network dynamic allocation of bandwidth for priority traffic at the
of a higher education institution where the behavior pattern router adapts according to the observed congestion level. This
of network traffic and the number of connected users is pre- proposal can be implemented with contemporary technolo-
dictable. We take the historical data of bandwidth and users gies like software-defined networks (SDNs) (Chang et al. 2019)
connected to the access points (APs) of the dense WiFi net- or traditional router techniques such as filters or time-based
work. These data were analyzed to obtain information on access control lists (TBACLs).
congestion times where the number of connected users and We validate our approach by assessing the packet loss of
the consumption of bandwidth in the network are high. The priority traffic in a segmented WiFi network using virtual lo-
problem of network congestion is packet loss, especially pri- cal area networks (VLANs) (Struhar et al. 2019). Bandwidth
ority traffic. To predict the congestion times in the dense WiFi distribution during various congestion periods is achieved

FACETS 10: 1–9 (2025) | dx.doi.org/10.1139/facets-2024-0128 1

 Canadian Science Publishing

Fig. 1. Bandwidth redistribution for periods of congestion time.

through the configuration of QoS and TBACLs. TBACLs, em- r Reduction of packet loss through predictive modeling: The
ploying criteria such as IP address, port number, or appli- research addresses the critical issue of network congestion
cation type within specific time intervals, contribute to im- and packet loss, especially in prioritized traffic. By using
proved traffic management during congestion, leading to re- predictive models based on the decision tree technique, the
duced packet loss in the dense WiFi network. Prediction mod- project successfully anticipates congestion times in a dense
els using the decision tree technique through machine learn- WiFi network. The dynamic allocation of bandwidth dur-
ing determine the time ranges for ACLs. Consequently, our ing these periods reduces packet loss for prioritized traffic,
proposed approach results in a reduction in lost packets for thereby improving efficiency and service quality.
prioritized network traffic. r Integration of time-based ACLs and QoS policies: The
The adjustments in bandwidth distribution for priority project proposes a comprehensive solution to the network
traffic align with predictions based on the consumption be- congestion problem by integrating TBACL and QoS policies.
havior of bandwidth and the number of users associated with The combination of these two strategies, along with conges-
the WiFi network. During periods of heightened traffic de- tion prediction through Machine Learning, enables more
mand, our method reduces packet loss and latency times, en- effective traffic management during critical periods. This
suring a more equitable distribution of network bandwidth provides a significant improvement in bandwidth distribu-
and an efficient Internet access service. tion and reduces packet loss in a dense WiFi network.
The article structure initially highlights the distinctions be-
tween our proposal and other approaches aiming to mini-
These contributions demonstrate how the application of
mize packet loss through load balancing or dynamic band-
advanced techniques, such as machine learning, coupled
width allocation. Subsequently, we detail our data collection
with the integration of specific network policies, can ef-
strategy and the employed prediction model, explaining how
fectively address challenges in bandwidth management and
the model’s results influenced the establishment of QoS poli-
packet loss in complex network environments.
cies for the network topology. Finally, we present packet
The document is organized as follows: Section 2 discusses
loss measurements using our method to demonstrate the en-
the background and related works. Section 3 explains the
hancement in bandwidth distribution and overall network
proposed system and its application in the campus network
efficiency.
of the higher education institution Unidades Tecnológicas de
The main contributions of our paper are the following:
Santander (UTS). Next, Section 4 analyzes the results through
the emulation of the model on the institution’s WiFi net-
r Dynamic bandwidth allocation with machine learning: The work. Finally, the conclusions are presented in Section 5.
project proposes an innovative method for dynamically al-
locating bandwidth in a dense WiFi network using ma-
chine learning techniques, specifically the decision tree
2. Related work
technique. This represents a significant contribution to the The concept of load balancing can be analyzed from var-
field of bandwidth management, as it leverages the predic- ious perspectives. When viewed from a server standpoint,
tive capabilities of machine learning to adapt bandwidth load balancing refers to the process of assigning and redis-
allocation in real-time. tributing load among available resources to achieve lower
2 FACETS 10: 1–9 (2025) | dx.doi.org/10.1139/facets-2024-0128
 Canadian Science Publishing

utilization, thereby minimizing costs, power consumption, the best machine prototype is a crucial consideration given
and response time (Radhika and Duraipandian 2021) (Wang the defined scope of our study.
et al. 2021). In wide area networks (WANs), load balancing is Load balancing in a LAN through QoS involves prioritizing
centered around packet routing mechanisms and algorithms and managing traffic based on its characteristics, ensuring ef-
(Wang et al. 2018). In our research, our aim is to implement ficient utilization of resources. Strategies include traffic clas-
load balancing in a wireless local area network (WLAN), en- sification, implementation of priority queues, allocation of
suring an efficient distribution of bandwidth and Internet ac- guaranteed bandwidth for critical services, bandwidth limi-
cess. tation for less prioritized traffic, congestion control, contin-
Some researchers have conducted load balancing investi- uous monitoring, QoS application on network devices, and
gations through SDN in emulated WLAN networks (Manzoor dynamic load balancing. The key lies in adapting these poli-
et al. 2020), while others have done it using physical devices cies to the specific needs of the network, coordinating de-
(Manzoor et al. 2020) (Manzanares-Lopez et al. 2018). SDN, vices, and maintaining constant monitoring for optimal per-
with its centralized architecture, facilitates the management, formance.
measurement, and control of dense WiFi networks. In these The implementation of dynamic QoS in a LAN with vari-
cases, load balancing is achieved by making user association able behavior involves adjusting policies according to chang-
decisions or controlling access to AP channels at the physical ing conditions. Key steps include continuous monitoring, his-
layer level. Although our proposal is not designed to operate torical analysis to anticipate changes, definition of dynamic
at the physical layer, it can be implemented in both tradi- thresholds, policy automation through automated systems,
tional networks that are not SDN-based and those utilizing dynamic prioritization during congestions, adaptation to us-
SDN. age patterns, consideration of machine learning, ensuring
Considering that our proposal incorporates machine learn- scalability and flexibility, periodic testing to validate effec-
ing and ACLs, there are load balancing studies where flows tiveness, and clear documentation to facilitate understand-
are redirected using ACL policies based on prediction models ing and collaboration. Flexibility and real-time adaptability
(Ibrar et al. 2021). In such instances, load balancing occurs at are essential for optimal performance in dynamic environ-
the network layer level, employing multiple routers. ments.
Context-based access control (CBAC) represents a critical se- An innovative method is introduced for automatically bal-
curity feature integrated into network devices such as routers ancing the load in a university campus network through
and firewalls, aimed at scrutinizing and managing network the dynamic adjustment of ACLs over time. This allows the
traffic based on contextual parameters. Unlike traditional bandwidth to adapt according to the specific needs of each
ACLs that rely solely on IP addresses or ports for packet fil- VLAN within the network. Traffic measurement campaigns
tering, CBAC takes into account the connection state and ap- are conducted on the university campus network, and ma-
plication layer information to inform its filtering decisions. chine learning is employed to predict the traffic patterns for
CBAC exhibits the capability of bidirectional traffic inspec- different VLANs. It is noteworthy that, in contrast to some
tion, enabling it to ascertain the legitimacy of initiated con- previously mentioned documents, our proposed model does
nections from the internal network and validate responses not necessitate the modification of network devices, thereby
originating from the external network as part of autho- avoiding additional costs or potential impacts on the effi-
rized connections. Additionally, CBAC conducts specific ap- ciency of information traffic and the quality of the user ex-
plication protocol inspections to bolster security measures perience.
(Hypolite et al. 2020).
In contrast, for bandwidth management, we use TBACL, a
specific type of ACL allowing the establishment of filtering
3. Proposed method
rules that apply only during specified time periods. These In a densely populated WiFi network, various types of vari-
ACLs enable the scheduling of when filtering rules are ap- ables can serve as indicators of congestion levels in the VLANs
plied or removed on a router, firewall, or other network de- of a network. Bandwidth and number of users are two in-
vice (Li et al. 2023). dicators of the traffic usage and congestion levels in a WiFi
Time-based ACLs prove useful in scenarios where network network. These indicators can be easily measured in a WiFi
behavior needs modification during specific times, days of network. We can represent the different values of bandwidth
the week, or months. For instance, you can configure a time- and the number of users in the VLANs of a network over time
based ACL to permit certain types of traffic during working through two vectors, as depicted in eqs. 1 and 2:
hours and restrict them outside of that timeframe, or limit ac-
cess to certain resources during weekends (Cao and Ai 2022). →
(1) μ (t ) = (u1, u2, u3, . . . uj)
The authors (Awad and Abuhasan 2016) utilized the
KMeans clustering algorithm to distribute bandwidth among →
nodes in a WLAN on a university campus. They identified con- (2) b (t ) = (b1, b2, b3, . . . bj)
gestion times for each node or AP, implementing load balanc-
ing through the efficient distribution of APs in the wireless Different APs within a WiFi network are used to gather in-
network. While our proposal may suggest relocating APs in formation and store it through two vectors: u(t), denoting the
the WiFi network or even controlling the activation or deacti- number of associated users, and b(t), representing the band-
vation of specific APs in the load balancing process, selecting width in each VLAN of the network. These time-dependent

FACETS 10: 1–9 (2025) | dx.doi.org/10.1139/facets-2024-0128 3

 Canadian Science Publishing

variables play a crucial role in determining the level of con- Fig. 2. Bandwidth distribution for priority traffic in periods
gestion within a dense WiFi network. of congestion time.
From these two indicators (u(t) and b(t)), the level of conges-
tion in a network can be determined. In this way, congestion
in a network occurs when the number of users is higher, and
all users are consuming a significant amount of bandwidth.
There should not be congestion in the network if the number
of users is small, and there are users who are not connected or
consuming bandwidth. This relationship can be represented
through:
→ → → →
(3) n (t ) = k (t ) ∗ u (t ) ∗ b (t )
Fig. 3. Example configuration of a time-based access control
The value of “k” represents the bandwidth usage index per list on a CISCO router.
user. This index changes over time. For example, in a uni-
versity, the “k” index is higher during laboratory classes that
require internet connectivity and may be lower during exam
periods when the use of internet-connected devices is prohib-
ited. Each of the VLANs in the network has different quanti-
ties of users, bandwidth consumption, and values of “k”. Each
VLAN exhibits distinct behaviors in terms of traffic and net-
work congestion over time.
With the variables n(t) indicating the congestion level in
In this example, traffic type 3 is non-prioritized. This traffic
the network, we construct a matrix S(t) with dimensions i × j,
is eliminated during the second congestion period, thereby
referred to as the congestion level matrix in the network. “i”
increasing the bandwidth for type 2 traffic that needs it.
represents the time intervals or the sampling time of the sys-
Generally, the traffic in a university campus network or any
tem, and “j” represents each of the VLANs used in the net-
company follows behavioral patterns. In this case, it is possi-
work.
⎛ ⎞ ble to predict the behavior of the traffic on the network.
n11 n12 n13 . . . nij The control to allow or deny traffic, as well as the configura-
⎜ ⎟ tion of QoS in a network, can be carried out through ACL. To
⎜ n21 n11 n31 . . . nij ⎟
⎜ ⎟ make this type of control dynamic and vary over time, tem-
⎜ ⎟
⎜· · ⎟
(4) ⎜
Sij (t ) = ⎜ ⎟ porary ACLs are used, which change over time.
⎟
⎜· · ⎟
⎜ ⎟
⎜n . . . uij ⎟ 3.1. Time-based access lists (TBACLs)
⎝ i1 ni2 ui3 ⎠
TBACLs are a type of ACL that allows the application of traf-
· fic filtering rules based on time. In other words, they enable
The purpose of matrix S(t) is to determine the time periods specifying when certain access rules should be applied and
T(t) during which congestion exists in the network. when they should be ignored. These lists are commonly used
During periods when many users in the network are con- in network devices such as routers and switches to control
suming a significant amount of bandwidth for a specific access to resources or services based on the schedule.
VLAN, it is important to implement QoS to increase its band- The functionality of TBACL provides flexibility in network
width or decrease the bandwidth in the other VLANs. These management by allowing the enforcement of security poli-
changes in the bandwidth assigned to each VLAN are made cies, QoS, or other access restrictions at specific times.
through QoS. In extreme cases, when congestion levels in a In Fig. 3, an example configuration of a TBACL on a
VLAN are excessive, certain non-priority traffic can be elimi- CISCO router is illustrated. In this example, a TBACL named
nated to utilize the available bandwidth. “WORK_HOURS” is created, allowing traffic from any source
This way, load balancing is achieved: using dynamic band- to any destination during the working hours from Monday
width distribution for each VLAN and implementing access to Friday. Subsequently, this access list is applied to the input
control for each type of traffic within the VLANs. In Fig. 2, interface of the router.
the behavior of three different types of traffic occurring in The use of TBACL provides an effective means to tailor net-
a VLAN is depicted. During a time period Xt1, ranging from work policies and bandwidth distribution to the specific time
an initial time Ti1 to a final time Tf1, traffic type 1 increases requirements of an organization.
and becomes congested. During this period, it is necessary to
implement QoS to enhance the bandwidth for that specific 3.2. Application of our dynamic QoS model on
VLAN and eliminate non-priority traffic from other VLANs in a university campus
the network. The dynamic load balancing method was implemented in
For a time period Xt2 occurring from an initial time Ti2 to the dense WiFi network of the educational institution UTS.
a final time Tf2, a second type of network traffic increases. To achieve this, software was installed to measure the band-
In this case, it is necessary to adjust the QoS configuration. width and the number of users associated with the WiFi APs

4 FACETS 10: 1–9 (2025) | dx.doi.org/10.1139/facets-2024-0128

 Canadian Science Publishing

Fig. 4. Users associated with the access point in the Wireless Fidelity network.

every 5 min. With the 288 samples of the two variables (num- Fig. 5. Decision tree with prediction of periods of congestion.
ber of users and bandwidth used), a matrix was constructed MSE, mean squared error.
indicating the congestion level in the network S(t) for each
VLAN.
A matrix S(t) is obtained for each day of the week. This
matrix contains information about the traffic behavior and
congestion level in the network. Figure 4 shows the scat-
ter plot of the network traffic and congestion level. The
network traffic exhibits a daily behavior pattern. This ma-
trix is used as input for the prediction model. This way,
the prediction of the traffic behavior in the network is ob-
tained, and the bandwidth control or QoS can be configured
dynamically.
To perform network traffic prediction, data analysis is con-
ducted using the Anaconda computational tool, which in-
cludes the Scikit-Learn library in Python (Singh 2015). Deci-
sion trees were chosen for prediction due to the characteris-
tics of the data. A decision tree in machine learning is a pre-
dictive model that uses a tree-like structure to represent deci- The WiFi network traffic at UTS is divided into two VLANs:
sions and outcomes based on specific conditions. Each node VLAN 9001 and VLAN 901. The priority for traffic in each
in the tree represents a question about the data’s features, VLAN changes over time. VLAN 9001 should have priority dur-
branches represent possible answers, and leaves contain fi- ing the morning congestion hours and is used by administra-
nal predictions. Constructing the tree involves selecting fea- tors and teachers. VLAN 901 is used by students, and its prior-
tures that best divide the data. These trees are valued for their ity is set during the afternoon hours. Table 1 shows the most
interpretability and are commonly used in classification and important characteristics of the two priority VLANs and the
regression tasks. congestion time periods.
Figure 5 displays the results of the model in prediction, Using the prediction model results, congestion time peri-
where nodes or terminal leaves represent network conges- ods were determined for the two VLANs in the UTS WiFi net-
tion times, the number of samples, and mean squared er- work. During time period T1, congestion is observed in VLAN
ror (MSE) values. MSE is a metric that measures the average 900, while in time period T2, VLAN 901 experiences conges-
squared difference between a model’s predictions and the tion.
actual values. In the context of a decision tree in machine Load balancing configuration was achieved by bandwidth
learning, MSE is used to assess the quality of splits at each distribution and restricting internet access. During conges-
node during tree construction. Minimizing MSE is the goal, tion periods, higher bandwidth was allocated to prioritized
aiming to select questions or conditions that minimize the traffic, and non-prioritized traffic faced internet access re-
discrepancy between the model’s predictions and the actual strictions. Bandwidth distribution was performed through
values at each step. In summary, MSE guides the tree con- QoS configuration, assigning higher bandwidth. In the ini-
struction process, seeking splits that optimize the model’s tial congestion period, T1, greater bandwidth is assigned to
accuracy. a specific prioritized traffic type in VLAN 900. In the second

FACETS 10: 1–9 (2025) | dx.doi.org/10.1139/facets-2024-0128 5

 Canadian Science Publishing

Table 1. Characteristics of virtual local area networks (VLANs) with priority traffic.
VLAN Name Security User Network congestion
VLAN 900 #UTSCONCORAZON WPA2 240 T1 = 6:45 am to 10:00 am
VLAN 901 #SOYUTEISTA Without password 3560 T2 = 5:00 to 9:00 pm
Note: UTS, Unidades Tecnológicas de Santander.

Fig. 6. Bandwidth distribution with Quality of Service (QoS) and internet access with time-based access control lists (ACLs): (a)
bandwidth increase for virtual local area network (VLAN) 900 in T1; (b) increased bandwidth for VLAN 901 and restriction of
non-priority traffic in T2.

congestion period, another type of traffic from VLAN 901 is Fig. 7. Network traffic behavior over a week.
prioritized. In other words, prioritized traffic and configured
QoS vary depending on the congestion periods. Internet ac-
Traffic Measures
cess restrictions were configured by denying non-prioritized 300
traffic. This dynamic QoS configuration in the UTS WiFi net- 250
work is illustrated in Fig. 6. During the initial congestion pe- 200
riod T1, prioritized traffic from VLAN 900 (marked in red).
Mbps

150
In this scenario, there is packet loss for VLAN 901 and non-
100
prioritized traffic (identified in black). In the subsequent time
50
period T2, prioritized traffic comes from VLAN 901 (identified
0
in blue). During this time, packet loss due to congestion af-
09:00
10:00
11:00
00:00
01:00
02:00
03:00
04:00
05:00
06:00
07:00
08:00

12:00
13:00
14:00
15:00
16:00
17:00
18:00
19:00
20:00
21:00
22:00
fects traffic from VLAN 900. In this T2 period, non-prioritized 23:00
Time
traffic is eliminated by an ACL. Monday Tuesday
Wednesday Thursday
Friday Average

4. Results
We present the results of the implementation of our dy- depicted for the days of a week. Data traffic behavior is not
namic bandwidth distribution system using TBACL in the UTS subject to control and is associated with the statistical pat-
WiFi network. A study of the traffic behavior on the network terns of network utilization. A tendency is exhibited during
was conducted for the two VLANs used. It was verified that periods of network congestion.
the traffic follows a daily pattern, experiencing congestion Over the course of 3 months, a comprehensive recording
at 6:45 a.m. and 5:00 p.m. Congestion in the 6:45 a.m. traf- and analysis of data regarding the behavior of traffic on the
fic occurs in VLAN 900, and congestion at 5:00 p.m. in VLAN UTS WiFi network was conducted. These meticulously gath-
901. This traffic behavior repeats every working day on the ered data were integrated into the prediction method using
UTS campus throughout the week. In Fig. 7, the traffic behav- an input matrix S(t). The choice of the decision tree technique
ior in VLAN 900 and VLAN 901 of the UTS WiFi network is for prediction is grounded in its proven effectiveness, as high-

6 FACETS 10: 1–9 (2025) | dx.doi.org/10.1139/facets-2024-0128

 Canadian Science Publishing

Table 2. Network information during congestion hours.

Drop packet (5 min) AP loked (15 min) BW_TPV900 am BW_TPV900 pm
No ACL, No Qos 45 Kbps 3 80 M 70 M
Time based ACL–QoS 3 Kbps 0 150 M 120 M
Note: ACL, access control list; QoS, Quality of Service.

lighted by Smith (2019). These authors have pointed out that Fig. 8. Packet loss on the network. ACL, access control list.
decision trees have proven to be valuable tools for modeling
and anticipating patterns in network traffic, thus providing Packet loss on the network
a deeper understanding of congestion schedules. 200
The decision tree-based approach implemented in this 180
160
study aligns with the research of Smith (2019), who empha- 140
sized that this technique is particularly effective for predic- 120
tion in the field of networks due to its ability to handle com- 100
80
plex and variable datasets. 60
The prediction results identified two periods of network 40
congestion: one in the morning and another in the afternoon. 20
0
The evaluation of the model’s accuracy, reaching a value of

09:00
00:00
01:00
02:00
03:00
04:00
05:00
06:00
07:00
08:00

10:00
11:00
12:00
13:00
14:00
15:00
16:00
17:00
18:00
19:00
20:00
21:00
22:00
23:00
0.655, was based on using 20% of the total data to train the al-
gorithm, following the methodology adopted by Wang et al. Average Traffic Appliying ACL model No ACL model
2018. These researchers highlighted the importance of deci-
sion tree pruning to reduce model complexity, decrease vari-
ance, and improve the predictive capacity of the system.
In line with the reviewed literature, these findings support posed model is attributable to the fact that APs do not un-
the conclusion that the implementation of the decision tree dergo shutdown when experiencing congestion.
technique, complemented by pruning and data analysis over The variables that were taken to measure the level of con-
an extended period, stands out as an effective strategy to fore- gestion in the dense WLAN network were: the number of
see and efficiently manage congestion moments in WiFi net- users associated with the APs and the bandwidth used by the
works. This approach significantly contributes to improved network. Some authors (Awad and Abuhasan 2016) decide to
planning and optimization of system performance, provid- take only the number of users associated with the network
ing valuable insights for decision-making in wireless network APs, this can be a source of error. There may be many asso-
management (Manzanares-Lopez et al. 2018). ciated users in the WLAN network, but without consuming
With the implementation of time-based ACLs, there are bandwidth, or there may be a few associated users consum-
considerable changes in the number of packets lost and the ing excessive bandwidth. For this reason, it is convenient to
number of APs blocked in times of congestion in the dense have at least two congestion indicators. Having many vari-
WiFi network. These values are observed in Table 2. If a QoS ables to obtain an indicator of network congestion increases
system is not implemented on the UTS WiFi network to dy- efficiency, but at a high cost of system resources.
namically allocate bandwidth to each of its VLANs, there is a There are many algorithms in machine learning that allow
high percentage of lost packets every 5 min. Approximately the prediction of times and congestion levels of the WiFi net-
45 kbps is discarded due to traffic congestion. Due to net- work. The decision tree method presented high precision val-
work congestion, 3 APs restart every 15 min. This restarting ues and varied the number of congestion periods depending
of wireless equipment in the network increases packet loss on the pruning of the tree. If the pruning of the decision tree
by a few kbps and significantly reduces network availability. is decreased, the sensitivity of the system increases and there
When our dynamic QoS system is implemented in the net- are a greater number of congestion periods.
work using TBACL, the number of lost packets decreases sig- We use time-based ACLs for bandwidth distribution
nificantly (3 kbps), and there are no blockages or AP restarts through QoS settings on the router. This type of ACL was also
in the network. used to restrict internet access for non-priority traffic in some
As illustrated in Fig. 8, the packet loss behavior in the UTS periods of congestion time. Unlike other authors (Ornatelli et
WiFi network is depicted, both with and without the ACL al. 2021), it was possible to establish different periods of con-
model. The packet loss occurs during periods of traffic con- gestion or schedules to perform load balancing in the WLAN
gestion, i.e., during the periods in which the institution offers network.
the greatest number of classes. The implementation of the Our proposal can be replicated without any problem, do-
ACL model implies a reduction of packet loss in the network ing load balancing in a dense WiFi network in SDN networks
by up to 50%. (Manzoor et al. 2020). We carried out the acquisition of the
Typically, APs undergo a shutdown and restart cycle when congestion variables through software installed in the UTS
experiencing congestion due to high traffic. The significant network that served as a probe to collect the associated user
reduction in packet loss observed in the presence of the pro- data and bandwidth of the dense WiFi network. The compar-

FACETS 10: 1–9 (2025) | dx.doi.org/10.1139/facets-2024-0128 7

 Canadian Science Publishing

ison of packet loss measurements, number of blocked APs, tion in any medium, provided the original author(s) and
transmission delay times and bandwidth allow determining source are credited.
an improvement in load balancing and improvement in pri-
ority traffic of each of the VLANs of the WiFi network. Data availability
The technical requirements for deployment in other dense The network traffic data in Mbps and the number of devices
WiFi networks are: routing devices that support temporary on the network are public. However, information about the
ACLs or SDN devices; and to guarantee the sampling require- type of traffic, AP locations and other security information is
ments described in Section 3.2. private to ensure data privacy and compliance with ethical
standards and institutional regulations.
5. Conclusions
Our research introduces an innovative perspective for ef- Author information
ficiently managing bandwidth distribution in densely popu-
lated WiFi network environments. Focusing on load balanc- Author ORCIDs
ing and congestion mitigation, our approach incorporates Ricardo Alvarado https://orcid.org/0000-0003-3096-2174
machine learning, specifically the decision tree technique.
The dynamic bandwidth allocation method adapts in real- Author contributions
time to congestion periods, ensuring that prioritized traffic Conceptualization: RA, BO, JT, VT
receives optimal bandwidth. We have observed significant Data curation: RA, BO, JT, VT
improvements in bandwidth distribution and the speed of Formal analysis: RA, BO, JT, VT
prioritized data traffic in two different VLANs within a dense Funding acquisition: RA, JT, VT
WiFi network. Investigation: RA, BO, JT, VT
These enhancements stem from configuring QoS through Methodology: RA, BO, JT, VT
TBACL or time-based ACLs. The predictable daily behavior of Project administration: RA, BO, JT, VT
traffic in the network facilitates a robust predictive capabil- Resources: RA, BO, JT, VT
ity. Each decision tree in our model demonstrated an accu- Software: RA, BO, JT, VT
racy of 0.655, and we verified a reduction in packet loss for Supervision: RA, JT, VT
prioritized traffic in the WiFi network VLANs. Validation: RA, BO, JT, VT
Notably, the number of blocked APs decreased consider- Visualization: RA, BO, JT, VT
ably during dense WiFi network congestion periods. Packet Writing – original draft: RA, JT, VT
loss was reduced by over 50%, and AP blocks were entirely Writing – review & editing: RA, JT, VT
eliminated, significantly improving network availability.
The established QoS exhibits dynamic behavior, enhancing Competing interests
WiFi network efficiency during congestion periods for the The author shave declared that no competing interests exist.
two utilized VLANs. The employed machine learning tech-
niques demonstrated satisfactory predictive capabilities, sup-
porting decision-making in QoS configuration. References
Awad, M., and Abuhasan, A. 2016. A smart clustering based approach to
dynamic bandwidth allocation In wireless networks. International
Acknowledgements Journal of Computer Networks and Communications, 8(1): 73–86.
The authors would like to thank Unidades Tecnológicas de doi:10.5121/ijcnc.2016.8106.
Cao, Y., and Ai, L. 2022. Experimental simulation and comparative anal-
Santander for their financial and in-kind support in this re- ysis of an access control list at different deployment locations. In
search. 2022 IEEE 2nd International Conference on Computer Communi-
cation and Artificial Intelligence (CCAI). pp. 115–120. doi:10.1109/
CCAI55564.2022.9807771.
Article information Chang, Y.-C., Chen, Y.-C., Chen, T.-H., Chen, J.-L., Chiu, S.-P., and Chang,
W.-H. 2019. Software-defined Dynamic bandwidth management. In
2019 21st International Conference on Advanced Communication
Editor Technology (ICACT). pp. 201–205. doi:10.23919/ICACT.2019.8702052.
Parminder Raina Farooq, F., Ahmed, W., Akbar, A., Aslam, F., and Alyousef, R. 2021. Pre-
dictive modeling for sustainable high-performance concrete from in-
dustrial wastes: a comparison and optimization of models using en-
History dates semble learners. Journal of Cleaner Production, 292: 126032. doi:10.
Received: 28 June 2024 1016/j.jclepro.2021.126032.
Accepted: 22 April 2025 Hypolite, J., Sonchack, J., Hershkop, S., Dautenhahn, N., DeHon, A., and
Version of record online: 31 July 2025 Smith, J.M. 2020. DeepMatch: practical deep packet inspection in the
data plane using network processors. In Proceedings of the 16th In-
ternational Conference on Emerging Networking Experiments and
Copyright Technologies. pp. 336–350. doi:10.1145/3386367.3431290.
Ibrar, M., Wang, L., Muntean, G.-M., Akbar, A., Shah, N., and Malik, K.R.
© 2025 The Author(s). This work is licensed under a Creative
2021. PrePass-flow: a machine learning based technique to minimize
Commons Attribution 4.0 International License (CC BY 4.0), ACL policy violation due to links failure in hybrid SDN. Computer
which permits unrestricted use, distribution, and reproduc- Networks, 184: 107706. doi:10.1016/j.comnet.2020.107706.

8 FACETS 10: 1–9 (2025) | dx.doi.org/10.1139/facets-2024-0128

 Canadian Science Publishing

Li, G., Ren, L., Fu, Y., Yang, Z., Adetola, V., Wen, J., et al. 2023. A critical re- com Technologies and Optimization (Trends and Future Directions)
view of cyber-physical security for building automation systems. An- (ICRITO). pp. 1–6. doi:10.1109/ICRITO51393.2021.9596289.
nual Reviews in Control, 55, 237–254. doi:10.1016/j.arcontrol.2023. Singh, H. 2015. Performance analysis of unsupervised machine learn-
02.004. ing techniques for network traffic classification. In 2015 Fifth Inter-
Manzanares-Lopez, P., Malgosa-Sanahuja, J., and Muñoz-Gea, J. 2018. A national Conference on Advanced Computing and Communication
software-defined networking framework to provide dynamic QoS Technologies. pp. 401–404. doi:10.1109/ACCT.2015.54.
management in IEEE 802.11 networks. Sensors, 18(7): 2247. doi:10. Smith, A.J. 2019. Predictors of active and remission substance use disor-
3390/s18072247. der: A decision tree analysis. Texas Tech University
Manzoor, S., Chen, Z., Gao, Y., Hei, X., and Cheng, W. 2020. Towards Struhar, V., Ashjaei, M., Behnam, M., Craciunas, S.S., and Papadopoulos,
QoS-aware load balancing for high density software defined wi-fi A.V. 2019. DART: Dynamic bandwidth Distribution Framework for
networks. IEEE Access, 8: 117623–117638. doi:10.1109/ACCESS.2020. virtualized software defined networks. In IECON 2019 - 45th Annual
3004772. Conference of the IEEE Industrial Electronics Society. pp. 2934–2939.
Manzoor, S., Karmon, P., Hei, X., and Cheng, W. 2020. Traffic aware doi:10.1109/IECON.2019.8927780.
load balancing in software defined WiFi networks for healthcare. In Syrigos, I., Sakellariou, N., Keranidis, S., and Korakis, T. 2019. On the
2020 Information Communication Technologies Conference (ICTC). employment of machine learning techniques for troubleshooting
pp. 81–85. doi:10.1109/ICTC49638.2020.9123278. WiFi networks. In 2019 16th IEEE Annual Consumer Communica-
Ornatelli, A., Tortorelli, A., Giuseppi, A., and Priscoli, F.D. 2021. Hierar- tions and Networking Conference (CCNC). pp. 1–6. doi:10.1109/CCNC.
chical RL for load balancing and QoS management in Multi-access 2019.8651823.
networks. In 2021 29th Mediterranean Conference on Control and Wang, B., Liu, F., and Lin, W. 2021. Energy-efficient VM scheduling based
Automation (MED). pp. 886–891. doi:10.1109/MED51440.2021. on deep reinforcement learning. Future Generation Computer Sys-
9480246. tems, 125: 616–628. doi:10.1016/j.future.2021.07.023.
Radhika, D., and Duraipandian, M. 2021. Load balancing in cloud com- Wang, X., Lv, J., Huang, M., Li, K., Li, J., and Ren, K. 2018. Energy-efficient
puting using support Vector machine and optimized dynamic task ICN routing mechanism with QoS support. Computer Networks, 131:
scheduling. In 2021 9th International Conference on Reliability, Info- 38–51. doi:10.1016/j.comnet.2017.12.002.

FACETS 10: 1–9 (2025) | dx.doi.org/10.1139/facets-2024-0128 9