Machine Learning based Frame Size Optimization in WLAN Downlink MU-MIMO Channel with the Least Cost of Delay

International Journal of Computer Networks & Communications (IJCNC) Vol.14, No.6, November 2022
DOI: 10.5121/ijcnc.2022.14601 1
MACHINE LEARNING BASED FRAME SIZE
OPTIMIZATION IN WLAN DOWNLINK MU-MIMO
CHANNEL WITH THE LEAST COST OF DELAY
Lemlem Kassa1
, Jianhua Deng1
, Mark Davis2
and Jingye Cai1
1
School of Information and Software Engineering,
University of Electronic Science and Technology China (UESTC),
Chengdu 610054, China
2
Communication Network Research Institute (CNRI),
Technological University Dublin, D08 NF82 Dublin, Ireland
ABSTRACT
The key enhancement in the medium access control (MAC) layer is frame aggregation introduced by the
IEEE 802.11n/ac standard to accommodate the growing traffic demand in the WLAN by allowing multiple
packets aggregated per transmission. Frame aggregation efficiently reduces control overhead in the MAC
layer, such as the MAC header and thus it helps to enhance transmission efficiency and throughput
performance of WLAN. However, heterogeneous traffic demand among streams in the WLAN downlink
MU-MIMO channel creates a challenge to efficiently utilize the benefits of frame aggregation.
Transmission efficiency is also compromised during frame size setting determination because when a
frame size is larger, the impact of the overhead frame can be lower, but they are also more vulnerable to
transmission errors. Thus, this trade-off between maximizing frame size and minimizing overhead frames
should be addressed by employing an adaptive frame aggregation technique to derive the optimal frame
size that would maximize the throughput in WLAN downlink MU-MIMO channel. Moreover, when frame
aggregation approach is employed, more frames must wait before transmission in a buffer which causes a
delay in the performance of WLAN. Thus, analysing the trade-off between maximizing throughput and
minimizing delay is a critical issue that should also be addressed to enhance the performance of WLAN.
However, the majority of the existing adaptive aggregation algorithms in the WLAN downlink MU-MIMO
channel are focused to maximize the throughput or minimize the delay. The main contribution of this paper
is to propose a machine learning-based frame size optimization algorithm by extending our earlier
approach in considering the cost of delay to maximize the system throughput of WLAN. The effectiveness of
the proposed scheme is evaluated over the FIFO Baseline Approach and earlier conventional approaches
under the effects of various traffic patterns, channel conditions, and the number of STAs.
KEYWORDS
Delay, downlink MU-MIMO, frame size optimization, WLAN, heterogeneous traffic, machine learning,
neural network, throughput optimization.
1. INTRODUCTION
Wireless technology has been continuously evolving to support basic coverage and satisfy the
advanced needs of today’s network [1]. To improve the system performance of WLAN, IEEE
802.11 introduced single-user MIMO (SU-MIMO) [2] from IEEE 802.11n [3] and multi-user
multiple in multiple out (MU-MIMO) in IEEE802.11ac [2-4]. SU-MIMO transmission is the
conventional approach that allows multiple transmission of data streams to a single station. The
SU-MIMO WLAN is effective if the connecting stations are equipped with multiple receiving

2
antennas. However, in a WLAN that connects many stations, due to the limitation of receiving
antennas at receiving stations, SU-MIMO cannot efficiently utilize the full advantage of multiple
transmission antennas found at access points (APs) [2]. As a result, this situation causes
degradation in the system performance in terms of system capacity and user throughput. As a
consequence of this background, by adopting MU-MIMO, simultaneous data transmission to
multiple stations (STAs) is achieved by forming virtual spatial transmission channels between the
AP and the multiple receiving STAs [2].
The other key enhancement in the medium access control (MAC) layer is frame aggregation
introduced by the IEEE 802.11n/ac to accommodate the growing traffic demand in the network.
By using frame aggregation technique, multiple packets can be aggregated for a single
transmission thus this reduces overhead frame consumption in the MAC layer, such as the MAC
header. This approach provides a throughput enhancement and transmission efficiency in WLAN
[5-8]. However, heterogeneous traffic demand among streams in the WLAN downlink MU-
MIMO channel creates a challenge to efficiently utilize the benefits of frame aggregation. This is
because of that, when shorter and longer streams are grouped in downlink MU-MIMO
transmission, wasted space channel time will occur which is a time duration where a part of
spatial streams carries a data frame whereas the others do not [9-11]. To accommodate the
airtime of the shorter streams, frame pad bits, called padding bits, can be added at the tail of the
shorter stream until it has the same transmission duration as the longer streams. However, this
approach severely degrades transmission efficiency [11]. Transmission efficiency is also
analyzed in terms of errors occurring during frame size setting because when a frame size is
larger the impact of the overhead frame can be lower, but they are also more vulnerable to
transmission errors. This trade-off is addressed by frame aggregation techniques to derive the
optimal frame size that would maximize transmission efficiency. The IEEE 802.11n/ac standard
introduces two basic aggregation methods: the aggregated MAC service data unit (A-MSDU) and
the aggregated MAC protocol data unit (A-MPDU) [9,12]. Thus, to enhance the system
throughput and transmission efficiency in every MU-MIMO transmission, signaling overhead
and frame error rate should also be reduced. Moreover, there is always a trade between
maximizing throughput and minimizing delay when frame aggregation is adopted [12-14]
because more frames must wait before transmission in a buffer. In this regard, the adaptive
aggregation mechanism (AAM) proposed by [14] studied that there is a trade-off between
maximizing throughput and minimizing delay when frame aggregation is adopted. This study
considered the varying nature of the packet size and the packet arrival time to assemble the target
aggregate packet size in considering the minimum delay. In the work [13], the delay lower limit
(DLL) in the IEEE 802.11 wireless network was studied using the DCF model. However, this
study considered best-case scenarios such as an ideal channel condition with only one active
station that always has frames to send and receive frames and send acknowledgments (ACKs).
However, both [13] and [14] were considered single-user WLAN scenarios.
In the case of an MU-MIMO-enabled WLAN, the majority of the existing studies have only
attempted to maximize the throughput or minimize the delay. For instance, work [16] proposed a
machine learning-based frame size optimization to improve the system throughput performance
of WLAN in the downlink MU-MIMO channel. Different channel conditions, traffic patterns,
and number of competing stations are considered in this study. However, the effect of delay is not
studied while maximizing the throughput. But the trade-off between maximizing throughput and
minimizing delay needs to be considered when frame aggregation approaches are adopted [14].
The main contribution of this paper is to propose a machine learning-based frame size
optimization algorithm by extending our earlier approach in considering the cost of transmission
delay. In this approach, the optimal frame size setting is achieved by adopting an adaptive
aggregation approach that can cope with the dynamic effects of the traffic pattern and channel
conditions.

3
Due to the increasing complexity of wireless networks along with unsophisticated deployment,
distributed management systems, and density of network systems, it is becoming a challenge to
efficiently operate the IEEE 802.11 networks. The dominant approach to solving these
performance-related problems is to apply machine learning (ML). ML is a type of artificial
intelligence where algorithms can learn from training data without being explicitly programmed
[15]. There are two types of learning strategies in machine learning such as online learning (also
called incremental learning) and offline learning (or batch learning) [15 -17]. In online learning,
the algorithm updates its parameters after learning from each training instance which allows the
learning algorithm to keep learning on the fly [16,17]. Our proposed ML approach adopts an
online learning approach to cope with the time-varying channel conditions and traffic patterns. In
this approach, the Access Point (AP) collects the “frame size-system throughput patterns” which
contain knowledge about the effects of transmission delay), traffic condition, channel condition,
and a number of stations (STAs). Based on these patterns, the neural network is used to correctly
model the system throughput as a function of the system frame size. Once the training is
completed by the neural network, the gradient information obtained is used to adjust the system
frame size. Finally, the effectiveness of the proposed scheme is validated under various traffic
patterns, channel conditions, and the number of STAs for WLAN downlink MU-MIMO
channels.
The rest of the paper is organized as follows, in Section 2, we introduce related works about the
frame aggregation schemes and the performance challenges of multi-user transmissions in the
WLAN downlink MU-MIMO channel. The description of the proposed approach is discussed in
Section 3. In Section 4, results and discussions are presented to evaluate the performance of the
proposed approach under various channel conditions, traffic models, and a number of stations.
Finally, the conclusions are given in Section 5.
2. RELATED WORK
In this section, some previous studies on frame size optimization problems in WLAN will be
discussed. According to [16], a machine learning-based frame size optimization approach is
proposed aiming to improve the system throughput performance of WLAN in the downlink MU-
MIMO channel. In this approach different channel conditions, traffic patterns, and a number of
competing stations are considered to evaluate the performance of the proposed approach.
However, this approach did not study the effect of delay while maximizing throughput thus, it
provides a suboptimal solution. The work in [17] proposed a machine learning-based frame size
optimization approach in considering both channel conditions and contention effects of users to
maximize the throughput performance of WLAN. However, this approach similarly ignored the
issue of delay, and it can not operate in the IEEE 802.11 MU-MIMO-enabled WLAN. In
addressing the energy-throughput trade-off, [18] proposed an online learning-based frame
aggregation, Intelligent Energy-Efficient Frame Aggregation (IE2FA), to design an energy-
efficient MAC for high throughput wireless local area network (HT-WLAN). According to the
simulation results, they improved network performance significantly compared to the other
existing studies. Aiming to maximize goodput, [19] proposed an adaptive ML-based approach for
frame size selection on a per-user basis by taking into account channel conditions and global
performance indicators. The main approach of this study is to propose ML techniques in the
specific case of Software-Defined Wireless Local Area Networks (SD-WLAN) particularly
focusing on frame length optimization. According to the results, by analyzing a multitude
scenarios, an average improvement of 18.36% is achieved in goodput over standard aggregation
mechanisms. However, the issue of delay is not addressed in both [18] and [19]. [20] proposed a
frame-aggregation size determination approach in the WLAN downlink MU-MIMO channel to
improve channel utilization. This approach considered the issue of delay for frames waiting in

4
transmission queues and attempted to reduce the delay by appropriately determining the
aggregation size according to the traffic variation. However, the authors did not elaborate on the
effects of channel errors, and different traffic situations, also the study mainly focused to enhance
channel utilization. The work in [21] proposed an adaptive aggregation algorithm in the downlink
MU-MIMO channel in considering the issue of minimizing the cost of increased delay. This
approach considers both queue delay and transmission delay to improve the system throughput
performic of WLAN. According to the results, maximum throughput performance is achieved
with a minimum delay over the baseline FIFO algorithm evaluated under the effects of traffic
patterns, channel conditions, and a number of stations. The main difference between our
approach and [21] is that we adopted a machine learning-based frame size optimization solution
to enhance the system throughput of WLAN in considering the cost of transmission delay in
particular.
According to [22], queueing length and number of active nodes have significant impacts on the
system throughput performance. Thus, their frame size-based frame aggregation scheme achieved
the maximum system throughput performance by generating the same frame length in all spatial
streams. However, the limitation of this approach is that the issue of delay is ignored, and the
frame aggregation policy is not adaptive to the change in the conditions of traffic and channel
conditions. By employing a novel user selection criterion that provides a high priority to the
stations expecting high throughput in the next MU-MIMO transmission and having a large
amount of data to reduce signaling overhead, [23] proposed a frame duration-based frame
aggregation scheme. The main approach of this study is equalizing the transmission time of all
spatial streams in all streams according to their Modulation and Coding (MCS) level thus, the
maximum system throughput performance is achieved by minimizing the space channel time in
the WLAN downlink MU-MIMO channel. However, the main focus of this approach is to
improve the system throughput of WLAN without studying the issue of delay. Moreover, in
focusing on the effects of adding a padding bit, [24,25] provides a strategy of replacing padding
bits with data frames from other users in one stream, thus the space of frame padding bits will be
filled by important frames aiming to improve the transmission efficiency. However, this approach
violates the rules of MU transmissions by increasing the complexity of both the transmission and
reception process to allow the transmission to multiple destinations within a special stream.
According to the literature, there is little research exploring the use of ML techniques to tackle
frame size optimization problems in WLAN, particularly in considering the trade-off between
maximizing the throughput and minimizing delay.
3. PROPOSED APPROACH
Aiming to tackle the effects of heterogenous traffic demand among streams in WLAN downlink
MU-MIMO, a machine-learning-based adaptive approach is proposed for frame-size optimization
that attempts to maximize the system throughput performance of WLAN in minimizing delay.
Different traffic models are adopted such as (Pareto, Weibull, or fractional Brownian Motion
(fBM) [9, 10] to generate different traffic scenarios. The AP collected the data set as a pattern of
“frame size–system throughput “which contains knowledge about the effects of traffic patterns,
channel conditions, number of stations, and minimum transmission delay using the simulation
environment [21] which considers both transmission delay and queue delay. However, the data
collection in this study particularly considers transmission delay. Transmission delay is defined in
this study as the time duration a station takes from sensing the channel state i.e., idle or busy to
transmitting a frame until it receives the Acknowledgment (ACK) of the frame [19]. A specific
aggregation algorithm cannot always contribute to the maximum throughput performance in
minimizing transmission delay under the actual dynamic conditions of heterogeneous traffic
patterns and channel conditions of WLAN. In addressing these challenges, different frame

5
aggregation policies such as (FIFO FA (baseline approach), Equal Frame Size FA, Equal MPDUs
Agg FA, and Avg Num MPDUs FA) are adopted [9,10, 19] in this experiment to achieve the
adaptive aggregation strategy. Moreover, the time required for completing a single MU-MIMO
transmission depends on the types of aggregation policies employed. Thus, the adaptive
aggregation approach is significant in determining the minimum delay. In downlink MU-MIMO
transmission, the longer frame determines the transmission delay in a single MU- MIMO
transmission [19] which is defined as max (TData) in formula (1) which presents the
mathematical expression used for the transmission delay considered in this study.
(1. (1)
Therefore, in the simulation, the average system transmission delay ( ) is
computed per MU-MIMO transmission as the ratio of the sum of the transmission delay and the
number of stations, as illustrated in formula (2). Where n represents a number of stations in the
network and t is the simulation time.
(2)
The average delay of each aggregation policy is evaluated to determine the optimal aggregation
policy that provides the minimum system delay. Thus, the optimal aggregation policy is always
the one that provides the minimum transmission delay [21].
Under the actual dynamic effects of channel conditions and traffic patterns, it is difficult to obtain
an accurate throughput function f(frm) in all network conditions. The main contribution of this
study is to study how the throughput Thr can be maximized by optimizing the frame size frm. To
solve such an optimization problem the well-known gradient ascent algorithm is adopted in this
study that can build the knowledge and accurately model the throughput Thr as a function of the
frame size frm. After the knowledge building, the gradient information obtained from the neural
networks is used to adaptively adjust the frame size based on the gradient information. In the
formation of the frame-size optimization problem, the throughput Thr is a complex function of
the frame size frm under the conditions of minimum delay, channel conditions, traffic patterns,
and a number of stations. Formula (3) shows the throughput function f varies to the frame size
due to the above-specified changing conditions.
(3)
Thus, the local maximum of the throughput function Thr = f(frm) can be found by adaptively
adjusting frame size frm using gradient ascent, by taking steps that are proportional to the
gradient. For instance, at the nth
time of adjustment, the frame size was frm(n), and the throughput
was Thr(n). At the next time of adjustment, the frame size frm will be set as shown in formula
(4).
(4)
Where depends on the gradient of the estimated throughput with respect to
, i.e.,
(5)

6
The parameter μ is a variable adjustment rate heuristically selected for different network
scenarios.
To cope with the effects of time-varying channel conditions and heterogeneous traffic patterns,
the online machine learning strategy is employed to achieve the data collection, knowledge
building, and frame-size adjustment kept online. The proposed MLP ML approach consists of
one hidden layer with four neurons and an output layer. The backpropagation algorithm only
consists of two passes: 1) a forward pass and 2) a backward pass [16,17]. To obtain the gradient
information which is used to adjust the frame size, we add a third pass, i.e., the
tuning pass that will be discussed in Section 3.1. In the proposed MLP approach, the
backpropagation algorithm is used to adjust the network and minimize the error between the
actual response and the desired (target). The description and summary of notations used in this
study are presented in Table 1.
Table 1. Simulation Parameter and Notation Summary
Parameters Symbol Value
# Of Antenna at AP NAnt 4
# Of Stations NumSTA 2–4
VoIP traffic payload size 100Byte
Video traffic payload size 1000Byte
Learning Rate ɳ 0.5
Mean Square Error Threshold MES 0.00001
Epoch Threshold 1000 times
Activation Function Sigmoid (σ)
Number of training patterns n
Indices of neurons in different layers i, j
Frame size(input) of nth
training pattern frm(n)
Target response for neuron j Thr(n)
Actual response of the nth
training pattern
Synaptic weight in layer l connecting the output
neuron of i to the input neuron j at iteration n
Weight sum of all synaptic inputs plus bias of
neuron j in layer l at iteration n.
Signal of output of neuron j in layer l at iteration n
Local gradient of neuron j in layer l in the tuning
pass of hidden layer
)
Local gradient of neuron j in layer l in the tuning
pass of the output layer
Adjustment rate µ
3.1. Tuning Pass Strategy
The diagram shown in Figure 1 illustrates the signal flow of the tuning pass in the machine
learning model to estimate the gradient and the key to adjusting the frame size to
maximize the throughput in minimizing delay. The initial weight is denoted as in the neural
network is randomly chosen. The synaptic weights that have been well adjusted in the backward
pass are set as fixed in the tuning pass. An adaptive learning rate is adopted to improve the
convergence speed [16].

7
Figure 1. Flow chart of the proposed machine learning approach consisting of tuning pass, which depicts
the derivation of the local gradients and the gradient for frame size adjustment.
In the following discussion, the procedure how the estimated gradient is computed by
adopting [16]. Considering the hidden layer, the local gradient for the tuning pass is
defined as follows:
(6)
Where in formula (6) is the weight sum of synaptic input plus bias of neuron j in layer l.
Similarly, considering the output layer, the local gradient is defined as follows:
(7)
While considering the hidden layer, the local gradient can be expressed as follows using
the chain rule:
(8)

8
Therefore, using results (7) and (8), the gradient can be written as follows:
(9)
Where can be defined as . Thus, the second term at the
rightmost side of equation (9) can be written as:
(10)
Therefore, the gradient (11)
The derivation of the local gradients at each layer and the gradient is depicted in Figure
1. Based on the result from equation (11), the frame size frm is adjusted as shown in formulas (4)
and (5).
4. RESULTS AND DISCUSSION
In this section, the experimental procedure is discussed. The performance of the proposed
machine learning-based adaptive approach will be evaluated by considering the effects of channel
conditions, heterogeneous traffic patterns, and a number of stations.
4.1. Experimental Procedure
This experiment is conducted to enhance the performance of our previous machine-leaning-based
adaptive approach for frame size optimization to maximize the system throughput performance of
WLAN in the downlink MU-MIMO channel [16]. However, this earlier work didn’t consider the
issue of delay. Thus, to collect the training data set as a pattern of “frame size - system
throughput”, we adopt the simulation environment [21]. Because, according to [21] both queue
delay and transmission delay are considered in the simulation, but in this experiment, the data
collection is conducted only by considering the transmission delay. The frame size in the data set
represents the input data set that represents the average offered traffic load generated in the
network to obtain the corresponding target system throughput. System throughput is the target
data set that determines the maximum system throughput values achieved [21]. The training data
set is collected once every 50 seconds in considering different network scenarios such as channel
conditions, traffic patterns, number of stations, and the minimum transmission delay. Thus 50

9
samples will be collected for each training to train the neural network until the stopping criteria
of Mean Square Error (MES) fall below 0.00001 or when the training epoch exceeds 1000 times
is satisfied by using Forward and backward passes iteratively. Then the weight is updated
following the procedure in the backward pass. The error threshold and the maximum number of
iterations determine the accuracy of the function and the computing cost. Finally, once the
training is over and the knowledge-building model is achieved, the tuning pass will be invoked to
adjust the frame size frm by using gradient information from the neural network.
We adopted the same simulation parameters considered in [16] as we are attempting to extend the
limitation of this earlier work in considering the issue of delay which is one of the significant
performance evaluation metrics in networking. To evaluate the performance of the proposed
approach, different comparative approaches in terms of system throughput performance are
considered in the experiment such as FIFO FA (Baseline Approach) which does not employ an
adaptive aggregation scheme, Adaptive FA Conv. Approach 1 [9] is an adaptive aggregation
algorithm that did not consider channel error and the cost of delay aiming to achieve only
maximize the throughput and Adaptive FA Conv. Approach 2 [10] is an adaptive approach that
can achieve the maximum system throughput by considering the effects of channel error and the
cost of increased delay. However, both [9, 10] didn’t adopt a machine learning-based
optimization solution. The abbreviations ‘FA’ and ‘Conv.’ in this paper refer to ‘frame
aggregation’ and ‘conventional’ respectively. Thus, the performance of our proposed ML
approach will be evaluated over the FIFO FA (Baseline Approach), Adaptive FA Conv.
Approach 1, and Adaptive FA Conv. Approach 2.
In general, the proposed machine-learning-based adaptive approach will be evaluated under the
following performance factors. In Section (4.2) the performance of the proposed ML approach is
evaluated under the effects of different traffic models such as Pareto, Weibull, and fBM. In
Section (4.3) the performance of the proposed approach is evaluated under the effect of channel
conditions where SNR= 5, 10, and 20 dB. In Section (4.4) the performance of the proposed
approach under a varying number of STAs (2,3,4) is evaluated. Finally, the performance of the
proposed ML approach is evaluated in terms of system throughput versus optimal system frame
size in Section (4.5). All experiments are conducted with a traffic mix of 50% VoIP and 50% [10]
video with a constant frame size of 100 Byte and 1000 Byte, respectively.
4.2. Performance Under the Effect of Various Traffic Models
In this section, the performance of the proposed ML approach is evaluated under the effects of
different traffic models such as Pareto, Weibull, and fBM [9], SNR = 10 dB, and NumSTA= 4.
According to the result, the proposed approach achieved different performances in different
traffic models while the same number of stations and channel conditions are considered. For
instance, the proposed approach copes better with the Weibull traffic model with a maximum
performance of 807Mbps than the Pareto traffic model 667Mbps which is the least even when
compared with fBM 705Mbps. These results demonstrated how the optimal system throughput
performance of WLAN in the downlink MU-MIMO channel is affected by the heterogeneous
traffic patterns among spatial streams in different traffic conditions. The average maximum
system throughput achieved by the proposed ML approach, FIFO FA (Baseline Approach),
Adaptive FA Conv. Approach 1, and Adaptive FA Conv. Approach 2 under the conditions of
different traffic models is illustrated in Table 2.

10
Table 2. Simulation result achieved by the FIFO FA (Baseline Approach), Adaptive FA Conv. Approach 1,
Adaptive FA Conv. Approach 2, and Proposed ML approach for average system throughput performance
in Mbps under the effects of different traffic models.
Comparative Approaches Traffic Models
Pareto Weibull fBM
FIFO FA (Baseline Approach) 692.87875 397.50775 482.20575
Adaptive FA Conv. Approach 1 868.09675 695.20775 724.73675
Proposed ML Approach 807.872 667.94675 705.45175
0
100
200
300
400
500
600
700
800
900
1000
Weibull Pareto fBM
Average
System
Throughput
[Mbps]]
Traffic Models
Throughput vs Traffic Model
FIFO FA (Baseline Approach) Adaptive FA Conv. Approuch 1
Adaptive FA Conv. Approuch 2 Proposed ML Approach
Figure 2. Performance of average system throughput under the effects of heterogenous traffic models when
SNR = 10dB.
As the result shows in Figure 2, Adaptive FA Conv. Approach 1 achieved the maximum
performance in all traffic models as it was only concerned with the maximum throughput with the
expense of delay and ideal channel conditions. Similarly, Adaptive FA Conv. Approach 2
achieved a better performance than the proposed approach. But this approach doesn’t consider
the issue of delay even though transmission error is considered. However, the proposed ML
approach achieved the maximum performance of 807Mbps by the Weibull traffic model which is
closer to the performance 809Mbps with Adaptive FA Conv. Approach 2. The FIFO (Baseline
Approach) achieved the worst performance of all traffic models particularly the least
performance of 397Mbps achieved using the Pareto traffic model due to its non-adaptive
aggregation policy employed in it. Thus, these results indicate that traffic patterns in the network
determine the system's performance.
4.3. Performance Under the Effects of Channel Conditions
The performance of the proposed approach under different channel conditions when SNR = 5, 10,
and 20dB, and NumSTA =4 is evaluated in this section. When the SNR value increases the channel
quality improves thus the throughput performance is enhanced as the occurrence of transmission
frame error rate will be less. As shown in Figure 3 (a), (b), and (c), for the case of different traffic
models such as Pareto, Weibull, and fBM, the system throughput performance of all approaches

11
increases when the channel quality improved from 5dB to 20dB. However, the proposed ML
approach achieved better performance than that of the FIFO (Baseline Approach) due to the
adaptive aggregation approach adopted in the Proposed ML Approach. On the contrary, the
Proposed ML approach achieved the lowest performance of 673Mbps in the Pareto traffic model
compared with 696Mbps achieved by Adaptive FA Conv. Approach 1 and Adaptive FA Conv.
Approach 2 when SNR=20dB. This is due to the reason that both Adaptive FA Conv. Approach 1
and Adaptive FA Conv. Approach 2 approach only focuses on maximizing throughput with a cost
of maximum delay which provides a suboptimal solution. On the contrary, under the worst
channel condition, e.g. SNR = 5dB in the figure, a better performance of 330Mbps is achieved by
the Weibull traffic, and the worst 266Mbps is achieved by fBM traffic. This result illustrates that
the system throughput performance is affected by traffic patterns even under the same channel
condition. However, the proposed approach achieved maximum performance better than FIFO
FA (Baseline Approach) as it employed an adaptive aggregation approach to realize the optimal
system throughput. Table 3 illustrates quantitative performance results of the average system
throughput performances achieved by the FIFO FA (Baseline Approach), Adaptive FA Conv.
Approach 1, and Adaptive FA Conv. Approach 2 and Proposed ML Approach, under the effects
of different channel conditions and traffic models.
Table 3. Simulation result achieved by the Proposed ML Approach, Maximum Throughput, and FIFO
(Baseline Approach) for average system throughput performance in Mbps under the effects of different
traffic models and channel conditions.
Comparative Approaches
Traffic
Models
SNR (dB)
Weibull
5(dB) 10(dB) 20(dB)
Adaptive FA Conv. Approach 1 Pareto 695.207 695.207 695.207
Adaptive FA Conv. Approach 1 fBM 724.736 724.7367 724.7367

12
a) Weibull Traffic Model b) Pareto Traffic Model
c) fBM Traffic Model
Figure 3. System throughput versus SNR for different traffic models such as Pareto, Weibull, and fBM
when NumSTAs= 4.
FIFO FA (Baseline Approach) achieved the worst performance 181Mbps and 180Mbps in the
worst channel condition of 5dB by the Pareto and fBM traffic models respectively. In general, the
FIFO FA (Baseline approach) aggregation policy is the worst compared to the proposed approach
in all scenarios because of the non-adaptive aggregation strategy it employs. However, the
Proposed ML Approach always achieved the maximum performance close to Adaptive FA Conv.
Approach 1, and Adaptive FA Conv. Approach 2 only considers the achievement of maximum
throughput with the cost of maximum delay [9,10].

13
4.4. Performance Under the Effects of Number of Stations
Figure 4 (a), (b), and (c), demonstrate the performance of the proposed approach evaluated under
the effect of a different number of stations (NumSTA =2, 3, and 4), and when the channel
condition is SNR=10dB for the case of Weibull, Pareto, and fBM traffic models. As the results
show, the system throughput performance significantly increases in all traffic models as the
traffic rate increase when the number of stations ranges from 2 to 4. However, the performance of
the proposed ML approach achieved different performances in different traffic models due to the
effect of heterogeneous traffic patterns even under the same number of stations. The quantitative
comparative results of the average optimal system throughput achieved by the FIFO FA (Baseline
Approach), Adaptive FA Conv. Approach 1, Adaptive FA Conv. Approach 2, and the Proposed
ML Approach under the effects of a variable number of STAs are illustrated in Table 4.
Table 4. Simulation results achieved by comparative approaches such as FIFO FA (Baseline Approach),
Adaptive FA Conv. Approach 1, Adaptive FA Conv. Approach 2, and Proposed ML approach for average
system throughput in Mbps under the effects of variable number of stations in Weibull, Pareto, and fBM
traffic models.
Comparative Approaches
Traffic
Models
Number of Stations
Weibull
2 3 4
Adaptive FA Conv. Approach 1 Pareto 529.9385 529.9385 695.2077
Adaptive FA Conv. Approach 1 fBM 564.47 564.47 724.7367

14
0
100
200
300
400
500
600
700
800
2 3 4
Asverage
System
Throughput
[Mbps]
Number of STAs
Throughput vs. Number of STAs (fBM)
FIFO FA (Baseline Approach) Adaptive FA Conv. Approach 1
Adaptive FA Conv. Approach 2 Proposed ML Approach
Figure 4. System throughput versus number of stations when the channel condition is SNR =10dB for the
Weibull, Pareto, and fBM traffic models.
As shown in the results in Figure 4, when the number of stations increases in all traffic models,
the system throughput performance increases as the traffic load increases with an increasing
number of stations from 2 to 4. However, different performance is achieved depending on the
traffic condition. For instance, 807Mbps is the maximum performance achieved by the Weibull
traffic model while 667Mbps is achieved by the Pareto traffic model. The Adaptive FA Conv.
Approach 1 outperforms all approaches as it always considers maximizing throughput in ideal
channel conditions and the expense of delay. However, the proposed approach always
outperforms the FIFO (Baseline Approaches) in all scenarios due to the adaptive aggregation
strategy it adopts. In this regard, the maximum performance of 807Mbps in the case of the
Weibull traffic model is achieved by the proposed approach whereas the lower performance of
667Mbps is achieved in the Pareto traffic with the same number of STAs equals 4. Likewise,
when the number of stations equals 2, the maximum performance of 428Mbps is achieved by the

15
Weibull and Pareto traffic models and the least 365Mbps by the fBM traffic model. These results
show that number of stations affects the performance of the system throughput under the
conditions of heterogeneous traffic patterns among streams in the downlink MU-MIMO channel.
However, the proposed approach always achieved the maximum system throughput performance
better than the FIFO FA (Baseline Approach) but closest to the Adaptive FA Conv. Approach 1,
Adaptive FA Conv. Approach 2 [9,10] which are only considered the maximum throughput with
the expense of delay.
4.5. Performance of System Throughput Vs. Optimal Frame Size
Figure 5 (a), (b), and (c) show the performance of system throughput with increasing System
frame size considering SNR= 10 dB, NumSTA = 4. To examine the effect of traffic conditions on
the performance of the proposed ML approach, traffic models such as Weibull, Pareto, and fBM
are considered. As the results show, the throughput performance of the proposed ML approach
increased when the size of the frame increased. However different traffic models achieved
different performances due to the effects of heterogeneous traffic patterns. For instance, the
maximum performance of 874Mbps is achieved when the optimal frame size is 1Mbyte by the
Weibull traffic model whereas the 816Mbps is the maximum performance by the Pareto when the
frame size is 1Mbyte. Adaptive FA Conv. Approach 1 achieved the maximum performance in all
traffic models because this approach only focuses on maximizing the throughput with the
expense of maximum delay and free from transmission error. However, the FIFO FA (Baseline
Approach) achieved the worst performance among all traffic models as this algorithm doesn’t
employ an adaptive aggreging strategy to maximize the throughput. For instance, 437Mbps is the
least performance achieved by fBM traffic model when the frame size is 0.93Mbyte.

16
Figure 5. Performance of system throughput versus optimal System frame size when NumSTAs = 4 and SNR
=10dB for the Weibull, Pareto, and fBM traffic models.
In general, the proposed ML approach provides a significant performance in optimizing the
system frame size of WLAN to maximize the system throughput performance with a minimum
cost of delay. This is achieved by employing an adaptive aggregation approach by considering
channel conditions, traffic patterns, and a number of stations than the baseline FIFO FA (Baseline
Approach). Moreover, the proposed approach achieved the closest performance to that of the
Adaptive FA Conv. Approach 1 and Adaptive FA Conv. Approach 2 only considers maximizing
the throughput without considering the issue of delay.
5. CONCLUSIONS
The one of key enhancements is MAC layer frame aggregation introduced by the IEEE
802.11n/ac to accommodate the growing traffic demand in the network by allowing multiple
packets aggregated at once in a single transmission. By adopting frame aggregation, the overhead
in the MAC layer is controlled and a tremendous throughput enhancement is achieved. However,
due to the heterogeneous traffic demand among streams in the WLAN downlink MU-MIMO
channel, it is challenging to efficiently utilize the benefits of frame aggregation. Because when
shorter and longer streams are grouped in downlink MU-MIMO transmission, wasted space
channel time will occur which is a time duration where a part of spatial streams carries data
frames whereas the others do not thus it would degrade the transmission efficacy. Thus, the trade-
off between maximizing frame size and minimizing overhead frames should be addressed by
adopting an adaptive frame aggregation technique to derive the optimal frame size that would
maximize the throughput performance of WLAN. Delay is another critical issue that needs to be
taken into account when frame aggregation is employed because frame aggregation experiences
more frames waiting in a buffer before transmission which degraded the performance of WLAN.
The main contribution of this paper is to propose a machine learning-based frame size
optimization algorithm to improve the throughput performance of WLAN in downlink Mu-
MIMO channel by extending our earlier approach in considering the cost of delay. In this
approach, the optimal frame size setting in the WLAN downlink MU-MIMO channel is achieved
by adopting an adaptive aggregation approach that considers the dynamic effects of traffic

17
patterns and channel conditions. The effectiveness of the proposed approach is evaluated over the
FIFO Baseline Approach and earlier conventional approaches under various traffic patterns,
channel conditions, and a number of STAs for WLAN downlink MU-MIMO channels. Future
work will extend our approach by considering real traffic data. Study the effects of other delay
factors such as processing delay and queuing delay in considering both uplink and downlink
transmission channels of WLAN.
CONFLICT OF INTEREST
The authors declare no conflict of interest.
ACKNOWLEDGMENTS
We would like to thank all authors for their contributions and the success of this manuscript and
all editors and anonymous reviewers of this manuscript.
REFERENCES
[1] Mugen Peng, Wenbo Wang, and Hsiao-Hwa Chen, (2010) “TD-SCDMA evolution,” IEEE Veh.
Technol. Mag., vol. 5, No. 2, pp. 28–41.
[2] IEEE Std 802.11ac-2013; Part11: Wireless LAN Medium Access Control (MAC) and Physical Layer
(PHY) Specifications: Amendment 4: Enhancements for Very High Throughput for Operation in
Bands below 6 GHz, 2013. DOI:10.1109/IEEESTD.2013.6687187
[3] Caire Giuseppe, and Shlomo Shamai, (2000) "On achievable rates in a multi-antenna broadcast
downlink." In Proc. 38th Annu. Allerton Conf. Communication, Control, and Computing, pp. 1188-
1193.
[4] Jiang Ming, and Lajos Hanzo,(2007) "Multiuser MIMO-OFDM for next-generation wireless
systems." Proceedings of the IEEE 95, No. 7 pp: 1430-1469.
[5] Li Tianji, Qiang Ni, David Malone, Douglas Leith, Yang Xiao, and Thierry Turletti. (2009)
"Aggregation with fragment retransmission for very high-speed WLANs." IEEE/ACM transactions
on networking, No. 2, pp: 591-604.
[6] Lin Yuxia, and Vincent WS Wong. (2006) "WSN01-1: Frame aggregation and optimal frame size
adaptation for IEEE 802.11 n WLANs." In IEEE Globecom, pp. 1-6. IEEE.
[7] Skordoulis Dionysios, Qiang Ni, Hsiao-Hwa Chen, Adrian P. Stephens, Changwen Liu, and Abbas
Jamalipour. (2008) "IEEE 802.11 n MAC frame aggregation mechanisms for next-generation high-
throughput WLANs." IEEE Wireless Communications, Vol. 15, no. 1, pp: 40-47.
[8] Ong Eng Hwee, Jarkko Kneckt, Olli Alanen, Zheng Chang, Toni Huovinen, and Timo Nihtilä. (2011)
"IEEE 802.11 ac: Enhancements for very high throughput WLANs." In 2011 IEEE 22nd International
Symposium on Personal, Indoor and Mobile Radio Communications, pp. 849-853. IEEE, 2011.
[9] Lemlem Kassa, Mark Davis, Jingye Cai, and Jianhua Deng, (2021) "A New Adaptive Frame
Aggregation Method for Downlink WLAN MU-MIMO Channels." Journal of Communication, Vol.
16, No. 8, pp 311-322.
[10] Lemlem Kassa, Mark Davis, Jianhua Deng, and Jingye Cai, (2022) "Performance of an Adaptive
Aggregation Mechanism in a Noisy WLAN Downlink MU-MIMO Channel." Electronics, Vol. 11,
No. 5, pp 754.
[11] Kim Sanghyun, and Ji-Hoon Yun, (2019) "Efficient frame construction for multi-user transmission in
IEEE 802.11 WLANs." IEEE Transactions on Vehicular Technology, Vol. 68, No. 6, pp 5859-5870.
[12] Teymoori Peyman, Nasser Yazdani, Seyyed Alireza Hoseini, and Mohammad Reza Effatparvar,
(2010) "Analyzing delay limits of high-speed wireless ad hoc networks based on IEEE 802.11n"
In 2010 5th International Symposium on Telecommunications, pp. 489-495. IEEE.
[13] Xiao Yang, and Jon Rosdahl, (2022) "Throughput and delay limits of IEEE 802.11." IEEE
Communications letters, Vol. 6, No. 8, pp: 355-357.
[14] Jianhua Deng, Mark Davis, (2013) "An adaptive packet aggregation algorithm for wireless
networks." In Proceedings of the 2013 International Conference on Wireless Communications and
Signal Processing, Hangzhou, China, 24–26 October 2013; pp. 1–6.

18
[15] Szott Szymon, Katarzyna Kosek-Szott, Piotr Gawłowicz, Jorge Torres Gómez, Boris Bellalta,
Anatolij Zubow, and Falko Dressler. (2022) "Wi-Fi Meets ML: A Survey on Improving IEEE 802.11
Performance With Machine Learning." IEEE Communications Surveys & Tutorials, Vol.24, No. 3,
pp : 1843-1893.
[16] Lemlem Kassa, Jianhua Deng, Mark Davis, and Jingye Cai. (2022) "Frame Size Optimization Using a
Machine Learning Approach in WLAN Downlink MU-MIMO Channel."In the 8th International
Conference of Networks, Communications, Wireless and Mobile Computing (NCWC). arXiv preprint
arXiv:2209.09687, 2022, Copenhagen, Denmark.
[17] Lin Pochiang, and Tsungnan Lin, (2009) "Machine-learning-based adaptive approach for frame-size
optimization in wireless LAN environments." IEEE transactions on vehicular technology, Vol. 58,
No. 9 pp 5060-5073.
[18] Karmakar Raja, (2019) "Online learning-based energy-efficient frame aggregation in high throughput
WLANs." IEEE Communications Letters, Vol. 23, No. 4, pp: 712-715.
[19] Coronado Estefanía, Abin Thomas, and Roberto Riggio. (2020) "Adaptive ml-based frame length
optimization in enterprise SD-WLANs." Journal of Network and Systems Management, Vol. 28, No.
4, pp: 850-881.
[20] Moriyama Tomokazu, Ryo Yamamoto, Satoshi Ohzahata, and Toshihiko Kato, (2017) "Frame
aggregation size determination for IEEE 802.11 ac WLAN considering channel utilization and
transfer delay." ICETE Proc. 14th Int Jt Conf E-Bus Telecommun. Vol. 6, pp 89–94.
[21] Lemlem Kassa, Jianhua Deng, Mark Davis and Jingye Cai, (2022). Performance of WLAN in
Downlink MU-MIMO Channel with the Least Cost in Terms of Increased Delay. Electronics, ,
11(18), 2851.
[22] Bellalta Boris, Jaume Barcelo, Dirk Staehle, Alexey Vinel, and Miquel Oliver, (2012) "On the
performance of packet aggregation in IEEE 802.11 ac MU-MIMO WLANs." IEEE Communications
Letters, Vol. 16, No. 10, pp 1588-1591.
[23] Nomura Yoshihide, Kazuo Mori, and Hideo Kobayashi, (2016) "High-Efficient Frame Aggregation
with Frame Size Adaptation for Downlink MU-MIMO Wireless LANs." IEICE Transactions on
Communications, Vol. 99, No. 7, pp 1584-1592.
[24] Lin Chi-Han, Yi-Ting Chen, Kate Ching-Ju Lin, and Wen-Tsuen Chen, (2018) "Fdof: Enhancing
channel utilization for 802.11 ac." IEEE/ACM Transactions on Networking, Vol. 26, No. 1, pp 465-
477.
[25] Lin Chi-Han, Yi-Ting Chen, Kate Ching-Ju Lin, and Wen-Tsuen Chen, (2017) "acPad: Enhancing
channel utilization for 802.11 ac using packet padding." In IEEE INFOCOM 2017-IEEE Conference
on Computer Communications, pp. 1-9.
AUTHORS
Lemlem Kassa has been a senior lecturer in Addis Ababa Science and Technology
university in Addis Ababa, Ethiopia since February 2013. She is currently pursuing her
PhD study in School of Information and Software Engineering in University of
Electronic Science and Technology of China (UESTC) in China. She received the B.S.
degree from Micro link Information Technology college in Addis Ababa Ethiopia in
2006 in software engineering and M.S. degree from University Putra Malaysia (UPM),
Malaysia in 2009. Her research interests are in the area of wireless communications,
artificial intelligence, mobile computing, and software designing.
Dr. Jianhua Deng graduated in information security from the University of Electronic
Science and Technology of China (UESTC), China, in 2006. After graduated, he joined
the School of Computer Science and Engineering at UESTC as a staff. From 2009 to
2014, he was a Ph.D. student in Dublin Institute of Technology (DIT) in Ireland and
received Ph.D. degree in electrical engineering from DIT in 2014. Now, he is a vice
professor in the School of Information and Software Engineering at UESTC. His
research interests are in the area of wireless communication, statistical machine learning, artificial
intelligence, deep learning. He is the reviewer of some SCI journals (e.g. wireless personal
communications).

19
Prof. Mark Davis received his BE, MEngSc, and PhD degrees from University
College Dublin in 1986, 1989 and 1992 respectively. He is currently the director of
the Communications Network Research Institute at Technological University Dublin
(TU Dublin). His research interests are in the area of radio resource management
techniques for wireless networks, specifically IEEE 802.11 WLANs.
Prof. Jingye Cai is Professor and Associate Dean at University of Electronic Science
and Technology of China (UESTC). He received BS from Sichuan University in
1983 with a major in radio electronics and Ph.D. from the University of Electronic
Science and Technology of China in 1990, with major in signal and information
processing. His research interests are in the area of intelligent computing, information
engineering, digital Information Processing, and signal processing.

Machine Learning based Frame Size Optimization in WLAN Downlink MU-MIMO Channel with the Least Cost of Delay

More Related Content

Similar to Machine Learning based Frame Size Optimization in WLAN Downlink MU-MIMO Channel with the Least Cost of Delay (20)

More from IJCNCJournal (20)

Recently uploaded (20)

Machine Learning based Frame Size Optimization in WLAN Downlink MU-MIMO Channel with the Least Cost of Delay