Content uploaded by Xiaomin Ma
Author content
All content in this area was uploaded by Xiaomin Ma on Sep 05, 2022
Content may be subject to copyright.
Model-based Deep Learning Optimization of IEEE
802.11 VANETs for Safety Applications
Shengli Ding, and Xiaomin Ma,
School of Engineering, Oral Roberts University, Tulsa, OK 74171, USA
Emails: dingshengli@oru.edu, xma@oru.edu
Abstract— IEEE 802.11p/bd driven Vehicular Ad Hoc
Networks (VANETs) have been investigated for safety-critical
applications with high reliability and low transmission
latency. However, due to dynamic vehicular environment and
various safety applications requiring different quality of
service (QoS), a fixed configuration of the communication
network parameters performs poorly in terms of balance
between QoS and channel spectrum efficiency. This paper
proposes a new real-time optimization scheme based on a
designated deep learning neural network (DLNN) working
with a stochastic model. In the scheme, the stochastic model is
adopted to predict the QoS of VANET given a set of
communication parameters. The DLNN is trained to
approach the inverse maps from the parameter sets to the
corresponding QoS by a sampled data set from running the
stochastic model. In the process of optimization, for a given
safety service, the DLNN and the stochastic model
complement each other to find an optimal solution of the
parameters that maximize the channel efficiency under the
constraints of QoS requirements in a fast and precise way.
The experiments on Google Colab with TensorFlow and
Python demonstrate the effectiveness of the scheme.
Keywords— Optimization, Deep Learning, Ad hoc networks,
Quality of Service, Safety
I. INTRODUCTION
Vehicular Ad Hoc Networks (VANETs) equipped with
wireless communications capability have been proposed
and investigated to provide vehicle-to-everything (V2X)
services for improving safety on the road [1]. Up to date,
three types of communication network standards are
considered as the promising candidates for next generation
of V2X technologies: IEEE 802.11p [2], IEEE 802.11bd [3],
and 3GPP NR-V2X [4]. Due to the development of new
wireless technologies such as higher-rate modulation and
coding schemes (MCS) with high channel bandwidth, Low-
Density Parity Check (LDPC) error-correction coding, Dual
Carrier Modulation (DCM), adaptive transmission schemes,
multi-input-multi-output (MIMO), etc., IEEE 802.11 driven
VANETs have big potential to meet the stringent
requirements on quality of service (QoS) for assuring safety
of human driving vehicles as well as autonomous driving
vehicles. In this paper, we focus on IEEE 802.11 driven
VANETs for Basic Safety Message (BSM) services. There
have been articles introducing IEEE 802.11p and IEEE
802.11bd for the safety services and describing the detailed
evolution of IEEE 802.11bd from IEEE 802.11p [3] [4]. The
analyses of the performance and reliability of IEEE 802.11p
are reported in numerous papers [5-7]. Insightful
comparisons of the two standards were conducted in [8-10].
Recently, an analytical model was proposed to evaluate the
QoS of SINR-based IEEE 802.11 p/bd VANETs for BSM
safety services [11].
In IEEE 802.11p/bd driven VANETs, most
communication parameters (MCS index, MAC layer
contention window size, channel bandwidth, etc.) are fixed
once they are configured. However, the vehicular
communication environment (e.g., density of vehicles,
channel shadowing/fading characteristics, current safety
application type, etc.) is dynamically changed. The fixed
communication settings could cause poor QoS or inefficient
channel spectrum usage. In the updated IEEE 802.11bd
standard, a retransmission scheme that dynamically adapts
the number of message repetitions to the channel busy status
is suggested. However, this scheme operates without
accounting for if the QoS requirement is satisfied. On the
other hand, a few dynamic optimization schemes have been
proposed to enhance the current VANETs in the literature
[12-17]. Y. Yao et al. [13] proposed a density-aware rate
adaptation (DARA) protocol to ensure reliable vehicle
safety 802.11p driven communication in a highway
environment. DARA adjusts the bit rate only in response to
the packet loss due to channel fading and the change of the
traffic density. C. Christopher et al. [14] focused on adaptive
modulation to improve the data throughput and efficiency
of channel spectrum in VANET. A. Bazzi [15] investigated
congestion control management in IEEE 802.11p VANET
through different approaches, including transmission power
control, packet generation frequency, and the adopted
modulation and coding scheme. J. Zhao et al. [16] and Y.
Wang et al. [17] proposed a multi-objective optimization
scheme with Particle Swarm Optimization to dynamically
adjust multi-transmission parameters for optimization of the
transmission capacity and the transmission delay. However,
these optimization schemes only adjust one or two
parameters without constraint of QoS requirements for
corresponding safety applications. Also, most of the
optimization algorithms cannot work in real-time.
In this paper, we propose a model-based optimization
scheme for VANET multi-parameter adjustments under the
constraints of QoS requirements for a given safety
application using a deep-learning neural network (DLNN).
In the scheme, a stochastic model is deployed to
characterize the IEEE 802.11 based channel access with the
help of real-time channel measurement and derive the
immediate quantity of QoS. The DLNN trained by the data
sampled from running the stochastic model helps find the
optimal solution with data. Compared with the other
optimization schemes for VANET safety applications and
DLNN based optimization approaches [18] [19], the main
contributions in this paper are: 1) The fusion of the DLNN
with generalization capability and the stochastic model can
reduce the searching hypothesis space to speed up the
optimization process. 2) The DLNN with discrete labels is
used for high-precision inverse mapping of IEEE 802.11
based VANETs. 3) New systematic optimization algorithms
are implemented and tested for real-time parameter tuning
to reach the optimal solution under constraints. Section II
gives a short overview of IEEE 802.11 driven VANETs, and
978-1-6654-6749-0/22/$31.00 ©2022 IEEE 835
analytic model that serves the optimization system. Section
III describes the optimization scheme and implementation
methods. Section IV shows the numerical results and
effectiveness of the proposed scheme along with
discussions. Conclusions are presented in Section V.
II. OVERVIEW OF VANETS FOR SAFETY SERVICES
A. Description of 802.11based VANET for BSM
Services
IEEE 802.11 driven VANETs are deployed to provide the
safety services through one-hop or multi-hop broadcasting
to disseminate real-time traffic information or safety-
related messages. The PHY layer of the communication
system is based on Orthogonal Frequency Division
Multiplexing (OFDM) working in the licensed frequency
band of 5.9 GHz (5.85–5.925 GHz) with bandwidth from
5MHz to 160MHz. In the PHY layer, Low-Density Parity
Check (LDPC) error-correction coding is introduced for
channel coding, which provides 2~3 dB of sensitivity gain
and offers increased spectral efficiency compared to the
Binary Convolutional Code (BCC). With help of channel
tracking using midamble symbols, it also supports higher-
rate modulation and coding schemes (MCS) up to 256-
QAM (MCS index k=8) and 1024-QAM (MCS index k=10)
with 52 data subcarriers. New multiple-input and multiple-
output (MIMO) and Dual Carrier Modulation (DCM) is
expected to bring ~3dB diversity gain (thus leading to
safety range extension). The MAC layer of the system uses
an enhanced distributed channel access (EDCA) method
with carrier sense multiple access with collision avoidance
(CSMA/CA). To enhance the reliability of safety
messaging, IEEE 802.11bd proposes an adaptive
retransmission scheme where decisions to retransmission
and the number of retransmissions (1~3) depend on the
measured congestion level. Specifically, in the adaptive
retransmission scheme, the number of retransmissions Nrp
depends on the dynamic occupancy of the channel, which
is designed by the following equation [9]:
(1)
The scheme is expected to offer a 4~7 dB performance
boost.
Equipped with the advanced high data-rate
communication technologies, the IEEE 802.11 driven
VANETs can potentially sustain data exchange between
high-speed vehicles and between the vehicles and the
roadside infrastructure with high QoS.
The safety messages can be classified in two categories:
Basic Safety Messages (BSMs) and event-driven safety
messages (ESMs) in US. ESMs are transmitted occasionally
in case of dangerous situations. BSMs are broadcasted
periodically to keep drivers informed about status of nearby
vehicles. Obviously, such safety-critical services are time-
sensitive and require high reliability. Potential safety
services and the corresponding QoS requirements are listed
in [11]. The major factors that degrade the reliability and the
performance of broadcast in VANETs are interferences
from other nodes’ transmissions, high vehicle mobility, and
adverse multi-path fading/shadowing channels.
B. Stochastic Model for DLNN Training and
Optimization
The analytical model being involved in the
optimization system is from a SINR based model [11] with
minor modifications.
We consider a highway wireless IEEE 802.11driven
broadcast VANET where each vehicle sends Basic Safety
Message (BSM) messages to its surrounding vehicles in its
transmission range regularly, and each vehicle receives the
broadcast messages from the surrounding vehicles. Given
message transmission rate λ, vehicle density β, and other
fixed communication parameters such as backoff window
size W0, MCS index k, Node transmission power Pt, carrier
sensing range rE, etc. the analytical model starts from the
analysis of IEEE 802.11 channel access for the evaluation
of transmission probabilities.
First, a SMPA (semi-Markov process with absorbing
state) model is deployed to characterize the IEEE 802.11
broadcast channel access behavior. Solving the model [11],
the following probabilities can be derived: the steady-state
probability πXMT that a vehicle is in the transmitting state, the
probability π0 that a vehicle starts to transmit in the
beginning of a time slot immediately after the backoff
process. Then, the probability that a hidden terminal
transmits during the vulnerable period of the transmission
from a tagged node.
, (2)
where T is the time duration for one packet transmission,
and AIFS is time duration for arbitration inter-frame spacing
of IEEE 802.11 MAC.
Fading/shadowing effect of wireless channels for
vehicular communication is described by the power Pr(d)
received from a receiver with distance d away from a source
node, and probability density function (PDF) of the power
Pr: , which can be measured from the practical
wireless channel, or abstracted by certain theoretical models
(e.g., Nakagami fading or Log-normal shadowing with
path-loss). Given a node distribution, distance ds between a
transmitter and a receiver, and mobility of vehicles, the
distribution of signal-to-interference-to-noise ratio (SINR)
on each receiver in the VANET can be derived as
(3)
where Psh(ds) is the hidden terminal probability that at least
one transmission from the hidden interference areas occurs;
Phc(ds) is the probability that at least two of nodes from the
hidden terminal areas transmit simultaneously; Pcc(ds) is
the probability that at least one concurrent collision occurs;
is the probability distribution of signal-to-
noise ratio (SNR) on the receiver.
Consequently, the following QoS metrics can be evaluated.
Packet (Message) Reception Probability (PRP) PRP is
defined as the probability that a receiver successfully
decodes a packet from a source node with distance ds to the
receiver,
The PRP evaluation based on the given PLR-SINR
curves can be derived from the above SINR distribution:
(4)
836
where Rth is the receiving power threshold, PLR(s) reflects
a Packet loss rate (PLR) function given a SINR value s, PLR
is the probability that a packet is received with error, which
is a function of bit error rate (BER)
, (5)
and
. (6)
Transmission Latency: the transmission latency is defined
as the time duration needed for a message (packet)
transmission on the wireless medium, which is evaluated as
, (7)
where tpre is the preamble duration, tAIFS is arbitrary inter-
frame space which is the waiting time for nodes after the
medium is sensed free, tsym is the OFDM symbol duration,
and nsym denotes the number of OFDM symbols required to
transmit a certain payload PL (including MAC header,
service, and tails bits) [11].
Packet Transmission Delay (ED) ED is defined as the
average delay a packet experiences from the time at which
the packet is generated, and the time at which the packet is
successfully received by all neighbors of the node that
generates the packet. The transmission delay ED in worst
case includes the average queuing delay E[Dq] and the
average medium service time E[S] (due to backoff, Packet
Transmission Latency , and propagation delay tp.).
. (8)
Channel Busy Rate (CBR) CBR is defined as the percentage
of busy time duration within certain observation period [11].
(9)
where Pdc is the probability that at least one more
transmission occurs in the carrier sensing range of an
observer in the channel and Pdh is the probability that at least
two hidden terminal transmissions are overlapped from
perspective of an observer in the channel.
APP Level Reliability PA(Nrp,ROI): PA(Nrp,ROI) is the
probability of successfully receiving at least one message
out of Nrp transmissions from a broadcast node within region
of interest (ROI) for a given safety application.
(10)
III. STRUCTURE AND IMPLEMENTATION OF THE
OPTIMIZATION SYSTEM
A. Optimization Problem Formulation
Consider an IEEE 802.11 driven VANET where each
node transmits messages regularly to its one-hop
neighbor(s). Each node in the transmitter’s ROI receives the
messages according to immediately measured SINRs. Since
communication environment including channel condition,
network topology and density, and node mobility is
dynamically changed, a fixed communication configuration
setting would lead to inefficient usage of communication
resources or incorrect conclusions on the network
capability. A stochastic model-based and DLNN driven
optimization platform is proposed to adjust the network
Deep Learning
fL (CBR)
Data Preprocessing
{Sp,CBR,QoS>QoSreq}
Stochastic Model
Trained DLNN
fL (CBR)
Stochastic Model
QoS*
Optimal Sp= Sp*
(a) Training (b) Optimization
-1
-1
-1
PHY Channel
f (p|d)
Pr|D
Training Data
CBR(Sp)
Sp
QoS
DLNN Configuration
FSINR
Constraint Data F us ion
-1
CBRmin*
Sp*
X*=CBR*
X*<CBRmin*?
CBRmin*=X*-ε
No
Yes
Figure 1 Structure of model-based DLNN optimization System
parameters for the robust and efficient deployment of the ad
hoc communication resources. Denote Sp as a set of cross-
layer communication adjustable parameters Sp={k, λ, W0,
Nrp} where k is the MCS index with respective channel data
bit rate, λ is the message generation rate, W0 is the 802.11
backoff window size, and Nrp is the number of message
repetitions. These parameter values are dropped in a discrete
set Scp due to nature of integer values:
). Then, setting up
U[Sp]=1/CBR indicates that the purpose of the optimization
is to minimize the channel usage under the condition that
the APP reliability and the transmission delay meet the QoS
requirements for the given safety application. Therefore, the
constraint utility-based optimization can be formulated as
subject to ,(11)
B. Structure and Implementation of Optimization
System
Figure 1 shows a structure of the proposed optimization
system by which the above optimization problem is
converted to DLNN-driven optimizations. In this structure,
the Stochastic Model described in Section II plays an
important role in training DLNNs. The PHY Channel
module provides data observed on wireless channels to
acquire channel characteristics regarding the effect of fading
and shadowing with power path-loss in the form of the
probability distribution of the received signal strength Pr.
Given the value ranges and possible values of Sp, the
Stochastic Model module runs the analytical model and
generates map data {QoS, CBR}= f (Sp, rE, β), which is used
to train the DLNNs to realize the inverse map between the
communication parameters and channel utility CBR. On the
other hand, the Stochastic Model module is cascaded with
the DLNN in the process of optimization to help improving
learning convergence (accuracy). The Data Preprocessing
module has the following three functions. 1) It carries out
data preprocessing from the Stochastic Model module,
including training data standardization, regroup, and
splitting of data. 2) It integrates the QoS requirements
constraints into the data set for training of the inverse map
DLNN. 3) The module is then designed to solve the model
family and the convergence theory of learning, and to
configure structure and the learning parameters of the
837
DLNN. Given an optimization problem, the optimization
solution search is carried out by running the trained DLNN
map inverse module along with constraint formalization and
a mechanism of approximation inference, as shown in
Figure 1(b). Initially, the channel model provides the IEEE
802.11 physical layer knowledge and parameter ranges and
channel fading/shadowing path-loss characteristics in the
form of the received power strength probability density
function (PDF) and receiving bit error rate (BER) as a
function of communication distance and SINR. Then, the
learned PDF is used for running the stochastic model with
all communication parameters to evaluate the distribution of
SINR. The DLNN module is deployed and configured to
learn an inverse function Sp=f-1 (CBR |QoS>QoSreq). Having
completed the training process, the trained DLNN and the
stochastic model are working together to find optimal
communication parameters Sp such that CBR reaches to its
minimum under the condition that the QoS requirements for
a selected safety service are met. In the optimization system,
the generalization capability of the DLNNs is leveraged to
find the optimal point that may not be observed as the
training samples. While the stochastic model helps to
narrow the search down to the valid communication
parameters in the search process. This supplementary fusion
allows small scale training data with low sampling rate and
accepts low convergency precision of DLNN training so
that the optimization can be performed with high accuracy
in real time and dynamically updated on-line.
Implementation of the optimization system can be
divided into two algorithms: Algorithm 1 for data
generation and DLNN training and Algorithm 2 for
constraint optimization. As shown from pseudo code for
Algorithm 1, the analytic model receives the
communication/network parameters and channel
fading/shadowing characteristics in the form of
(which can be acquired from theoretical equation for typical
vehicular communication channel or summarized from real
channel measurements). Set the adjustable parameters in
their possible value ranges, then run the model to obtain a
group of mapping data set {Sp, QoS}. During the generation
of the data set, a data preprocessing is performed to build
Algorithm 1 Data Generation and DLNN Training
1: Initialization: variables range of k, λ, , Nrp
2: for parameters in ranges Sp and rE, β do
3: input Sp, rE, β into the QoS generator model; and
4: execute the model to obtain PA, ED, CBR
5: return the results as training data
6: end for
7: preprocess and standardize data{CBR| QoS>QoSreq, Sp}
8: build and train by following parameters:
9: x=Dense(200, 200, 200, activation=ReLU)(Input)
10: y=Dense(64, 64, activation= ReLU)(Input)
11: k=Dense(activation=softmax)(y)
12: λ=Dense(activation=softmax)(y)
13: W0=Dense(activation=softmax)(x)
14: Nrp=Dense(activation=softmax)(y)
15: compile model with optimizer=rmsprop, Loss=categorical
crossentropy, Metrics=accuracy
16: train model with data, epochs=100, batch size=2000; and
17: return DL model from training
18: searching the minimum CBR from training data as a
local minimum
Algorithm 2 Constraint Optimization
1: while not find the Optimal Solution do
2: list1=take a series of numbers in this range from 0 to
local minimum CBR with random interval ϵ
3: for CBR in list do
4: run DL model predict (k,λ,W0,Nrp)=model
(CBR)
5: obtain PA and ED by analytic model:
6: {PA, ED, CBR}=Function QoS generator
model (k,λ,W0,Nrp,β)
7: search global minimum CBR:
8: if (CBR<current CBR min) and
(QoS>QoSreq) do
9: put PA, ED and CBR in a list2
10: end for
11: if (CBR in list2<current CBR min) do
12: replace the current CBR min with this CBR in
list2
13: back to while loop
14: else do
15: find the true Optimal Solution
16: end while
new standardized training data set {CBR|QoS>QoSreq, Sp}.
Then, four DLNNs are configured in Google Tensorflow as
multi-level classifiers with parameter outputs coded as one-
hot vectors considering discrete integers of parameter
values (dimension of output vectors: 3 for k; 19 for λ; 1023
for W0; and 10 for Nrp. 3 hidden layers with 200 neurons in
each layer for λ, W0 and Nrp, respectively, 2 hidden layers
with 64 neurons in each layer for k). Having completed the
DLNN training in Google Colab, as depicted in Algorithm
2, the optimization methodology is applied to find the
optimal parameters to minimize CBR under the QoS
requirement constraints. The optimization starts with
finding the minimum CBR from the training set, which can
be done during the training process. However, this CBRmin
may not be the true optimal solution. Run the trained DLNN
model with the current minimum CBRmin-ϵ, where ϵ is a set
of random numbers between 0 and CBRmin. The outputs Sp
will be tested for validity through the stochastic forward
model QoS=f(f-1(Sp)). All derived QoSs that meet the
specified QoS requirements will then be put into the DLNN
model as inputs until the minimum CBR is founded.
The model-based optimization algorithms can be run on
an edge computing platform with data from the covered
vehicles. Then, the optimized parameters are reported to the
individual vehicles.
IV. NUMERICAL RESULTS AND DISCUSSIONS
We apply the proposed optimization scheme to a
highway VANET for a slow vehicle warning (SVW) safety
application. An SVW application can provide alerts to the
driver about potential hazards if a slow vehicle is detected.
IEEE 802.11 based communication system following
updated standard for PHY layer and MAC layer [20] with
the communication parameters shown in Table I is adopted
to drive the VANET. Notice that four parameters will be
optimized for best channel efficiency under the constraint
QoS requirements. The QoS requirements of the SVW
safety application are listed in Table II.
838
Figure 2 Packet Loss Rate (PLR) of IEEE 802.11 in AWGN channel,
Payload P=1600 bytes
TABLE I Communication parameter settings
Parameters
Values
Parameters
Values
Average sensing
range rE
500 m
Packet generation
rate λ
2~40 packets/s
Slot time
13 µs
No. of subcarriers
52
Preamble duration
4 µs
Bandwidth B
10~160 MHz
AIFS
64 µs
Packet length PL
1600 bytes
CW W0
2~1024
Node trans. power Pt
0.28183815
Symbol duration tSy
1~8 µs
MCS index k
6~10
Coding rate r
¾, 5/6
Packet Rep. no. Nrp
1~10
MAC header
64 bits
Node density β
0.1~0.3 v/m
Table II QoS Requirements for SVW safety applications
Safety Apps
SVW
ROI（dROI）
350m
Tolerance time
0.5s
APP probability requirement ()
99%
Figure 2 shows packet loss rates (PLR) of typical MCSs
(k=6: 64-QAM and k=8: 256-QAM) as functions of SINR
in the wireless AWGN channel with packet length
E[PL]=1600 bytes. A Log-normal shadowing channel
model is assumed for the vehicular communication
channel, which is described by the model [21]
(12)
where denotes the signal strength as calculated with
the usage of the model of free space path loss at the
reference distance d0; dc stands for the critical distance.
and are the path-loss exponents. Xδ describes the
random shadowing effects, which is Gaussian distribution
with mean 0 and variance δ12 and δ22, respectively. For
urban area traffic environment [21], d0=1 meter, dc=102,
δ1=3.9dB, δ2=5.2dB α1=2.56, α2 =6.34, and Noise floor
Nf=1.2589x10-17watts. The communication nodes are
distributed with constant densities on the highway with a
length of 5000m.
Given the communication parameters ranges and the QoS
requirements for the SVW safety application, we run the
analytical model to generate the data for training the DLNN
Figure 3 DLNN training Loss convergency performance β=0.2 vehicles/m
Figure 4 DLNN training Accuracy convergency performance β=0.2
vehicles/m
as an inverse mapping Sp={k, λ, W0, Nrp}=f-1
(CBR|QoS>QoSreq). Figure 3 shows the DLNN
convergence performance of normalized Loss values that
evaluate how the DLNN behaves over each iteration of
training time epochs for four parameters being optimized
as the vehicle density is 0.2 vehicles/meter. From Figure 3,
we can see that the parameters can quickly converge to its
low loss values within 100 epochs (2.5 seconds)
(k_loss=0.9, λ_loss=0.62, W0_loss=2.99, Nrp_loss=1.3).
Figure 4 shows the DLNN convergence performance of
accuracy that measures generalization performance of the
DLNN over each iteration of training time epochs for four
parameters being optimized as the vehicle density is 0.2
vehicles/meter. From Figure 4, we can see that the
parameters can quickly converge to its high accuracy
within 100 epochs (k_accuracy=0.47, λ_accuracy=0.65,
W0_accuracy=0.05, Nrp_accuracy=0.32).
Then, Algorithm 2 searches and finds the optimal
parameter set Sp={k, λ, W0, Nrp}={10, 2, 151, 2}with
minimum CBR=0.05789 and achieved QoS PA=0.9943674
and ED=0.158ms.
In order to verify the robust adaptivity of the model-
based DLNN optimization to various training data
sampling interval, we change the data sample interval of
message generation rate (λ) and backoff window size (W0)
in the system. We observe that high resolution of training
data leads to long training time and total optimization time,
839
Table III Impact of sampling data rate on the optimization
Performance
Sampling intervals
Training
Time
Search
Time
Total
Time
Optimum
CBR error
λ: (5:5:40)
W0: (50:50:1000)
42.924s
59.782s
102.70s
0%
λ: (5:20:40)
W0: (50:50:1000)
15.90 s
59.44s
75.34 s
0%
λ: (5:10:40)
W0:(50:200:1000)
7.09s
59.246s
66.336s
5%
λ: (0:2:40)
W0: (50:50:1000)
97.204s
30.179s
127.38s
0%
λ: (0:2:40)
W0: (0:10:1000)
506.62s
30.059s
536.68s
0%
Table IV Comparisons with other systems (β=0.2 v/m)
QoS Values
Systems
PA
ED
CBR
802.11bd
0.99101
0.00041
0.64749
Fixed Parameters
0.98046
0.00032
0.52261
Optimal
0.99437
0.00015
0.05789
which causes poor performance of adaptivity of the
adjusting parameters to the dynamic communication
environment. Due to fusion of the analytical model and the
DLNN with function generalization capability in the
proposed system, lowering the sample rate of training data
to some extent allows real-time optimization with high
accuracy (Please see second line in Table III: find the
optimal CBR with 0 error and total optimization time 75.34
seconds). However, excessively lowering the sampling rate
of training data could cause error of optimization. It is also
observed that high resolution of training data can reduce the
time of seeking the optimal parameters in the price of
taking much longer training time for same accuracy.
Table IV shows comparisons of the QoS values from the
proposed optimization scheme and that derived in a fixed
parameter setting (k = 6, λ = 10, W0 = 15, Nrp = 2) and IEEE
802.11bd with the adaptive repetition algorithms. From
Table IV, we can see that the optimization scheme
proposed in this paper outperforms the other two
communication systems in terms of both QoS and channel
efficiency.
V. CONCLUSIONS
This paper proposes and tests a model-based deep learning
optimization system for cross-layer parameter optimization
of IEEE 802.11 driven vehicular communication networks.
The proposed optimization system can perform the
constraint optimization and come up with the parameters
and utility values in real-time with high accuracy.
Compared with the other IEEE 802.11 VANETs, this
approach is more adaptable to the dynamic communication
environments with lower computation complexity. The
system can be potentially extended to the optimization of
VANET with practical channel model and QoS model in
which the QoSs are analyzed by integrating analytical
models and wireless channel measurements. The future
work could be extension of the algorithm to more general
cases for more general applications, and configuration of
DLNN in terms of scale of the training data.
REFERENCES
[1] V2X Functional and Performance Test Procedures – Selected
Assessment of Device to Device Communication Aspects, 5GAA P-
180092, 2018.
[2] J. B. Kenney, Dedicated short-range communications (DSRC)
standards in the united states, Proceedings of the IEEE, 99(7), July
2001.
[3] B. Erdem, IEEE 802.11bd – A seamless evolutionary access layer
for ITS-G5/DSRC, CAR 2 CAR Journal, Page 21-27, Oct. 2019.
[4] 3GPP, 3GPP TR 22.886: study on enhancement of 3GPP support for
5G V2X services (v16.2.0, Release 16), Dec. 2018.
[5] X. Ma, J. Zhang, X. Yin, K. S. Trivedi, Design and Analysis of a
Robust Broadcast Scheme for VANET Safety-Related Services,
IEEE Trans. on Vehicular Technology, 61(1):46-61, 2012.
[6] X. Yin, X. Ma, and K. S. Trivedi, An interacting stochastic models
approach for the performance evaluation of DSRC vehicular safety
communication. IEEE Trans. Computers, 62(5): 873-885, 2013.
[7] X. Ma, X. Yin, M. Wilson, and K. S. Trivedi, MAC and application-
level broadcast reliability in VANETs with channel fading,
International Conference on Computing, Networking and
Communications, Mobile Computing Symposium (ICNC’13), Jan.
28-31, 2013, San Diego, USA.
[8] W. Anwar, N. Franchi, and G. Fettweis, Physical Layer Evaluation
of V2X Communications Technologies: 5G NR -V2X, LTE-V2X,
IEEE 802.11bd, and IEEE 802.11p, IEEE 90th Vehicular
Technology Conference (VTC2019-Fall), Sept. 2019.
[9] G. Naik, B. Choudhury, and J. Park, IEEE 802.11bd & 5G NR V2X:
Evolution of Radio Access Technologies for V2X
Communications, IEEE Access, Vol. 7, pp70169-70184, May 2019.
[10] X. Ma, S. Ding, C. R.Busse, and I. S. Esley, Multi-layer QoS
Analysis of IEEE 802.11bd Based VANET for Safety Applications,
IEEE Consumer Communications & Networking Conference
(CCNC), 9-12 January 2021.
[11] X. Ma, and K. S. Trivedi, SINR-based analysis of IEEE 802.11p/bd
broadcast VANETs for safety services, IEEE Trans. Netw. Serv.
Manag. 18(3): 2672-2686, 2021.
[12] D. B. Rawat, D. C. Popescu, G. Yan, and S. Olariu, Enhancing
VANET performance by joint adaptation of transmission power and
contention window size, IEEE Transactions on Parallel and
Distributed Systems, 22(9): 1528-1535, Sept. 2011.
[13] Y. Yao, X. Zhou, and K. Zhang, Density-aware rate adaptation for
vehicle safety communications in the highway environment, IEEE
Communications Letters, 18(7), July 2014.
[14] C. Christopher et al., Optimizing wireless channel using adaptive
modulation to improve QoS In VANET, Malaysian Journal of
Computer Science, 26(1), 2013.
[15] A. Bazzi, Congestion control mechanisms in IEEE 802.11p and
Sidelink C-V2X, arXiv:2001.08495, 2020.
[16] J. Zhao et al., Adaptive optimization of QoS constraint transmission
capacity of VANET, Elsevier Veh. Commun., V.17, pp. 1–9, 424,
Jun. 2019.
[17] Y. Wang, et al., Analysis and adaptive optimization of vehicular
safety message communications at intersections, Elsevier Ad Hoc
Networks, V. 107, Oct. 2020.
[18] J. Snoek, et al., Scalable Bayesian optimization using deep neural
networks, Proceedings of the 32nd International Conference on
International Conference on Machine Learning, v. 37, pp 2171–
2180, July 2015.
[19] A. Kumar and S. Levine, Model inversion networks for model-based
optimization, 34th Conference on Neural Information Processing
Systems (NeurIPS 2020), Vancouver, Canada.
[20] IEEE 802.11, “IEEE standard - part 11: Wireless LAN medium
access control (MAC) and physical layer (PHY) specifications,”
IEEE Std 802.11-2016 (Revision of IEEE Std 802.11-2012), pp. 1–
3534, Dec 2016.
[21] Y. Yao, et al., On MAC access delay distribution for IEEE 802.11p
broadcast in vehicular networks, IEEE Access, Volume 7, 2019.
840
