1. Introduction

• Development of an Advanced Multi-Metric Scoring Mechanism: we introduce a novel scoring algorithm that uses the properties of Massive MIMO and the 802.11ax standard. This algorithm evaluates routes based on a comprehensive set of QoS metrics, providing a nuanced and context-aware route selection process optimized for urban environments.
• Introduction of Machine Learning Techniques: by incorporating sophisticated machine learning algorithms, MMS-DSR predicts future route qualities based on historical data, enabling proactive route optimization. This is particularly useful in smart cities where predicting and adapting to changing conditions can significantly improve network performance.
• Integration of a CNN-LST Based Beamforming Algorithm: we enhance MMS-DSR with a beamforming optimization algorithm that dynamically adjusts beamforming vectors. This integration helps in managing the spatial-temporal variations in the channel, supporting the complex beamforming needs of Massive MIMO systems in 6G.
• Empirical Validation in a Smart City Simulation Environment: we validate the performance of MMS-DSR through extensive simulations using the INET framework along with the OSG Earth visualizer in OMNeT++. This simulation environment provides a realistic backdrop that highlights the protocol’s potential benefits for smart city applications, demonstrating its robustness and efficiency in a variety of challenging urban 6G scenarios.

2. Related Works

3. Preliminaries

4. Architectural Overview

4.5. Mathematical Formalization

MMS-DSR is designed with a unique multi-metric scoring function that adapts in real-time to the dynamic nature of Mobile Ad-Hoc Networks, incorporating beamforming strategies to enhance signal quality and reliability. This scoring function serves as the backbone for making robust and efficient routing decisions, especially in the complex and rapidly changing environments of smart cities. The scoring function $S(r,t)$ in MMS-DSR is defined as:

$$S(r,t)=\alpha ·\sum _{i=1}^n{w}_i\left(t\right)·m_i\left(r\right)+\delta ·P_{\mathrm{prediction}}(r,t)+\lambda ·BF(r,t)$$

(2)

where $m_i\left(r\right)$ are the route metrics, $w_i\left(t\right)$ the time-dependent weights, $\alpha$ a weighting factor for the summation real-time term, $\delta$ a factor integrating the predictive reliability score $P_{prediction}(r,t)$, and $\lambda$ a weighting factor for the beamforming score $BF(r,t)$.

The weights $w_i\left(t\right)$ are dynamically adjusted based on real-time network analytics and predictions from the Machine Learning-Based Prediction Unit to reflect the changing urban conditions. This adjustment is given by:

$$w_i(t+1)=w_i\left(t\right)-\eta {\displaystyle \frac{\partial J(w,t)}{\partial w_i\left(t\right)}}$$

(3)

where $\eta$ is the learning rate and ${\displaystyle \frac{\partial J(w,t)}{\partial w_i\left(t\right)}}$ is the gradient of the cost function with respect to weight $w_i\left(t\right)$.

The cost function $J(w,t)$ is designed to minimize the difference between the current metric values and their predicted values, ensuring that the weights reflect the most relevant network conditions. The cost function is defined as:

$$J(w,t)=\sum _{i=1}^n{\left({\displaystyle \frac{m_i(r,t)-{\widehat{m}}_i(r,t)}{{\widehat{m}}_i(r,t)}}\right)}^2$$

(4)

where $m_i(r,t)$ is the network metric for route r at time t, and ${\widehat{m}}_i(r,t)$ is the predicted value of the metric provided by the Machine Learning-Based Prediction Unit.

In order to understand how the parameters $P_{prediction}(r,t)$ and $BF(r,t)$ are obtained we must talk about Hybrid CNN-LSTM models. Hybrid Deep Learning models combining CNN and LSTM can improve the prediction accuracy [46,47,48,49]. The spatial and temporal features can be thoroughly extracted using hybrid models, where CNN is utilized to capture the spatial features of traffic data, while LSTM is employed to extract the temporal features. Let’s consider that we have traffic state data of K locations $s_i$ ($i=1,2,\dots ,K$) as inputs to predict the traffic states at times $t,t+1,\dots ,t+h$. The real-time measured data can be arranged, as explained in [49,50,51], into a matrix:

$$S=\left[\begin{array}{c}{s}_1\\ s_2\\ ⋮\\ s_K\end{array}\right]=\left[\begin{array}{cccc}{s}_{1,t-N}& s_{1,t-N+1}& \cdots & s_{1,t-1}\\ s_{2,t-N}& s_{2,t-N+1}& \cdots & s_{2,t-1}\\ ⋮& ⋮& \ddots & ⋮\\ s_{K,t-N}& s_{K,t-N+1}& \cdots & s_{K,t-1}\end{array}\right]$$

(5)

In $P_{prediction}(r,t)$, the element $s_{k,t}$ in the matrix is a vector which includes latency L, bandwidth B, reliability R, and beamforming efficacy $BE$. In $BF(r,t)$, the element $s_{k,t}$ in the matrix is a vector that contains the channel state information (CSI). The real-time measured data matrix S is first parallelized in the time domain and then input into the CNN, which captures high-level spatial features. Finally, the high-level spatial feature map is fed into LSTM models to generate the final prediction.

The high-level spatial feature map output by the CNN can be expressed as:

$$\mathbf{X}=\left[\begin{array}{c}{x}_1\\ x_2\\ ⋮\\ x_L\end{array}\right]=\left[\begin{array}{cccc}{x}_{1,t-N}& x_{1,t-N+1}& \cdots & x_{1,t-1}\\ x_{2,t-N}& x_{2,t-N+1}& \cdots & x_{2,t-1}\\ ⋮& ⋮& \ddots & ⋮\\ x_{L,t-N}& x_{L,t-N+1}& \cdots & x_{L,t-1}\end{array}\right]$$

(6)

where: $\mathbf{X}$ is the high-level spatial feature map, $x_{l,t}$ is the l-th high-level feature at the t-th time instant, L is the number of high-level features, and N is the length of the time window considered.

Each feature $x_{l,t}$ is obtained using the following equation:

$$x_{l,t}=f_l({\mathbf{w}}_t\otimes {\mathbf{S}}_t+b_t)$$

(7)

where: ${\mathbf{w}}_t$ is the filter with $K-L+1$ coefficients, ${\mathbf{S}}_t$ is the t-th column of $\mathbf{S}$, $b_t$ is the bias term, $f_l$ is the l-th activation function, ⊗ denotes the convolution operation.

To extract temporal features, the high-level spatial feature vector is selected as the input of LSTM, denoted as:

$$\mathbf{F}=[{\mathbf{F}}_0{\mathbf{F}}_1\cdots {\mathbf{F}}_{N-1}]$$

(8)

where ${\mathbf{F}}_n$ is the high-level spatial feature map for the n-th LSTM network, represented as:

$${\mathbf{F}}_n=\left[\begin{array}{cccc}{x}_{1,t+n-N}& x_{1,t+n-N+1}& \cdots & x_{1,t+n-N+M-1}\\ x_{2,t+n-N}& x_{2,t+n-N+1}& \cdots & x_{2,t+n-N+M-1}\\ ⋮& ⋮& \ddots & ⋮\\ x_{L,t+n-N}& x_{L,t+n-N+1}& \cdots & x_{L,t+n-N+M-1}\end{array}\right]$$

(9)

where M is the adjustable input window size.

The spatio-temporal features output by LSTM are denoted as:

$$\mathbf{H}=[{\mathbf{H}}_0{\mathbf{H}}_1\cdots {\mathbf{H}}_{N-1}]$$

(10)

where ${\mathbf{H}}_n$ is a $K\times T$ matrix, and T is the adjustable output window size with $M+T\le N$.

The generated spatio-temporal features are iteratively determined by:

$$\begin{array}{cc}\hfill f_n& =\sigma \left(W_{f,F}·\mathrm{vec}\left(F_n\right)+W_{f,H}·\mathrm{vec}\left(H_{n-1}\right)+b_f\right)\hfill \\ \hfill i_n& =\sigma \left(W_{i,F}·\mathrm{vec}\left(F_n\right)+W_{i,H}·\mathrm{vec}\left(H_{n-1}\right)+b_i\right)\hfill \\ \hfill c_n& =f_n\odot c_{n-1}+i_n\odot tanh\left(W_{c,F}·\mathrm{vec}\left(F_n\right)+W_{c,H}·\mathrm{vec}\left(H_{n-1}\right)+b_c\right)\hfill \\ \hfill o_n& =\sigma \left(W_{o,F}·\mathrm{vec}\left(F_n\right)+W_{o,H}·\mathrm{vec}\left(H_{n-1}\right)+b_o\right)\hfill \\ \hfill H_n& =o_n\odot tanh\left(c_n\right)\hfill \end{array}$$

(11)

where: $W_{*,F},W_{*,H}$ are weighting matrices for current input high-level spatial feature matrix $F_n$ and previous spatio-temporal feature matrix $H_{n-1}$. $\mathrm{vec}(·)$ denotes vectorization due to different sizes of $F_n$ and $H_{n-1}$. $\sigma$ and tanh are sigmoid and hyperbolic tangent functions applied element-wise.

MMS-DSR uses the infinite-horizon discounted Markov Decision Process (MDP) framework for route optimization. The MDP, as detailed in [52,53], is defined by a tuple $(S,A,P,R)$, where:

• S is the set of states representing different network conditions, including the varying urban scenarios.
• A is the set of actions corresponding to route selections.
• P is the state transition probability matrix, which is particularly complex in urban environments due to the myriad of potential interactions and obstructions.
• R is the reward function, aligned with the scoring function $S(r,t)$, and is enhanced to factor in urban dynamics and machine learning predictions.

The optimization objective in MDP, as well described in is formulated as:

$${\pi }^{*}=arg\underset{\pi }{max}\mathbb{E}\left[\sum _{t=0}^{\infty }{\gamma }^{t}R(s_t,\pi \left(s_t\right))\right]$$

(12)

where ${\pi }^{*}$ is the optimal policy we want to find, $\pi$ is a policy that maps states to actions, $\mathbb{E}$ is the mathematical expectation, representing the average value considering all possible sequences of states and actions, ${\sum }_{t=0}^{\infty }{\gamma }^{t}R(s_t,a_t)$ is the discounted sum of rewards over time, t is the time index, $\gamma$ is the discount factor, $R(s_t,a_t)$ is the reward obtained by taking action $a_t$ in state $s_t$. The discount factor $\gamma$ (0 < $\gamma$ < 1) reduces the weight of future rewards. Values close to 1 give more importance to future rewards, while values close to 0 emphasize immediate rewards.

In brief, the prediction model can be represented as:

$$Y_t=f_{\mathrm{CNN}-\mathrm{LSTM}}\left(X_t\right)$$

(13)

where: $X_t$ is the input feature vector at time t, $Y_t$ is the predicted future state at time t, $f_{\mathrm{CNN}-\mathrm{LSTM}}$ represents the CNN-LSTM model.

This prediction is then used to adjust the beamforming vectors $BF(r,t)$ and to update the multi-metric scoring engine $P_{prediction}(r,t)$. The CNN component extracts spatial features from the channel state information (CSI), while the LSTM component captures temporal dependencies, enabling the model to predict optimal beamforming strategies.

The beamforming optimization can be expressed as:

$$BF(r,t)=f_{\mathrm{opt}}(CS{I}_t,Y_t)$$

(14)

where: $CS{I}_t$ is the current channel state information at time t, $BF(r,t)$ is the beamforming vector at time t for route r, $f_{\mathrm{opt}}$ is the optimization function.

Here, $Y_t$ is the predicted future state from the CNN-LSTM model, and $f_{\mathrm{opt}}$ calculates the optimal beamforming vectors to enhance communication quality and reliability.

The optimization function typically aims to maximize signal strength and minimize interference. One way to model this is by maximizing the signal-to-interference-plus-noise ratio (SINR) for each beamforming vector. Here is a general formula for $f_{\mathrm{opt}}$:

$$BF(r,t)=arg\underset{BF}{max}\left({\displaystyle \frac{H_t·BF(r,t)}{{\sigma }^2+{\sum }_{i\ne t}{I}_i\left(BF(r,t)\right)}}\right)$$

(15)

where: $H_t$ represents the channel gain matrix for the current time t, $BF(r,t)$ is the beamforming vector at time t, ${\sigma }^2$ is the noise power, $I_i\left(BF(r,t)\right)$ represents the interference from other users or sources when using the beamforming vector $BF(r,t)$. The optimization seeks to maximize the SINR by adjusting the beamforming vectors $BF(r,t)$.

The Beamforming Optimization Unit dynamically adjusts beamforming vectors to maximize signal strength and minimize interference, based on both current and predicted channel state information (CSI). This optimization ensures that each node in the network can maintain high-quality communication links, which is essential in complex urban environments.

We conclude this part by explaining the reasons why we introduced the $\alpha$, $\delta$, and $\lambda$ weights. To ensure that as experience increases, the predictive terms ($P_{\mathrm{prediction}}(r,t)$ and $BF(r,t)$) become more significant while the real-time term (${\sum }_{i=1}^n{w}_i\left(t\right)·m_i\left(r\right)$) becomes less significant, we define a regulation formula. This formula adjusts the ratio between $\alpha$, $\delta$, and $\lambda$ over time or experience E:

$${\displaystyle \frac{\alpha }{\delta +\lambda }}=f\left(E\right)$$

(16)

where $f\left(E\right)$ is a decreasing function of experience. A possible choice for $f\left(E\right)$ could be an exponential decay function:

$$f\left(E\right)=e^{-\beta E}$$

(17)

where $\beta$ is a constant that controls the rate of decay. Thus, the relationship becomes:

$${\displaystyle \frac{\alpha }{\delta +\lambda }}=e^{-\beta E}$$

(18)

As experience E increases, $e^{-\beta E}$ decreases, indicating that $\alpha$ should decrease relative to $\delta$ and $\lambda$. This ensures the predictive terms gain more importance over time.

To summarize, the final formulation of the scoring function $S(r,t)$ with the regulation formula is:

$$S(r,t)=\alpha \left(E\right)·\sum _{i=1}^n{w}_i\left(t\right)·m_i\left(r\right)+\delta ·P_{\mathrm{prediction}}(r,t)+\lambda ·BF(r,t)$$

(19)

where

$$\alpha \left(E\right)=(\delta +\lambda )·e^{-\beta E}$$

(20)

This adjustment ensures that the real-time metrics ${\sum }_{i=1}^n{w}_i\left(t\right)·m_i\left(r\right)$ become less influential as the system gains more experience, while the predictive metrics $P_{\mathrm{prediction}}(r,t)$ and $BF(r,t)$ become more influential.

4.5.1. Proof of Effectiveness through Simulation and Analysis

The effectiveness of MMS-DSR is substantiated through:

• Simulation studies that demonstrate rapid convergence, adaptability, and efficiency, particularly in urban settings where the network topology and conditions can change unpredictably.
• Comparative analysis with existing protocols, showcasing superior performance in varied network conditions, with a focus on urban environments where traditional protocols often fail to maintain high performance.
• Statistical analysis of network performance metrics such as latency, throughput, reliability, and beamforming efficacy over multiple simulation runs, illustrating significant improvements in smart city applications.

4.5.2. Numerical Illustration

To provide a concrete understanding of the MMS-DSR scoring mechanism, we present a numerical example with a network topology, as illustrated in Figure 3. This hypothetical scenario considers several candidate routes, each with different metrics for latency, bandwidth, reliability, and beamforming efficacy (BE), typical of diverse urban paths.

The routes and their respective metrics are summarized in Table 1.

Table 2. Example routes with associated metrics.

Table 2. Example routes with associated metrics.

Route	Latency (ms)	Bandwidth (Mbps)	Reliability	Beamforming Efficacy (BE)
$r_1$	10	50	0.9	15
$r_2$	8	40	0.95	18
$r_3$	12	60	0.85	12
$r_4$	15	45	0.92	17
$r_5$	9	55	0.88	14
$r_6$	10	50	0.9	15
$r_7$	8	40	0.95	18
$r_8$	12	60	0.85	12
$r_9$	15	45	0.92	17
$r_{10}$	9	55	0.88	14

Using the scoring function we assume the following parameters: $\delta =0.2$, $\lambda =0.3$, $\beta =0$, Experience $E=10$, dynamic weights at time t: $w_1\left(t\right)=0.35$, $w_2\left(t\right)=0.4$, $w_3\left(t\right)=0.2$, $w_4\left(t\right)=0.05$.

The Predictive reliability scores at time t: $P_{prediction}(r_1,t)=0.92$, $P_{prediction}(r_2,t)=0.87$, $P_{prediction}(r_3,t)=0.90$, $P_{prediction}(r_4,t)=0.91$, $P_{prediction}(r_5,t)=0.89$, $P_{prediction}(r_6,t)=0.92$, $P_{prediction}(r_7,t)=0.87$, $P_{prediction}(r_8,t)=0.90$, $P_{prediction}(r_9,t)=0.91$, $P_{prediction}(r_{10},t)=0.90$.

The Beamforming Factor (BF) scores at time t: $BF(r_1,t)=20$, $BF(r_2,t)=22.5$, $BF(r_3,t)=15$, $BF(r_4,t)=21$, $BF(r_5,t)=19$, $BF(r_6,t)=20$, $BF(r_7,t)=22.5$, $BF(r_8,t)=20$, $BF(r_9,t)=22.5$, $BF(r_{10},t)=15$.

To ensure fair comparison, the metrics are normalized to a [0, 1] scale using the formula:

$$x_{\mathrm{norm}}={\displaystyle \frac{x-x_{\min }}{x_{\max }-x_{\min }}}$$

(21)

Table 3. Normalized metrics for each route.

Table 3. Normalized metrics for each route.

Route	Latency Norm	Bandwidth Norm	Reliability Norm	BE Norm
r1	0.286	0.50	0.5	0.500
r2	0.000	0.00	1.0	1.000
r3	0.571	1.00	0.0	0.000
r4	1.000	0.25	0.7	0.833
r5	0.143	0.75	0.3	0.333
r6	0.286	0.50	0.5	0.500
r7	0.000	0.00	1.0	1.000
r8	0.571	1.00	0.0	0.000
r9	1.000	0.25	0.7	0.833
r10	0.143	0.75	0.3	0.333

The Predictive reliability scores at time t were also normalized obtaining the following values: $P_{prediction}Norm(r_1,t)=1.0$, $P_{prediction}Norm(r_2,t)=0.0$, $P_{prediction}Norm(r_3,t)=0.6$, $P_{prediction}Norm(r_4,t)=0.8$, $P_{prediction}Norm(r_5,t)=0.4$, $P_{prediction}Norm(r_6,t)=1.0$, $P_{prediction}Norm(r_7,t)=0.0$, $P_{prediction}Norm(r_8,t)=0.6$, $P_{prediction}Norm(r_9,t)=0.8$, $P_{prediction}Norm(r_{10},t)=0.6$.

The Beamforming Factor (BF) scores at time t were normalized obtaining the following values: $BFNorm(r_1,t)=0.667$, $BFNorm(r_2,t)=1.000$, $BFNorm(r_3,t)=0.000$, $BFNorm(r_4,t)=0.800$, $BFNorm(r_5,t)=0.533$, $BFNorm(r_6,t)=0.667$, $BFNorm(r_7,t)=1.000$, $BFNorm(r_8,t)=0.667$, $BFNorm(r_9,t)=1.000$, $BFNorm(r_{10},t)=0.000$.

The scoring function $S(r,t)$ for each route is calculated using the normalized values.

For example for $r1$ we have:

$$S(r1,t)=0.5\times \left(0.35·0.286+0.4·0.50+0.2·0.5+0.05·0.5\right)+0.2·1.0+0.3·0.667$$

$$S(r1,t)=0.5\times (0.1001+0.2+0.1+0.025)+0.2+0.2$$

$$S(r1,t)=0.21255+0.2+0.2$$

$$S(r1,t)=0.6125$$

Repeating the same calculations for each route we will obtain the following values:

Table 4. Scoring Function values for each route after normalization.

Table 4. Scoring Function values for each route after normalization.

Route	Scoring Function (SF)
r1	0.612500
r2	0.425000
r3	0.420000
r4	0.715833
r5	0.453333
r6	0.612500
r7	0.425000
r8	0.620000
r9	0.775833
r10	0.333333

Even if by applying the scoring function, the path $r9$ emerges as the optimal route with the highest score of $0.775833$, we must remember that in the classic DSR protocol, the primary metric for selecting the optimal path is typically the number of hops. Carrying out the route analysis of the graph illustrated in Figure 4, the possible paths from source to destination and the related Scoring Function (SF) values are shown in Table 5:

Table 5. Paths, Routes, and Total Scoring Functions (SF).

Table 5. Paths, Routes, and Total Scoring Functions (SF).

Path	Routes	Total SF
A → B → J	r1 + r6	$0.612500+0.612500=1.225$
A → C → J	r2 + r7	$0.425000+0.425000=0.850$
A → D → J	r3 + r8	$0.420000+0.620000=1.040$
A → B → E → J	r1 + r4 + r9	$0.612500+0.715833+0.775833=2.104166$
A → C → F → J	r2 + r5 + r10	$0.425000+0.453333+0.333333=1.211666$

Considering that MMS-DSR prioritizes the route with the highest cumulative scoring function (SF) and minimum number of hops, the chosen optimal path would be:

$A->B->J$ with a total $SF$ of $1.225$.

In conclusion this numerical illustration demonstrates the effectiveness of MMS-DSR’s scoring engine and shows how it can determine the optimal path in a network topology, considering various metrics and beamforming efficacy.

5. CNN-LSTM Model Architecture

5.2. Dataset Preparation and Validation

This subsection details the procedures involved in assembling and validating a dataset that encapsulates the dynamic and complex environment of urban networks. Such dataset construction serves to train the CNN-LSTM model to effectively manage and predict urban traffic and network conditions.

5.2.1. Data Collection

Comprehensive data collection is fundamental for training robust models. The dataset includes data from multiple sources within the urban infrastructure:

• Traffic Sensors: installed at various intersections and highways, these sensors collect real-time data on vehicle counts, speeds, and queue lengths, which will be used to compute various metrics. This data helps in identifying congestion patterns and traffic flow dynamics.
• Network Performance Metrics: mobile network operators provide data on network usage including latency, bandwidth and reliability. These metrics are collected at minute intervals and are essential for assessing the quality of communication and identifying potential network failures.

5.2.2. Data Preprocessing

Data preprocessing involves several steps designed to prepare the raw data for effective machine learning processing:

• Data Cleaning: initial cleaning processes remove outliers and correct errors in the data to improve the quality and reliability of the input.
• Normalization: each metric is normalized using the min-max scaling technique, ensuring that the data falls within a specified range (usually 0 to 1), which helps in speeding up the learning process and improves model convergence.
• Tensor Structuring: the normalized data is then structured into tensors of dimensions $90\times 3$, where each row corresponds to one minute of data. This tensor format is specifically designed to align with the CNN-LSTM model’s input requirements.

5.2.3. Data Structuring for Machine Service Enhancement

Structured data input is crucial for the effective operation of machine learning models. The following Python code snippet illustrates how to structure the data into a 90x3 tensor, suitable for feeding into our CNN-LSTM model. Each row in this tensor represents one minute of collected data across three critical network metrics: latency, bandwidth, and reliability. This structure supports the model’s requirement to analyze spatial-temporal dynamics for network management.

Listing 1: Structured Data Preparation for CNN-LSTM Model

This structured approach to data not only provides the CNN-LSTM model with the required format for optimal processing but also encapsulates the essence of the network’s dynamic conditions within each minute interval. Each metric is carefully normalized and aligned within a structured array to maintain consistency and improve the predictive accuracy of the model. The listing includes detailed normalization steps, where each network performance metric is adjusted to a specific range. This is essential to maintain uniformity in input scales and to prevent any feature from dominating the model’s training process due to scale discrepancies.

By stacking the normalized data into a 90x3 tensor with ‘np.stack’, we ensure that the data retains its chronological order across different metrics. This structure allows the CNN to extract spatial features effectively while the LSTM can track temporal patterns, both critical for making informed predictions in a network environment.

The choice of stacking along ‘axis=1’ ensures that each row of the tensor corresponds to a single time interval, with columns representing different metrics. This alignment is crucial for the CNN-LSTM model to process each time slice as a unified snapshot of network conditions, enhancing the detection of underlying patterns and dependencies.

5.4. Model Training and Preparation

The CNN-LSTM model is trained using historical traffic data, encompassing a variety of urban traffic scenarios. Training is conducted using MATLAB, known for its robust environment for machine learning model development and simulation.

Listing 2: Training the CNN-LSTM Model in MATLAB

In the Listing 2 the data variable could include metrics such as latency, bandwidth, and reliability which are crucial for understanding traffic flow and network performance. The convolutional layers are designed to extract spatial features from the sequential input data, which are crucial for identifying patterns. LSTM layers follow to capture temporal dependencies, important for forecasting traffic conditions. Training Options: parameters such as MaxEpochs, MiniBatchSize and InitialLearnRate are tuned to optimize the learning process, ensuring the model adequately learns from historical data to make accurate predictions about traffic patterns.

6. Adapting MMS-DSR from MANETs to VANETs for Smart Cities

6.2. Mathematical Formalization for VANET

To adapt the previous Mathematical Formalization from MANETs to VANETs, we need to consider the specific characteristics and requirements of VANETs. In particular, in mathematical formalization for VANETs we must incorporate additional parameters such as the Density of the Vehicles, the Distance of the Vehicle from the Destination, and the Trajectory of the Vehicle [54,55]. These parameters are crucial for ensuring efficient and reliable routing in highly dynamic vehicular environments.

Vehicle Density (De): vehicle density measures the concentration of vehicles within a specific area of the network. It is a critical metric for evaluating network congestion, as high vehicle density can lead to increased communication delays and packet collisions.

$$De(r,t)={\displaystyle \frac{\mathrm{TNVA}}{\mathrm{AS}}}$$

(22)

where: $TNVA$ is the total number of vehicles in the area, and $AS$ is the area size. High vehicle density indicates potential congestion, which can negatively impact communication reliability and latency. Understanding vehicle density helps in optimizing routing decisions to avoid congested areas and improve overall network performance.

Vehicle Distance (Di): vehicle distance is the Euclidean distance of the vehicle in the route from the destination.

$$Di(r,t)=exp\left(-{\left({\displaystyle \frac{d_{\mathrm{i}}}{d_{\mathrm{ref}}}}\right)}^{\zeta }\right)$$

(23)

where: $d_{\mathrm{i}}$ is the Euclidean distance of the vehicle from the destination, $d_{\mathrm{ref}}$ is a reference distance, $\zeta$ is an attenuation factor.

Vehicle distance provides insight into the potential longevity of communication links, allowing for better route stability. Routes with shorter distances to the destination are preferred, as they are likely to offer more reliable communication.

Vehicle Trajectory (T): vehicle trajectory measures the predicted path towards the destination for the vehicle in the route, considering its current and future positions. This metric helps in assessing the alignment of vehicle movements with the desired route.

$$T(r,t)=exp\left(-\left({\displaystyle \frac{\Delta d{\left(t\right)}^{\zeta }}{d_{\mathrm{ref}}}}\right)\right)$$

(24)

where: $\Delta d{\left(t\right)}^{\zeta }=d{\left(t\right)}^{\zeta }-d{\left(0\right)}^{\zeta }$, $d\left(t\right)$ is the future distance of the vehicle in the route to the destination at time t, $d\left(0\right)$ is the current distance to the destination, $d_{\mathrm{ref}}$ is a reference distance, $\zeta$ is an attenuation factor.

Vehicle trajectory helps in predicting the future alignment of vehicles with the intended route, ensuring that the selected path remains optimal over time. Routes with stable trajectories are preferred as they minimize deviations and potential disruptions in communication.

To predict the next position $(x_{i+1},y_{i+1})$ of the vehicle, we use the current velocity vector $(v_x,v_y)$ and the time interval $\Delta t$:

$$x_{i+1}=x_i+v_x·\Delta t$$

(25)

$$y_{i+1}=y_i+v_y·\Delta t$$

(26)

where: $(x_i,y_i)$ are the current coordinates of the vehicle, $v_x$ and $v_y$ are the velocity components in the x and y directions, respectively, and $\Delta t$ is the time interval.

Vehicle trajectory helps in predicting the future alignment of vehicles with the intended route, ensuring that the selected path remains optimal over time. Routes with stable trajectories are preferred as they minimize deviations and potential disruptions in communication.

To avoid making the article too heavy, we refer to the articles cited [54,55] for more details on the formulas relating to the parameters.

Below is the adapted mathematical formalization for VANETs:

VANETs Score Computation: the overall score for each route, S, is computed as:

$$\begin{array}{cc}\hfill S(r,t)& =w_1·{\displaystyle \frac{1}{L(r,t)}}+w_2·B(r,t)+w_3·R(r,t)+w_4·BE(r,t)\hfill \\ \hfill & +w_5·{\displaystyle \frac{1}{De(r,t)}}+w_6·{\displaystyle \frac{1}{Di(r,t)}}+w_7·{\displaystyle \frac{1}{T(r,t)}}\hfill \end{array}$$

(27)

where $w_1,w_2,w_3,w_4,w_5,w_6,$ and $w_7$ are the weights for latency, bandwidth, reliability, beamforming efficacy, density, distance, and trajectory respectively.

The scoring function $S(r,t)$ for VANETs is defined as:

$$\begin{array}{cc}\hfill S(r,t)& =\alpha ·\sum _{i=1}^n{w}_i\left(t\right)·v_i\left(r\right)+\delta ·P_{prediction}(r,t)+\lambda ·BF(r,t)\hfill \end{array}$$

(28)

where: $v_i\left(r\right)$ are the route metrics, $w_i\left(t\right)$ are the time-dependent weights, $\alpha$ is a weighting factor for the summation real-time term, $\delta$ integrates the predictive reliability score $P_{prediction}(r,t)$, $\lambda$ is a weighting factor for the beamforming score $BF(r,t)$.

When $De$ increases (i.e., more vehicles in a given area), it indicates higher congestion, leading to increased communication delays and potential packet collisions. Consequently, the score for routes through such areas would decrease, signaling less optimal paths. As $Di$ increases, the vehicle is farther from the destination, likely leading to less reliable communication links. The score for such routes decreases because longer distances can result in higher latency and potential disconnections. An increase in T (indicating a less predictable or more erratic trajectory) implies that the vehicle’s future positions are less aligned with the optimal path, leading to potential disruptions in communication. This decreases the route score.

Integrating these new metrics is critically important. Vehicle Distance ($Di$) and Vehicle Trajectory (T) metrics account for the relative position and movement of vehicles towards the destination, helping to predict the stability and longevity of communication links. Vehicle Density ($De$) assesses the congestion level around each vehicle, influencing the likelihood of packet collisions and network performance.

6.2.1. Numerical Illustration for VANET

To provide a concrete understanding of the MMS-DSR scoring mechanism for VANETs, we present a numerical example with a network topology, as illustrated in Figure 8. This hypothetical scenario considers several candidate routes, each with different metrics for latency, bandwidth, reliability, vehicle density, vehicle distance, vehicle trajectory, and beamforming efficacy (BE), typical of diverse urban paths.

The routes and their respective metrics are summarized in Table 5.

Table 6. Example routes with associated metrics.

Table 6. Example routes with associated metrics.

Route	L (ms)	B (Mbps)	R	De	Di (m)	T (m)	BE
$r_1$	10	50	0.9	20	100	10	15
$r_2$	8	40	0.95	30	120	8	18
$r_3$	12	60	0.85	25	150	12	12
$r_4$	15	45	0.92	15	140	15	17
$r_5$	9	55	0.88	22	130	9	14
$r_6$	10	50	0.9	20	100	10	15
$r_7$	8	40	0.95	30	120	8	18
$r_8$	12	60	0.85	25	150	12	12
$r_9$	15	45	0.92	15	140	15	17
$r_{10}$	9	55	0.88	22	130	9	14

Using the scoring function we assume the following parameters: $\delta =0.2$, $\lambda =0.3$, $\beta =0$, Experience $E=10$, dynamic weights at time t: $w_1\left(t\right)=0.3$, $w_2\left(t\right)=0.2$, $w_3\left(t\right)=0.1$, $w_4\left(t\right)=0.05$, $w_5\left(t\right)=0.1$, $w_6\left(t\right)=0.15$, $w_7\left(t\right)=0.1$.

The Predictive reliability scores at time t: $P_{prediction}(r_1,t)=0.92$, $P_{prediction}(r_2,t)=0.87$, $P_{prediction}(r_3,t)=0.90$, $P_{prediction}(r_4,t)=0.91$, $P_{prediction}(r_5,t)=0.89$, $P_{prediction}(r_6,t)=0.92$, $P_{prediction}(r_7,t)=0.87$, $P_{prediction}(r_8,t)=0.90$, $P_{prediction}(r_9,t)=0.91$, $P_{prediction}(r_{10},t)=0.90$.

The Beamforming Factor (BF) scores at time t: $BF(r_1,t)=20$, $BF(r_2,t)=22.5$, $BF(r_3,t)=15$, $BF(r_4,t)=21$, $BF(r_5,t)=19$, $BF(r_6,t)=20$, $BF(r_7,t)=22.5$, $BF(r_8,t)=20$, $BF(r_9,t)=22.5$, $BF(r_{10},t)=15$.

To ensure fair comparison, the metrics are normalized to a [0, 1] scale using the formula:

$$x_{\mathrm{norm}}={\displaystyle \frac{x-x_{\min }}{x_{\max }-x_{\min }}}$$

(29)

Table 7. Normalized metrics for each route.

Table 7. Normalized metrics for each route.

Route	L Norm	B Norm	R Norm	De Norm	Di Norm	T Norm	BE Norm
r1	0.286	0.50	0.5	0.20	0.0	0.20	0.500
r2	0.000	0.00	1.0	0.60	0.4	0.00	1.000
r3	0.571	1.00	0.0	0.40	1.0	0.40	0.000
r4	1.000	0.25	0.7	0.00	0.8	0.60	0.833
r5	0.143	0.75	0.3	0.35	0.6	0.25	0.333
r6	0.286	0.50	0.5	0.20	0.0	0.20	0.500
r7	0.000	0.00	1.0	0.60	0.4	0.00	1.000
r8	0.571	1.00	0.0	0.40	1.0	0.40	0.000
r9	1.000	0.25	0.7	0.00	0.8	0.60	0.833
r10	0.143	0.75	0.3	0.35	0.6	0.25	0.333

The Predictive reliability scores at time t were also normalized obtaining the following values: $P_{prediction}Norm(r_1,t)=1.0$, $P_{prediction}Norm(r_2,t)=0.0$, $P_{prediction}Norm(r_3,t)=0.6$, $P_{prediction}Norm(r_4,t)=0.8$, $P_{prediction}Norm(r_5,t)=0.4$, $P_{prediction}Norm(r_6,t)=1.0$, $P_{prediction}Norm(r_7,t)=0.0$, $P_{prediction}Norm(r_8,t)=0.6$, $P_{prediction}Norm(r_9,t)=0.8$, $P_{prediction}Norm(r_{10},t)=0.6$.

The Beamforming Factor (BF) scores at time t were normalized obtaining the following values: $BFNorm(r_1,t)=0.667$, $BFNorm(r_2,t)=1.000$, $BFNorm(r_3,t)=0.000$, $BFNorm(r_4,t)=0.800$, $BFNorm(r_5,t)=0.533$, $BFNorm(r_6,t)=0.667$, $BFNorm(r_7,t)=1.000$, $BFNorm(r_8,t)=0.667$, $BFNorm(r_9,t)=1.000$, $BFNorm(r_{10},t)=0.000$.

The scoring function $S(r,t)$ for each route is calculated using the normalized values.

For example for $r1$ we have:

$$S(r1,t)=0.5\times \left(0.3·0.286+0.2·0.50+0.1·0.5+0.05·0.2+0.1·0.2+0.15·0.0+0.1·0.2\right)$$

$$+0.2·1.0+0.3·0.667$$

$$S(r1,t)=0.5\times (0.0858+0.1+0.05+0.01+0.02+0.0+0.02)+0.2+0.3·0.667$$

$$S(r1,t)=0.1429+0.2+0.2001$$

$$S(r1,t)=0.5429$$

Repeating the same calculations for each route we will obtain the following values:

Table 8. Scoring Function values for each route after normalization.

Table 8. Scoring Function values for each route after normalization.

Route	Scoring Function (SF)
r1	0.5429
r2	0.5500
r3	0.4700
r4	0.675833
r5	0.530000
r6	0.5429
r7	0.5500
r8	0.4700
r9	0.675833
r10	0.530000

Even if by applying the scoring function, the paths $r4$ and $r9$ emerge as the optimal route with the highest score of $0.675833$, we must remember that in the classic DSR protocol, the primary metric for selecting the optimal path is typically the number of hops. Carrying out the route analysis of the graph illustrated in Figure 9, the possible paths from source to destination and the related Scoring Function (SF) values are shown in Table 9:

Table 9. Paths, Routes, and Total Scoring Functions (SF).

Table 9. Paths, Routes, and Total Scoring Functions (SF).

Path	Routes	Total SF
A → B → J	r1 + r6	$0.5429+0.5429=1.0858$
A → C → J	r2 + r7	$0.5500+0.5500=1.1000$
A → D → J	r3 + r8	$0.4700+0.4700=0.9400$
A → B → E → J	r1 + r4 + r9	$0.5429+0.675833+0.675833=1.894566$
A → C → F → J	r2 + r5 + r10	$0.5500+0.530000+0.530000=1.610000$

Considering that MMS-DSR prioritizes the route with the highest cumulative scoring function (SF) and minimum number of hops, the chosen optimal path would be:

$A\to C\to J$ with a total $SF$ of $1.1000$.

In conclusion, this numerical illustration demonstrates the effectiveness of MMS-DSR’s scoring engine and shows how it can determine the optimal path in a network topology, considering various metrics and beamforming efficacy.

7. Considered Scenario for MMS-DSR

8. Simulation and Evaluation

8.1. Simulation Setup

Simulations utilize OMNeT++ integrated with SUMO (Simulation of Urban MObility) and the inet framework to emulate realistic urban MANET and VANET scenarios. SUMO manages the vehicular traffic within the city, reflecting accurate vehicle movements and interactions, while inet supports advanced network simulations using the 802.11ax HE mode that we designed in [56]. This setup allows for the detailed assessment of the MMS-DSR protocol in managing communication in densely populated urban areas.

Figure 11. SUMO simulation scenario in Reggio Calabria.

Figure 11. SUMO simulation scenario in Reggio Calabria.

8.2. Simulation Parameters

We configure our simulation to reflect the unique traffic and networking conditions of the Metropolitan area of Reggio Calabria. Special attention is given to the placement of Road-Side Units (RSUs) equipped with MU-MIMO technology, optimized for beamforming capabilities to enhance communication efficiency across the network.

Table 10. Simulation Parameters tailored for Reggio Calabria’s urban environment.

Table 10. Simulation Parameters tailored for Reggio Calabria’s urban environment.

Parameter	Value
Tone Configuration	26, 242, 2x996
Protocols	DSR, MMS-DSR, SOL-DSR, Enhanced OLSR
MIMO Configurations	10x10
Seed Variability	Multiple sets
Confidence Interval	95%
Simulation Time	1000 seconds
Number of Nodes	20 to 100
Data Rate	Up to 9607.5 Mbps
Network Standard	IEEE 802.11ax
Mobility Model	Random Waypoint, Manhattan
Traffic Type	UDP, CBR, VBR
TX Data App	UDPBasicApp
RX Data App	UDPSink
Packet Size	1500 bytes
Node Sensitivity	-120 dBm
Simulation Tools	OMNeT++, SUMO, Veins
RSU Placement	Strategically placed at major intersections
RSU Quantity	15 RSUs covering key traffic nodes

The RSUs are strategically placed at major intersections and points of high vehicle density to maximize the MU-MIMO gains and beamforming effectiveness, ensuring optimal coverage and minimal communication disruption. This placement is based on traffic flow data, which indicates areas of frequent congestion and high vehicular mobility, making them critical points for effective traffic management and emergency response facilitation.

8.4. Implementation of Comparative Protocols

In the evaluation of MMS-DSR, we compare its performance not only with the traditional DSR protocol but also with two advanced adaptations: SOL-DSR and Enhanced OLSR. These comparisons are crucial to validate the effectiveness and efficiency of MMS-DSR in handling urban vehicular communications. Here we detail the implementation of these protocols within the OMNeT++ simulation framework, alongside the classic DSR protocol provided by the inet framework. Classic DSR, a well-established routing protocol in ad hoc networks, is utilized as a baseline for our simulations. This protocol is integrated into the inet framework, which provides a robust implementation suitable for both MANETs and VANETs. The inet framework’s version of DSR follows the standard routing mechanism where nodes dynamically discover and maintain source routes to all other nodes in the network. SOL-DSR has been Implemented in Python, the DRL model is interfaced with OMNet++ through the Veins framework, which allows for the integration of machine learning algorithms with traditional network simulation components. The DRL algorithm continuously updates its policy based on the feedback from the network, selecting routes that minimize delays and maximize throughput. The model is trained offline using a synthetic dataset generated from simulated network scenarios, and the trained model is then deployed in the simulation to make real-time routing decisions. The implementation of Enhanced OLSR in OMNeT++ is carried out using the inet framework, with modifications to the standard OLSR protocol to include "greedy forwarding" techniques. These techniques utilize dynamic assessment of neighbor proximity, based on the real-time vehicular data provided by SUMO, to make forwarding decisions. This ensures that the routing is highly responsive to the urban dynamics, providing improved performance in terms of latency and network overhead in densely populated city scenarios. All protocols are simulated under identical network conditions in OMNeT++ using the same vehicular mobility patterns generated by SUMO. This ensures that the comparison is fair and highlights the relative strengths and weaknesses of each protocol in managing urban vehicular communications. Metrics such as end-to-end delay, packet delivery ratio, and routing overhead are measured to evaluate and compare the performance of the protocols.

The diagram in Figure 12 illustrates the integration of the simulation tools and routing protocols used in our study. The inet framework provides foundational network functionalities and implements the classic DSR protocol, which serves as our baseline comparison. SUMO generates realistic vehicular movement data for simulating urban traffic patterns. Veins play a crucial role by bridging the gap between OMNet++ simulations and real-world applications by incorporating machine learning models, such as those used in SOL-DSR, into network simulations. This setup ensures that our simulations are not only accurate but also reflective of actual urban vehicular environments. Each protocol, including our enhanced MMS-DSR, interacts within the OMNeT++ environment to assess its effectiveness against urban mobility challenges, providing a comprehensive evaluation of routing efficiency and communication reliability.

Figure 12. Visualization of protocol implementation and interactions with simulation tools.

Figure 12. Visualization of protocol implementation and interactions with simulation tools.

8.5. Performance Metrics

We focus on throughput, end-to-end latency, and route discovery time to assess the MMS-DSR protocol’s performance under varying network loads and conditions. Throughput measures the network’s capacity to deliver data successfully, reflecting the efficiency of data handling by the protocol.

Figure 13. Throughput comparison across varying number of vehicles.

Figure 13. Throughput comparison across varying number of vehicles.

Throughput is a critical metric for assessing the performance of routing protocols in Vehicular Ad Hoc Networks. The MMS-DSR protocol showcases an exemplary integration of machine learning techniques, specifically Long Short-Term Memory networks, to optimize routing decisions and enhance network performance effectively. This integration is evident in the near-linear growth of throughput observed in our simulations, where MMS-DSR accelerates from 1200 Mbps with 10 vehicles to a remarkable 9400 Mbps with 100 vehicles. This growth not only underscores the effectiveness of LSTM-based predictive modeling in optimizing data paths but also highlights the protocol’s ability to dynamically minimize congestion, even as the network scales.

In comparison to other protocols like SOL-DSR and Enhanced OLSR, MMS-DSR demonstrates superior scalability and efficiency. For instance, at 100 vehicles, MMS-DSR achieves 9400 Mbps, significantly outperforming SOL-DSR’s 6600 Mbps and Enhanced OLSR’s 5000 Mbps. This substantial throughput advantage is attributable to the multi-metric scoring engine of MMS-DSR, which dynamically adjusts the weights of various routing metrics based on current network conditions, combined with the predictive power of LSTM that enhances route selection and management. This approach enables MMS-DSR to maintain high throughput levels, even as network conditions change, by proactively adapting to shifts in bandwidth availability, vehicle density, and network reliability.

Urban environments, characterized by high vehicle mobility and fluctuating traffic densities, pose unique challenges for routing protocols. MMS-DSR addresses these challenges head-on by utilizing its LSTM networks to predict and adapt to dynamic network conditions. For example, MMS-DSR’s throughput remains stable and high in urban scenarios, a testament to its ability to predict potential bottlenecks and disruptions in data flow. The protocol’s predictive capabilities allow it to reroute data preemptively around congested areas and potential points of disruption, ensuring smooth data transmission. For instance, even as the number of vehicles increases, MMS-DSR maintains a throughput increase from 1200 Mbps at 10 vehicles to 9400 Mbps at 100 vehicles, showcasing less than a 1% drop in efficiency compared to the initial performance, unlike traditional protocols which exhibit larger declines.

Furthermore, MMS-DSR’s efficiency in managing network resources is highlighted by its intelligent routing decisions that balance load across the network, preventing the over-utilization of any single path and supporting uniform data distribution. This is crucial in maintaining high throughput levels, as seen in our simulations where MMS-DSR consistently outperforms other protocols across various vehicle densities and conditions. For example, when the vehicle density increases from 50 to 100 vehicles, MMS-DSR manages to keep its throughput nearly optimal, showcasing only a minimal decrease in performance, which is significantly lower than the declines observed in protocols like SOL-DSR and Enhanced OLSR.

Also, the urban scenario analysis further demonstrates the robustness of MMS-DSR in handling the dynamic and often unpredictable environments of city scenario. The protocol’s machine learning models enable it to effectively navigate frequent topology changes and varying traffic densities, ensuring optimal route selection and minimal delays. This capability is particularly important during peak traffic hours in cities, where MMS-DSR can predict increased vehicle density and adjust routes in real-time to avoid slow-moving areas, thereby maintaining high throughput levels.

Figure 14. Throughput comparison across varying vehicle speeds.

Figure 14. Throughput comparison across varying vehicle speeds.

MMS-DSR also demonstrates superior performance in terms of throughput across varying vehicle speeds. At a speed of 100 km/h, MMS-DSR achieves 9000 Mbps, which is significantly higher than SOL-DSR’s 7300 Mbps and Enhanced OLSR’s 6300 Mbps. This can be attributed to MMS-DSR’s adaptive routing strategies and predictive capabilities that effectively manage the dynamic nature of vehicular movement.

As vehicle speed increases, MMS-DSR’s ability to predict and adapt to changes in the network ensures that data packets are routed efficiently, avoiding potential disruptions and maintaining high throughput. For instance, even at high speeds, MMS-DSR’s throughput only decreases slightly from 1100 Mbps at 10 km/h to 9000 Mbps at 100 km/h, demonstrating its robustness and efficiency in high mobility scenarios. This minimal decrease in performance at higher speeds contrasts sharply with the more significant drops observed in other protocols.

Furthermore, the efficiency of MMS-DSR in managing network resources and avoiding congestion is evident as vehicle speeds increase. The protocol’s intelligent routing decisions, informed by LSTM predictions, ensure that data is transmitted smoothly even in high-speed environments, reducing the likelihood of packet loss and maintaining consistent throughput.

8.5.1. Average End-to-End Latency

Latency measures the time it takes for packets to travel from the source to the destination, indicating the responsiveness of the network.

Figure 15. Latency comparison in function of vehicle density.

Figure 15. Latency comparison in function of vehicle density.

Figure 16. Latency comparison in function on vehicle speed.

Figure 16. Latency comparison in function on vehicle speed.

Latency is an essential metric for evaluating the responsiveness and efficiency of routing protocols in Vehicular Ad Hoc Networks. The MMS-DSR protocol demonstrates a significant advantage in managing latency, particularly evident in our simulation results, where latency increases from 20 ms at a density of 10 vehicles/km² to only 60 ms at 100 vehicles/km². This increase is notably lower compared to other protocols, such as SOL-DSR, which jumps from 25 ms to 75 ms, and Enhanced OLSR, which escalates from 30 ms to 85 ms under similar conditions. This remarkable performance by MMS-DSR can be primarily attributed to the integration of Long Short-Term Memory networks, which empower the protocol to proactively adjust routes in response to potential congestion and vehicle mobility, thereby avoiding common causes of delay.

The machine learning impact on MMS-DSR is profound, as its use of LSTM networks allows for predictive adjustments that preemptively mitigate potential delays. For instance, as vehicle density increases, leading to higher chances of congestion and complex vehicle interactions, MMS-DSR effectively uses its LSTM models to foresee and navigate around these potential bottlenecks. This proactive routing ensures that even in scenarios where vehicle density significantly increases, the latency remains optimally low. For example, when vehicle density increases from 50 to 100 vehicles/km², MMS-DSR exhibits only a minimal increase in latency from 40 ms to 60 ms, demonstrating an exceptional ability to manage and adapt to increased traffic without a corresponding steep rise in delay.

However, MMS-DSR’s adaptive routing capability is further highlighted in its consistent performance across various network configurations. Its ability to maintain lower latency under increasing vehicle density showcases the effective use of LSTM predictions to anticipate and adjust to potential delays dynamically. This ensures faster packet delivery, which is particularly beneficial in urban scenarios where quick response times are crucial. For instance, in a dense urban environment characterized by frequent stop-and-go traffic and variable vehicle speeds, MMS-DSR’s latency remains around 60 ms at 100 vehicles/km², significantly lower than the 75 ms for SOL-DSR and 85 ms for Enhanced OLSR. This is crucial for applications requiring timely data delivery, such as dynamic traffic light control and emergency vehicle routing, where delays can have significant implications.

The consistent low latency of MMS-DSR underscores its suitability for urban environments, where the dynamic movement of vehicles necessitates quick adaptability and responsive routing. The protocol’s machine learning-based predictive adjustments ensure timely data delivery, essential for safety and traffic management applications. For example, in scenarios involving emergency responses where vehicles need to communicate quickly with each other and with roadside infrastructure, MMS-DSR’s ability to keep latency low, despite increasing vehicle density and mobility, ensures that critical information is relayed without delay, supporting efficient emergency handling and enhancing overall urban mobility.

The detailed analysis of latency in MMS-DSR, with its integration of LSTM networks and adaptive routing strategies, reveals a protocol that is not only responsive and efficient but also highly adaptable to the varying dynamics of urban network environments. The protocol’s performance in maintaining low latency across different vehicle densities and speeds, and urban scenarios, combined with its robust machine learning capabilities, positions MMS-DSR as an advanced solution for future developments in VANET routing protocols, particularly for applications demanding quick and reliable communication in complex urban settings.

8.5.2. Route Discovery Time

The time required to discover a route reflects the protocol’s efficiency in establishing connectivity, especially in dynamic networks where rapid changes are common.

Figure 17. Route discovery time VS vehicle density.

Figure 17. Route discovery time VS vehicle density.

Figure 18. Route discovery time VS vehicle speed.

Figure 18. Route discovery time VS vehicle speed.

Route discovery time is a crucial performance metric for routing protocols in Vehicular Ad Hoc Networks, particularly in high-mobility urban environments where rapid topological changes are common. MMS-DSR demonstrates superior performance in this aspect, with route discovery times increasing from only 15 ms at a density of 10 vehicles/km² to 55 ms at 100 vehicles/km². This progressive increase is significantly slower compared to other protocols like SOL-DSR and Enhanced OLSR, where times escalate more steeply from 20 ms to 70 ms and from 25 ms to 80 ms respectively under similar conditions. This highlights the effectiveness of the integrated machine learning model in MMS-DSR, which uses Long Short-Term Memory networks to predict and adjust routes dynamically, facilitating quick and efficient pathfinding even as network complexity increases.

The impact of machine learning on MMS-DSR is significant, particularly in how it optimizes route discovery times. The LSTM networks enable the protocol to preemptively adjust routes in response to real-time changes in the network topology. For instance, as vehicle density increases and the network becomes more congested, MMS-DSR efficiently predicts potential bottlenecks and reroutes data packets through less congested paths. This ability is reflected in the moderate increase in discovery times from 30 ms at 50 vehicles/km² to only 55 ms at 100 vehicles/km², demonstrating an exceptional capacity to handle increased traffic and complexity without a corresponding steep rise in discovery times.

Furthermore, MMS-DSR’s efficient pathfinding is evident across various network densities. Its lower discovery times, compared to other protocols, underscore its advanced algorithms and the features provided by LSTM predictions. These algorithms enable MMS-DSR to quickly adapt to topological changes and optimize routes without incurring excessive overhead. For example, in urban scenarios characterized by high mobility and frequent topological changes, MMS-DSR maintains discovery times around 55 ms for 100 vehicles/km², much lower than the 70 ms for SOL-DSR and 80 ms for Enhanced OLSR. This rapid route discovery is crucial for maintaining continuous and reliable communication in dynamic urban settings, supporting operational efficiency and safety in smart city applications.

In dense urban networks, where the interaction between vehicles is continuous and complex, MMS-DSR’s rapid route discovery capability ensures that communication delays are minimized. For instance, in scenarios involving real-time traffic information sharing or coordination among emergency vehicles, MMS-DSR’s ability to maintain low discovery times ensures that information is shared promptly, enhancing the responsiveness of urban transport systems and emergency services.

In summary, the detailed analysis of route discovery time in MMS-DSR, with its integration of LSTM networks and efficient pathfinding algorithms, reveals a protocol that is not only quick and responsive but also highly adaptable to the varying dynamics of urban environments. The protocol’s performance in maintaining low discovery times across different vehicle densities and speeds, urban scenarios, and under dynamic topological changes, combined with its robust machine learning capabilities, positions MMS-DSR as an advanced solution for future developments in VANET routing protocols, especially for applications demanding fast and reliable route establishment in complex and high-mobility urban settings.

To further validate the effectiveness of MMS-DSR, additional simulation results focusing on key performance metrics such as Packet Delivery Ratio (PDR), Routing Overhead, and Scalability can be discussed. These metrics provide a comprehensive view of the protocol’s performance in urban VANET environments, highlighting its robustness and efficiency.

8.5.3. Routing Overhead

Routing Overhead measures the extra communication required for maintaining the routing information, indicating the efficiency of the protocol.

MMS-DSR exhibits lower routing overhead compared to SOL-DSR, Enhanced OLSR, and traditional DSR, as depicted in Figure 19 and Figure 20. The overhead increases from 200 control packets at 10 vehicles/km² to 1600 control packets at 100 vehicles/km² and from 200 control packets at 10 km/h to 1550 control packets at 100 km/h. This is significantly lower than the 2000 control packets required by Enhanced OLSR and DSR. MMS-DSR’s on-demand routing strategy, combined with the predictive capabilities of LSTM networks, minimizes unnecessary control packet transmissions, thus reducing overhead and improving overall network efficiency.

Figure 19. Routing overhead comparison in function of vehicle density.

Figure 19. Routing overhead comparison in function of vehicle density.

Figure 20. Routing overhead comparison in function of vehicle speed.

Figure 20. Routing overhead comparison in function of vehicle speed.

A key advantage of MMS-DSR is its ability to predict vehicle movements without relying on GPS systems. By utilizing predictive models that anticipate vehicle trajectories, MMS-DSR reduces the need for frequent route updates, which significantly lowers the routing overhead. This approach ensures efficient use of network resources and enhances the protocol’s scalability and adaptability in urban VANET environments.

The machine learning model in MMS-DSR, particularly the LSTM networks, plays a crucial role in this optimization. These networks forecast vehicle movements based on historical and real-time data, allowing the protocol to adjust routes dynamically and avoid congested areas. As vehicle density and speed increase, MMS-DSR maintains lower overhead by efficiently managing control packet transmissions and preventing unnecessary route discoveries. For example, at a vehicle density of 50 vehicles/km², MMS-DSR requires only 1000 control packets compared to 1500 for Enhanced OLSR, demonstrating its superior efficiency.

Furthermore, in high-density urban scenarios, MMS-DSR’s ability to predict and adapt to vehicle movements without GPS significantly reduces overhead. This is crucial for applications where network efficiency and resource management need to be handled. The protocol’s predictive capabilities enable it to maintain low overhead even as the number of vehicles and their speeds increase, ensuring reliable and efficient communication in dynamic and complex urban environments.

Basically, the detailed analysis of routing overhead in MMS-DSR, with its integration of predictive models and efficient control packet management, reveals a protocol that is not only efficient and scalable but also highly adaptable to the varying dynamics of urban VANET environments. The protocol’s performance in maintaining low overhead across different vehicle densities and speeds, combined with its robust machine learning capabilities, positions MMS-DSR as an advanced solution for future developments in VANET routing protocols, particularly for applications demanding efficient and reliable communication in complex urban settings.

8.5.4. Scalability

Scalability examines the protocol’s ability to handle increasing network sizes without significant performance degradation.

MMS-DSR demonstrates excellent scalability performance, as illustrated in Figure 21 and Figure 22. The Normalized Performance Index decreases from 0.95 with 10 vehicles/km² to 0.63 with 100 vehicles/km² and from 0.95 at 10 km/h to 0.63 at 100 km/h. This decline is less pronounced compared to SOL-DSR, Enhanced OLSR, and DSR. MMS-DSR’s machine learning-enhanced predictive capabilities and efficient routing strategies allow it to manage increasing network sizes effectively, ensuring consistent performance across various vehicle densities and speeds. This robustness is crucial for maintaining reliable communication in densely populated urban VANET environments.

Figure 21. Scalability comparison performance across increasing vehicle density.

Figure 21. Scalability comparison performance across increasing vehicle density.

Figure 22. Scalability comparison performance across increasing vehicle speed.

Figure 22. Scalability comparison performance across increasing vehicle speed.

In high-density scenarios, MMS-DSR’s predictive modeling using Long Short-Term Memory networks allows it to anticipate and adapt to changes in vehicle movements and network conditions dynamically. This capability ensures that even as vehicle density and speed increase, MMS-DSR can maintain high performance with minimal degradation. For example, at a density of 100 vehicles/km² and speed of 100 km/h, MMS-DSR achieves a performance index of 0.63, compared to 0.48 for traditional DSR. This significant difference highlights MMS-DSR’s ability to manage higher densities and speeds without compromising on efficiency.

The MMS-DSR protocol benefits from its multi-metric scoring engine, which dynamically adjusts the weights of various routing metrics based on current network conditions. This adaptability is particularly beneficial in urban environments where vehicle densities and speeds can fluctuate rapidly. By expoliting predictive models, MMS-DSR can preemptively adjust routes to avoid congestion, ensuring that data transmission remains smooth and efficient.

In addition, the protocol’s efficiency is further enhanced by its reduced reliance on frequent route updates. Unlike traditional protocols that may rely on constant GPS data, MMS-DSR predicts vehicle movements and adjusts routes accordingly. This approach reduces the need for continuous control packet transmissions, lowering overhead and preserving network resources.

In essence, the detailed analysis of scalability in MMS-DSR, with its integration of machine learning models and dynamic routing strategies, reveals a protocol that is not only scalable and efficient but also highly adaptable to the varying dynamics of urban VANET environments. The protocol’s performance in maintaining high scalability across different vehicle densities, speeds, and numbers, combined with its robust predictive capabilities, positions MMS-DSR as an advanced solution for future developments in VANET routing protocols, particularly for applications demanding efficient and reliable communication in complex urban settings.

9. Conclusions and Future Goals

9.1. Conclusions

The Multi-metric Scoring Dynamic Source Routing protocol has been demonstrated to be a highly efficient and adaptable solution for routing in Vehicular Ad-hoc Networks within smart city environments. Through extensive simulations and empirical validation, MMS-DSR has shown superior performance in key metrics such as throughput, latency, and route discovery time compared to traditional DSR, SOL-DSR, and Enhanced OLSR protocols.

MMS-DSR’s integration of machine learning, particularly Long Short-Term Memory networks, for predictive routing, allows it to dynamically adapt to rapidly changing urban traffic conditions. This proactive approach ensures reliable and efficient communication by optimizing data paths and adjusting to network dynamics in real-time. The use of advanced beamforming techniques further enhances communication quality by improving signal directionality and reducing interference, which is critical in dense urban environments.

The simulation results underscore MMS-DSR’s robust performance across various urban scenarios, demonstrating its ability to handle high mobility and frequent topological changes inherent in VANETs. Thanks to the capabilities of Massive MIMO and IEEE 802.11ax, MMS-DSR provides a scalable and future-proof routing protocol suitable for the evolving demands of 6G networks and smart cities.

Table 11 provides a detailed comparison of the main aspects of the MMS-DSR, SOL-DSR, Enhanced OLSR, and DSR protocols. The MMS-DSR approach stands out for its advanced technological features, incorporating machine learning, particularly LSTM networks, to optimize routing decisions. This protocol is versatile, suitable for both MANET and VANET environments, and employs beamforming with MU-MIMO technology, making it well-suited for the most recent network standards such as IEEE 802.11ax and future 6G networks. Notably, MMS-DSR does not rely on GPS, which significantly reduces overhead, enhancing its scalability and efficiency in urban scenarios with high vehicle density and dynamic network conditions.

Table 11. Summary table of main features of each considered approach.

Table 11. Summary table of main features of each considered approach.

Approach	Technical Features	Application Fields	Additional Features
MMS-DSR	Uses machine learning Suitable for both MANET and VANET Employs beamforming with MU-MIMO technology Adapts to real-time network conditions Supports most recent network standards	Good for urban environments with high vehicle density Suitable for high-mobility scenarios Effective in dynamic and complex environments Optimized for 6G networks	Does not rely on GPS, reducing overhead Maintains high throughput and low latency Predictive modeling for efficient routing Scalable to large network sizes
SOL-DSR	Self-organizing and adaptive Learning algorithms for routing decisions On-demand route discovery	Suitable for moderate mobility networks Effective in both urban and suburban areas	Higher overhead due to frequent route discoveries Limited scalability in high-density scenarios Less effective predictive capabilities
Enhanced OLSR	Link-state routing protocol Greedy forwarding techniques Suitable for both MANET and VANET	Suitable for static and low-mobility networks Effective in well-defined urban layouts	High control packet overhead Struggles with rapid topology changes Less scalable in high-mobility environments
DSR	Source-routing protocol Simple and well-established On-demand route discovery	Suitable for low-mobility and low-density networks Effective in small-scale ad hoc networks	High latency and low throughput in dense networks Significant overhead in dynamic environments Poor scalability and limited predictive capabilities

In contrast, the SOL-DSR protocol, while adaptive and incorporating learning algorithms for improved routing, tends to generate higher overhead due to frequent route discoveries. This makes it less scalable in high-density scenarios compared to MMS-DSR. Enhanced OLSR, being a link-state routing protocol with greedy forwarding techniques, is effective in static or low-mobility networks but struggles with rapid topology changes and incurs high control packet overhead, limiting its scalability in high-mobility environments.

The traditional DSR protocol, known for its simplicity and on-demand route discovery, performs well in low-mobility and low-density networks but faces challenges in dense and dynamic environments. It experiences high latency, low throughput, and significant overhead in such scenarios, making it less suitable for large-scale or high-mobility applications compared to MMS-DSR. Overall, MMS-DSR’s integration of predictive modeling, adaptive routing, and reduced reliance on continuous control packet transmissions positions it as a highly advanced and efficient solution for modern urban VANET and MANET environments, addressing the limitations observed in the other protocols.

References

1. Al-shareeda, M.A.; Alazzawi, M.A.; Anbar, M.; Manickam, S.; Al-Ani, A.K. A comprehensive survey on vehicular ad hoc networks (vanets). 2021 International Conference on Advanced Computer Applications (ACA). IEEE, 2021, pp. 156–160.
2. Yogarayan, S. Wireless ad hoc network of manet, vanet, fanet and sanet: a review. Journal of Telecommunication, Electronic and Computer Engineering (JTEC) 2021, 13, 13–18. [Google Scholar]
3. Reddy, A.S. Performance of VANET over MANET in Mobile Computing Environment. 2022 7th International Conference on Communication and Electronics Systems (ICCES). IEEE, 2022, pp. 659–664.
4. Quy, V.K.; Nam, V.H.; Linh, D.M.; Ban, N.T.; Han, N.D. Communication solutions for vehicle ad-hoc network in smart cities environment: A comprehensive survey. Wireless Personal Communications 2022, 122, 2791–2815. [Google Scholar] [CrossRef]
5. Amir, E.A.; Altabbakh, S.M.; Beshay, S.T.; Gomaa, I. Performance Evaluation of VANET Statistical Broadcast Protocols for Smart Cities. 2023 20th ACS/IEEE International Conference on Computer Systems and Applications (AICCSA). IEEE, 2023, pp. 1–8.
6. Sumit.; Chhillar, R.S.; Dalal, S.; Dalal, S.; Lilhore, U.K.; Samiya, S. A dynamic and optimized routing approach for VANET communication in smart cities to secure intelligent transportation system via a chaotic multi-verse optimization algorithm. Cluster Computing 2024, pp. 1–26.
7. Abir, M.A.B.S.; Chowdhury, M.Z.; Jang, Y.M. Software-defined uav networks for 6g systems: Requirements, opportunities, emerging techniques, challenges, and research directions. IEEE Open Journal of the Communications Society 2023. [Google Scholar]
8. Vaigandla, K.K.; Bolla, S.; Karne, R. A survey on future generation wireless communications-6G: requirements, technologies, challenges and applications. International Journal 2021, 10. [Google Scholar]
9. Gururaj, H.; Natarajan, R.; Almujally, N.A.; Flammini, F.; Krishna, S.; Gupta, S.K. Collaborative energy-efficient routing protocol for sustainable communication in 5G/6G wireless sensor networks. IEEE Open Journal of the Communications Society 2023. [Google Scholar] [CrossRef]
10. Raddo, T.R.; Rommel, S.; Cimoli, B.; Vagionas, C.; Perez-Galacho, D.; Pikasis, E.; Grivas, E.; Ntontin, K.; Katsikis, M.; Kritharidis, D.; others. Transition technologies towards 6G networks. EURASIP Journal on Wireless Communications and Networking 2021, 2021, 100. [Google Scholar] [CrossRef]
11. Salameh, A.I.; El Tarhuni, M. From 5G to 6G—challenges, technologies, and applications. Future Internet 2022, 14, 117. [Google Scholar] [CrossRef]
12. Dorothy, A.B.; Sreeja, B.; Chamundeeswari, V.; Madala, R.; Singh, D.P.; Sucharitha, K. 6G Networks in Unmanned Ariel Vehicle Intelligent Internet of Things Secured Communication. 2023 6th International Conference on Contemporary Computing and Informatics (IC3I). IEEE, 2023, Vol. 6, pp. 1374–1378.
13. Al Ridhawi, I.; Aloqaily, M.; Karray, F. Solutions for Moving Internet of Things Devices. IEEE ROBOTICS & AUTOMATION MAGAZINE 2022, 1070. [Google Scholar]
14. Gupta, R.; Nair, A.; Tanwar, S.; Kumar, N. Blockchain-assisted secure UAV communication in 6G environment: Architecture, opportunities, and challenges. IET communications 2021, 15, 1352–1367. [Google Scholar] [CrossRef]
15. Albattah, W.; Habib, S.; Alsharekh, M.F.; Islam, M.; Albahli, S.; Dewi, D.A. An overview of the current challenges, trends, and protocols in the field of vehicular communication. Electronics 2022, 11, 3581. [Google Scholar] [CrossRef]
16. Almazok, S.A.; Bilgehan, B. A novel dynamic source routing (DSR) protocol based on minimum execution time scheduling and moth flame optimization (MET-MFO). EURASIP Journal on Wireless Communications and Networking 2020, 2020, 1–26. [Google Scholar] [CrossRef]
17. Quy, V.K.; Ban, N.T.; Han, N.D. A multi-metric routing protocol to improve the achievable performance of mobile ad hoc networks. In Modern Approaches for Intelligent Information and Database Systems; Springer, 2018; pp. 445–453.
18. Zarzoor, A.R. Enhancing dynamic source routing (DSR) protocol performance based on link quality metrics. 2021 international seminar on application for technology of information and communication (iSemantic). IEEE, 2021, pp. 17–21.
19. Natarajan, B.; Shanmugam, A. A trust based model to mitigate black hole attacks in DSR based manet. European Journal of Scientific Research 2011, 50, 6–15. [Google Scholar]
20. Liang, Q.; Lin, T.; Wu, F.; Zhang, F.; Xiong, W. A dynamic source routing protocol based on path reliability and link monitoring repair. Plos one 2021, 16, e0251548. [Google Scholar] [CrossRef] [PubMed]
21. Gao, S.; Li, H.; Zhang, Q. An optimized routing protocol for MANETs based on multi-metric dynamic weighting. 2022 IEEE 5th International Conference on Electronics Technology (ICET). IEEE, 2022, pp. 954–958.
22. Shams, E.A.; Rizaner, A.; Ulusoy, A.H. Trust aware support vector machine intrusion detection and prevention system in vehicular ad hoc networks. Computers & Security 2018, 78, 245–254. [Google Scholar]
23. Laanaoui, M.; Raghay, S. Enhancing OLSR protocol by an advanced greedy forwarding mechanism for VANET in smart cities. Smart Cities 2022, 5, 650–667. [Google Scholar] [CrossRef]
24. Bajracharya, R.; Shrestha, R.; Kim, S.; Jung, H. 6G NR-U based wireless infrastructure UAV: Standardization, opportunities, challenges and future scopes. IEEE Access 2022, 10, 30536–30555. [Google Scholar] [CrossRef]
25. Sharma, S.; Mahajan, A.N.; Poonia, R.C. SA-DSR: a Bandwidth Optimizing Technic for Dynamic Source Routing Protocol. International Journal of Engineering and Advanced Technology (IJEAT) 2020, 9. [Google Scholar] [CrossRef]
26. Nurwarsito, H.; Aziz, A. Implementation of the Friis free space propagation model in the Dynamic Source Routing (DSR) routing protocol in the Vehicular Ad-hoc Network (VANET) with variations of road models. Journal of Physics: Conference Series. IOP Publishing, 2021, Vol. 1962, p. 012063.
27. Kim, K.; Lee, J.; Nam, J. A study on efficient Direction and Speed in Dynamic Source Routing (DS-DSR) protocol for highly mobile VANET. J. Digit. Contents Soc 2021, 22, 1521–1528. [Google Scholar] [CrossRef]
28. Marten, R.; Kartikasari, D.P.; Bakhtiar, F.A. Analisis Pengaruh Node Density pada Vehicular Ad Hoc Network (VANET) dengan menggunakan Protokol Routing Dynamic Source Routing (DSR) di Model Jalan Perkotaan. Jurnal Pengembangan Teknologi Informasi dan Ilmu Komputer 2022, 6, 5123–5128. [Google Scholar]
29. Maahin, H.; Kondamuri, D.; Polisetty, S.; Yaddanapudi, S.D. VANET-Based Communication in Vehicles to Control Accidents Using an Efficient Routing Strategy. International Conference on Paradigms of Communication, Computing and Data Analytics. Springer, 2023, pp. 829–838.
30. Li, Y.; Chen, Y.; Yan, D.; Guan, K.; Han, C. Channel characterization and ray-tracing assisted stochastic modeling for urban vehicle-to-infrastructure terahertz communications. IEEE Transactions on Vehicular Technology 2022, 72, 2748–2763. [Google Scholar] [CrossRef]
31. Mustari, N.; Karabulut, M.A.; Shah, A.S.; Tureli, U. Terahertz Communication with MIMO-OFDM in FANETs for 6G. The Open Transportation Journal 2023, 17. [Google Scholar] [CrossRef]
32. Saoud, B.; Shayea, I.; Yahya, A.E.; Shamsan, Z.A.; Alhammadi, A.; Alawad, M.A.; Alkhrijah, Y. Artificial Intelligence, Internet of things and 6G methodologies in the context of Vehicular Ad-hoc Networks (VANETs): Survey. ICT Express 2024. [Google Scholar]
33. Wang, Z.; Xu, Y.; Liu, J.; Li, Z.; Li, Z.; Jia, H.; Wang, D. An Efficient Data Sharing Scheme for Privacy Protection Based on Blockchain and Edge Intelligence in 6G-VANET. Wireless Communications and Mobile Computing 2022, 2022, 5031112. [Google Scholar] [CrossRef]
34. Zhang, X.; Pan, S.; Miao, Q. Adaptive beamforming-based gigabit message dissemination for highway VANETs. IEEE Transactions on Intelligent Transportation Systems 2021, 23, 7666–7679. [Google Scholar] [CrossRef]
35. Reddy, K.S.K.; Tamizhazhagan, V.; Murugan, V.S. VANET’s: Performance analysis of fixed beamforming techniques for vehicular Environments. 2021 International Conference on Advance Computing and Innovative Technologies in Engineering (ICACITE). IEEE, 2021, pp. 474–481.
36. Doddalinganavar, S.S.; Tergundi, P.; Patil, R.S. Survey on deep reinforcement learning protocol in VANET. 2019 1st International Conference on Advances in Information Technology (ICAIT). IEEE, 2019, pp. 81–86.
37. Khan, Z.; Fan, P.; Fang, S.; Abbas, F. An unsupervised cluster-based VANET-oriented evolving graph (CVoEG) model and associated reliable routing scheme. IEEE Transactions on Intelligent Transportation Systems 2019, 20, 3844–3859. [Google Scholar] [CrossRef]
38. Liu, X.; Amour, B.S.; Jaekel, A. A reinforcement learning-based congestion control approach for V2V communication in VANET. Applied Sciences 2023, 13, 3640. [Google Scholar] [CrossRef]
39. Murshed, R.U.; Ashraf, Z.B.; Hridhon, A.H.; Munasinghe, K.; Jamalipour, A.; Hossain, M.F. A cnn-lstm-based fusion separation deep neural network for 6g ultra-massive mimo hybrid beamforming. IEEE Access 2023. [Google Scholar] [CrossRef]
40. Nguyen, C.; Hoang, T.M.; Cheema, A.A. Channel estimation using CNN-LSTM in RIS-NOMA assisted 6G network. IEEE Transactions on Machine Learning in Communications and Networking 2023. [Google Scholar] [CrossRef]
41. Tian, Z.; Yu, D.; Bai, Y.; Lei, S.; Wang, Y. Application of Spectrum State Prediction Method based on CNN-LSTM Network in Communication Interference. IEEE Access 2023. [Google Scholar] [CrossRef]
42. Wang, P.; Li, Z.; Wei, Z.; Wu, T.; Luo, C.; Jiang, W.; Hong, T.; Pedersen, G.F.; Shen, M. Space-Time-Coding Digital Metasurface Element Design Based on State Recognition and Mapping Methods with CNN-LSTM-DNN. IEEE Transactions on Antennas and Propagation 2024. [Google Scholar] [CrossRef]
43. Daldoul, Y.; Meddour, D.E.; Ksentini, A. Performance evaluation of OFDMA and MU-MIMO in 802.11 ax networks. Computer Networks 2020, 182, 107477. [Google Scholar] [CrossRef]
44. Triwinarko, A.; Dayoub, I.; Cherkaoui, S. PHY layer enhancements for next generation V2X communication. Vehicular Communications 2021, 32, 100385. [Google Scholar] [CrossRef]
45. Aslam, M.; Jiao, X.; Liu, W.; Mehari, M.; Havinga, T.; Moerman, I. A novel hardware efficient design for IEEE 802.11 ax compliant OFDMA transceiver. Computer Communications 2024, 219, 173–181. [Google Scholar] [CrossRef]
46. Duan, Z.; Yang, Y.; Zhang, K.; Ni, Y.; Bajgain, S. Improved deep hybrid networks for urban traffic flow prediction using trajectory data. Ieee Access 2018, 6, 31820–31827. [Google Scholar] [CrossRef]
47. Liu, Y.; Zheng, H.; Feng, X.; Chen, Z. Short-term traffic flow prediction with Conv-LSTM. 2017 9th international conference on wireless communications and signal processing (WCSP). IEEE, 2017, pp. 1–6.
48. Sainath, T.N.; Vinyals, O.; Senior, A.; Sak, H. Convolutional, long short-term memory, fully connected deep neural networks. 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). Ieee, 2015, pp. 4580–4584.
49. Wu, Y.; Tan, H. Short-term traffic flow forecasting with spatial-temporal correlation in a hybrid deep learning framework. arXiv preprint 2016, arXiv:1612.01022. [Google Scholar]
50. Wang, S. Traffic State Prediction and Traffic Control Strategy for Intelligent Transportation Systems. In Intelligent Electronics and Circuits-Terahertz, ITS, and Beyond; IntechOpen, 2022.
51. Zheng, H.; Lin, F.; Feng, X.; Chen, Y. A hybrid deep learning model with attention-based conv-LSTM networks for short-term traffic flow prediction. IEEE Transactions on Intelligent Transportation Systems 2020, 22, 6910–6920. [Google Scholar] [CrossRef]
52. Mammeri, Z. Reinforcement learning based routing in networks: Review and classification of approaches. Ieee Access 2019, 7, 55916–55950. [Google Scholar] [CrossRef]
53. Shani, L.; Efroni, Y.; Mannor, S. Exploration conscious reinforcement learning revisited. International conference on machine learning. PMLR, 2019, pp. 5680–5689.
54. Tripp-Barba, C.; Urquiza-Aguiar, L.; Igartua, M.A.; Rebollo-Monedero, D.; De la Cruz Llopis, L.J.; Mezher, A.M.; Aguilar-Calderón, J.A. A multimetric, map-aware routing protocol for VANETs in urban areas. Sensors 2014, 14, 2199–2224. [Google Scholar] [CrossRef] [PubMed]
55. Aliyu, A.; Abdullah, A.H.; Isnin, I.F.; Radzi, R.Z.R.M.; Kumar, A.; Darwish, T.S.; Joda, U.M. Road-based multi-metric forwarder evaluation for multipath video streaming in urban vehicular communication. Electronics 2020, 9, 1663. [Google Scholar] [CrossRef]
56. Inzillo, V.; Ariza Quintana, A. Implementation of 802.11ax and cell free massive MIMO scenario for 6G wireless network analysis extending omnet++ simulator. Simulation: Transactions of the Society for Modeling and Simulation International 2024, FORTHCOMING, -, –.