The application of advanced computational algorithms used for cooperative communication transmission of vehicular networks: a proposed method

Collaborative communication in vehicular networks

The basic idea of collaborative communication is based on a multiuser communication environment. A single-antenna mobile user forms a virtual multi-antenna array by sharing the antennas of multiple mobile terminals nearby, thereby gaining diversity and effectively improving the reliability of information transmission (Dias et al., 2020). The transmission method of collaborative communication is not a simple relay technology or the previously applied spatial diversity technology but a fusion of the two technologies. Therefore, collaborative communication technology has the advantages that both have. So, communication coverage has been expanded and shadow fading, path loss, and multipath effects have been effectively overcome. The introduction of collaborative communication technology into telematics represents a virtual multi-antenna environment where the source vehicle communicates directly with the destination vehicle. At the same time, the relay vehicle also forwards information to the source (Zhang et al., 2018), which undoubtedly brings diversity gains and extends the range of information transmission. However, such a high-speed in-vehicle environment leads to problems. The introduction of relaying is bound to cause more latency and reduced bandwidth utilization. Therefore, adaptive collaboration models and relay selection algorithms in vehicular networking should receive extensive attention.

The adaptive collaboration model in telematics consists of two levels of meaning. One of which is the adaptive relay collaboration policy, i.e., the judgment of relay “collaboration timing” (Fotohi, Nazemi & Aliee, 2020), which can effectively resolve the problem of “to forward or not to forward” and “when to forward” for relay nodes. The second is the adaptiveness of the relay forwarding policy, i.e., the choice of the relay node’s information forwarding method when choosing a relay collaboration communication method between vehicles. Amplify-and-forward and decode-and-forward are two of the simplest yet widely implemented fundamental forwarding strategies in collaborative communication systems. The relay selection algorithm in telematics is employed to choose the best-performing relay among all potential relay nodes when picking a relay collaboration method for communication between vehicles (Rivoirard et al., 2018).

In summary, the research of adaptive collaboration modes and relay selection algorithms in telematics can effectively resolve the problem of reduced bandwidth utilization and increased latency brought about by the introduction of collaborative communication technologies, thus improving the reliability performance of the system.

Vehicles in telematics can support a range of secure and non-secure applications with vehicle-to-vehicle and vehicle-to-infrastructure communications. Considering that a single wireless communication method cannot simultaneously meet the QoS requirements of multiple telematics applications. The current telematics research has focused on heterogeneous architectures of telematics (Wang et al., 2022). In China, Dating and Huawei have been improving and testing the V2X architecture based on LTE-V, which is based on TD-LTE technology and supports both centralized LTE-V-Cell and distributed LTE-V-Direct communication modes, which could reuse the available cellular network infrastructure without allocating dedicated spectrum, and could effectively extend the vehicle sensing range with the help of LTE-V-Direct and ensure network connectivity. The LTE-V enables ubiquitous network coverage with cellular communications, enabling the interconnection of vehicle-vehicle, vehicle-road, and vehicle infrastructure, thus effectively supporting telematics applications.

Advanced computational algorithms in the vehicular networks

In recent years, vast amounts of data have been accumulated in various fields and the human capacity used for collecting (Ismael et al., 2021), storing, transmitting, and processing data has increased rapidly. These developments lead to advanced computational algorithms, which are important in enabling revolutionary advances in artificial intelligence (AI).

Advanced computational algorithms (Bangui & Buhnova, 2021) are classified in terms of dataset labeling: unsupervised, semi-supervised, supervised, and reinforcement learning algorithms (Prasad et al., 2021). Where supervised learning algorithms master a prediction function from a training dataset and estimate the output based on the new dataset. Commonly employed supervised learning algorithms include regression analysis and statistical classification methods. The training set for supervised learning algorithms consists of feature inputs and target outputs, where humans annotate the targets and all training data has annotation information. In contrast to supervised learning algorithms, unsupervised learning algorithms model the features of the training set directly and identify the underlying class rules based on the same features and targets of the training set. The difference comes from whether humans annotate the targets of the training set. Training sets of the unsupervised learning algorithms (Islam et al., 2021) have no human annotation information and common unsupervised learning algorithms are used for visualization, cluster analysis, and density estimation.

On the other hand, semi-supervised learning algorithms are somewhere between supervised and unsupervised learning algorithms, using partially labeled and partially unlabeled data to master predictive functions. Finally, reinforcement learning algorithms are about achieving a goal, gradually adjusting behavior as the environment changes, and evaluating whether the feedback obtained after each action is positive or negative. Reinforcement learning algorithms are classified as instantaneous difference, dynamic programming, and Monte Carlo methods. There is still much room for research into the generalizability of those algorithms, i.e., their ability to generalize needs further improvements (Al-Shareeda et al, 2022).

In recent years, many technologies have contributed significantly to the development of connected vehicles, where autonomous driving by computer is an ideal solution in the field of connected vehicles. Abstract autonomous driving is an advanced computational-based research, where the inputs are various types of information received by multiple onboard sensors that sense and acquire information on vehicle conditions in real-time, and the outputs are control actions such as throttle, direction, braking, and other functions. Also, due to the cyclical nature of urban traffic flows, data traffic in telematics has a regularity in the spatio-temporal dimension (Ashish & Prakash, 2021). When vehicles are associated with base stations of heterogeneous networks, one of the most important issues is to exploit the regularity of vehicle traffic data for load balancing between these base stations. For certain tasks, the collected data is stored and processed into the training and test sets, respectively, and the error is calculated separately. The error of the test set is then employed to represent the generalizability of the model and algorithm and the optimal strategy is again mastered to prove the algorithm’s effectiveness.

When the performance of telematics is analyzed, collaborative communication technology involves algorithms of advanced computations. Middleware technology is the core technology in the current software development field, which realizes the middle program of data transmission, data filtering, format conversion, and other operations between radio frequency identification (RFID) hardware equipment and the user’s application system and imports various data information read by RFID reader into telematics application through the operation of software extraction, decryption, filtering, format conversion and so on, and makes it available to telematics users through the user interface of application system (Adnan et al., 2021). Specifically involved are middlewares such as vehicle path navigation, emergency event handling, vehicle-assisted driving, and traffic signal control. Each one is developed regarding the standards for telematics application services. Convolutional neural networks (CNN), which are sweeping the autonomous driving technology and connected vehicle fields, employ a greedy layer-by-layer training approach to master deeply, perform feature extraction, estimate unknown functions, and utilize multiple convolutional layers to generate deeper feature maps to reach more accurate estimates of network functions. The main applications based on deep learning techniques are neural network models, sensor fusion methods, and path planning methods (Hossain et al., 2021). The reinforcement learning-based distributed machine-to-machine communication packet scheduling algorithm is employed in telematics communication to support million-level connectivity requirements, and the accuracy of the algorithm’s predictions is evaluated using Pearson’s linear correlation coefficient to measure the degree of relationship in the training set. While objective evaluation algorithms are measured, the degree of deviation of subjective evaluation algorithms is a function of the root mean square error (RMSE). Using different advanced computational algorithms for different telematics needs to study the behavior of vehicle-vehicle (Elgendy et al, 2021), vehicle-roadside unit, and vehicle-pedestrian interactions in telematics and learning and exploring the overall perceptual environment of telematics. All involve the effective application of advanced computations.

The introduction of a collaborative communication system in telematics, while bringing diversity gains and improving the system’s reliability, faces the problem of introducing relaying that will inevitably lead to more delays and a reduction in bandwidth utilization. In addition, the adaptive nature of the relay forwarding strategy, i.e., the relay selects the forwarding method based on the channel transmission characteristics of the source-relay when a relay collaborative communication method is chosen between vehicles. Thus, amplify-and-forward and decode-and-forward are two of the simplest yet widely implemented fundamental forwarding strategies in collaborative communication systems.

The research of adaptive collaboration modes and relay selection algorithms in telematics can effectively resolve the problem of reduced bandwidth utilization and increased latency brought about by the introduction of collaborative communication technologies, thus improving the reliability performance of the system. Therefore, a multi-intelligence advanced computational algorithm is proposed to construct a joint communication transmission mechanism in the telematics communication system model in the manuscript and is simulated. The outcomes suggest that it has a better cooperative communication transmission.

A communication model of vehicular networks

Telematics is essentially a multi-source, multi-relay, multi-destination, and multiuser network. Still, analyzing the three types of nodes concurrently when they are moving simultaneously is very complex, down to the data link layer, network layer, and other protocols. Only the physical layer performance analysis is considered here to incorporate collaborative communication techniques, and the system model is shown in Fig. 1. A two-hop transmission model consisting of a source end s, several relay ends r, and a destination end d is only considered.

First, the source broadcasts a message to the destination and the relay, and the received signals y_s,d y_s,r for both are calculated in Eqs. (1) and (2). (1)${\displaystyle y_{s,d}=\sqrt{P_s}{h}_{s,d}x+z_{s,d}}$ (2)${\displaystyle y_{s,r}=\sqrt{P_s}{h}_{s,r}x+z_{s,r}}$

Then, the instantaneous channel transmission characteristics of the source–destination, source-relay, and relay–destination channels are taken into account to determine how the source–destination carries out information transmission, i.e., it shows when to use the collaborative communication model and when to use the direct transmission model. When the direct transmission model is employed, the relay is not involved in the communication transmission, and the instantaneous received signal-to-noise ratio of the system is calculated in Eq. (3). (3)${\displaystyle r^d=\frac{P}{N_0}|h_{s,d}{|}^2.}$ When the collaborative communication mode is implemented, the relay end is involved in the communication transmission. Assuming that the maximum ratio merging method is used at a destination end, the instantaneously received signal-to-noise ratio of the system at this time is calculated in Eq. (4). (4)${\displaystyle r^{{}^{\prime }}=\frac{P_s}{N_0}|h_{s,d}{|}^2+\frac{P_s}{N_0}|h_{r,d}{|}^2}$

Advanced computational algorithms when multi-agent exists

The manuscript employs a multi-intelligence advanced computational algorithm to construct a collaborative communication transmission mechanism in the telematics communication system model. The framework of the model consists of K intelligence. A deep deterministic policy gradient algorithm implements each of these, and the framework is shown in Fig. 2.

The deep deterministic policy gradient (DDPG) algorithm (Bao et al., 2021) combines the benefits of policy gradients and DQNs. It consists of two types of neural networks: a policy-based actor-network and a value-based critic network. The actor-network collects the state of the environment and then performs the appropriate action based on the policy. The Critic network employs a state-action value function to evaluate the goodness of the action chosen by the actor-network based on the policy $Q_k\left(\cdot \right).S_t^k$ represents the input state of the intelligence k, γ denotes the discount factor of the immediate reward $R_t^k$. The state-action value function of the DDPG algorithm is calculated in Eq. (5). (5)${\displaystyle Q_k\left(S^k,A^k\right)=E\left[R^k+\gamma Q\left(S^{k^{\prime }},A^{k^{\prime }}\right)\right]}$

As a policy gradient algorithm, the main idea of DDPG is to obtain an optimal policy ${\pi }_k^{\ast }$ and learn the state-action value function corresponding to the optimal policy ${\pi }_k^{\ast }$ until convergence occurs. DDPG’s Actor-Critic network employs a dual network structure. One for the evaluation network and the other for the target network are employed, where the ${\theta }_k^{\upsilon }$ and ${\theta }_k^Q$ of the evaluation network are updated in real time. The update process starts with a small randomly selected sample of experiences from the experience pool and feeds them into the intelligence one by one. In the training phase, the Actor and Critic networks update the parameters of the evaluation network based on small batches of samples for each input. The Critic network adjusts the parameters of the evaluation network by reducing the losses, and the loss function is calculated in Eq. (6). (6)${\displaystyle L\left({\theta }_k^Q\right)=E\left[{\left(R_t^k+\gamma Q_k^{{}^{\prime }}\left(S_t^{k^{\prime }},A_t^{k^{\prime }}|{\theta }_k^{Q^{\prime }}\right)-Q_k\left(S_t^k,A_t^k|{\theta }_k^Q\right)\right)}^2\right]}$ When the Actor makes an action decision for each observation and each intelligence aims to maximize the cumulative payoff. The evaluation network parameters of the Actor are updated by maximizing the policy objective function, which is calculated in Eq. (7). (7)${\displaystyle J\left({\theta }_k^{\upsilon }\right)=E\left[Q_k\left(S_t^k,A^k\right)|A^k=\upsilon \left(S_t^k\right)\right]}$ where u_k(⋅) denotes an Actor evaluation network function for a deterministic strategy π_k of state-mapped actions. Since the action space is continuous, $J\left({\theta }_k^u\right)$ is continuously differentiable and ${\theta }_k^u$ can adjust the direction of gradient descent ${\nabla }_{{\theta }_k^u}J\left({\theta }_k^u\right)$. As the parameters ${\theta }_k^u$ and ${\theta }_k^Q$ of the evaluation network are continuously updated, the parameters ${\theta }_k^{u^{\prime }}$ and ${\theta }_k^{Q^{\prime }}$ of the target network are updated using a soft update as follows: (8)${\displaystyle {\theta }_k^{{\upsilon }^{\prime }}=\tau {\theta }_k^u+\left(1-\tau \right){\theta }_k^{u^{\prime }}}$ (9)${\displaystyle {\theta }_k^{Q^{\prime }}=\tau {\theta }_k^Q+\left(1-\tau \right){\theta }_k^{Q^{\prime }}}$

Power distribution

When a transmitting node sends a message to a receiving node, the transmitting node generates channel contention. Only some nodes can communicate with the receiving node, at which point the other nodes are treated as interfering nodes. The power allocation in a collaborative communication system is divided into two steps: the determination of the collaborative transmitting node i that communicates with the receiving node r and the optimization of the power of the transmitting node s and the collaborative node i to increase the channel capacity of the receiving node r.

Here, the SINR capture effect model is implemented to select collaborating nodes. When the ratio of node signal i strength to node s channel strength is greater than the threshold value β, node r can receive the transmit signal from node i. Then, such transmit node i is defined in Eq. (10). (10)${\displaystyle \left\{C_r|SINR=\left(pC{R}_{i,r}GC{R}_{i,r}\right)/\left(p_{i,r}{G}_{s,r}\right)\ge \beta \right\}}$ where C_r denotes the set of collaborating nodes and Z_r denotes the set of all vehicle nodes that cause interference to the receiving node r, except C_r.

In summary, the amount of interference the receiving node receives can be expressed in Eq. (11). (11)${\displaystyle I=\sum _{j\in Z_r}\sum _{k=1,k\ne r}^N{G}_{j,r}{P}_{j,k}+n_0.}$

In a collaborative communication system, when a receiving node is close to a transmitting node, its SINR is easily satisfied by that transmitting node, and the power relationship between p_i,r of other collaborative transmission nodes and that transmitting node can be expressed in Eq. (12). (12)${\displaystyle P_{i,r}=P_{m,r}{G}_{i,r}/G_{m,r}}$

Simulation environment

The telematics wireless communication scenario presented in the article helps define the telematics simulator in the city, including vehicles, lanes, and a wireless communication network model. One of the V2V links corresponds to one of the V2I links. Each intelligence’s Actor and Critic networks consisted of three fully connected hidden layers containing 256, 64, and 16 neurons, respectively. The activation function is a modified linear unit, and an adaptive moment estimation optimizer is employed to train the weights of the updated neural network iteratively. The algorithm is trained for 2,000 episodes, adding a variable Gaussian noise to the actions selected by the intelligences. The probability of exploration is handled by a linear annealing algorithm, starting from 1 and annealing to 0.02 at 1,600 episodes, with the likelihood of exploration remaining constant in the subsequent training steps. In the article, a V2V link load of 1,060 megabytes is chosen for the algorithm’s training phase, and a different load size is utilized when compared with other benchmark algorithms to verify the algorithm’s robustness.

Simulation results

The convergence of the algorithm

Figure 3 shows the variation of rewards with an episode, which can observe the convergence performance of the multi-antenna deep deterministic policy gradient (MADDPG) algorithm. In the article, 100 training steps are set in each episode, and the cumulative reward sum of all training steps is utilized as the reward. The analysis shows that the reward score gradually increases as the training iterations increase. The proposed algorithm starts to converge when the training reaches 500 episodes, where the network topology changes rapidly due to the higher mobility of the vehicles. The channel fading fluctuates, leading to numerical fluctuations in the convergence of the proposed algorithm.

The channel capacity at different nodes

Figure 4A represents the relationship between transmit node power and channel capacity for different communication schemes based on when vehicle nodes n = 20 are picked. The optimized channel capacity for collaborative power allocation and the allocation of cooperative communication power decrease as the power of the transmitting node increases, and the channel capacity with no collaboration keeps increasing with the power of the transmitting node. However, the channel capacity of the first two still far exceeds that of the latter. As the power of the transmitting node increases, the gap between the collaborative power allocation’s channel capacity and the collaborative communication’s optimized power allocation keeps decreasing. Figure 4B indicates that the channel capacity of the cooperative power allocation is slightly better than the cooperative communication power optimization allocation when the transmit node power is 5-6 dBm when the vehicle node n = 40 is chosen. However, when the transmit power exceeds 7 dBm, the channel optimized capacity by collaborative communication power allocation gradually exceeds that of cooperative power allocation. The channel capacity without collaboration is still much smaller than the first two. Figure 4C indicates that the communication quality of the optimized allocation of collaborative communication power is consistently more significant than that of the cooperative power allocation when the vehicle node n = 60 is picked.

Transmission success rates under different loads

Figure 5 shows the performance of each algorithm in optimizing V2I link capacity and performance under different V2V link payloads. The experimental results show that as the V2V link payload size increases, all algorithms’ optimized V2I capacity and performance show a decreasing trend. To increase the probability of a successful V2V payload transmission, the increase in V2V payload leads to more substantial interference on the V2I link for a longer period, thus jeopardizing its capacity performance. Thus, an increase in the V2V link load will result in a longer time for the V2V link to transmit data and a higher transmit power, leading to more interference with the V2I link and consequently causing a lower capacity. Under the same load conditions, the total capacity of the V2I link optimized by the proposed algorithm is still higher than that of the other three baseline strategies. As the load increases, the total V2I link capacity of the three baseline strategies tends to decrease continuously. At the same time, the performance of the proposed algorithm shows a slight fluctuating decrease with better robustness.

Figure 6 shows the relationship between the transmission success probability of the V2V link and the V2V link transmission load size. The likelihood of the transmission success of all four optimization strategies decreases as the load increases since the amount of transmission load is fixed at the same transmission rate and the same transmission time slot, while the larger the total load size required for the V2V link is, the lower the transmission success probability of the V2V link load would be.

Simulation results and analysis of connectivity

To evaluate the connectivity of collaborative communication transmission in-vehicle networks with the introduction of advanced computational algorithms in the article, a simulation scenario is set up to analyze algorithms dealing with the connectivity of different mobile access point selections, in which vehicles obey a Poisson point process with a density of 0.02 m-1 on a 10 km road, and the speed follows a uniform distribution within [50, 80] km/h. The speed of each vehicle remains constant during the simulation. Assume that the communication channel obeys Rayleigh fading, the path loss index is assigned to 4, the transmit power of the mobile access point is assigned to 2W, and the vehicle receives an additive noise power of −100 dBm. In the simulation, mobile access points are selected according to different selection algorithms, and adjacent mobile access points communicate with the vehicles in between, employing a collaborative communication mode. The impact of different mobile access point selection algorithms is evaluated by comparing the received signals with the interference-to-noise ratio (SINR) and the SINR threshold to calculate the overall probability of the downlink connectivity.

Figure 7 shows the relationship between the probability of the downlink connectivity PC and the SINR’s threshold β received by the vehicle when communicated collaboratively with two collaborating mobile access points. Three different mobile access point selection methods are compared in the simulation: independent random, sequential, and distance-based selections. The simulation results also show that the sequential and distance-based selection methods are better adapted to the actual distribution of vehicles, and their connectivity is significantly better than that of independent random selection.