1. Introduction

Vehicular edge computing (VEC)is an enabling technology for Intelligent Transportation Systems (ITS), and the Internet of Vehicles (IoV) facilitates feasible solutions enabling computation capabilities among vehicles ^[1]. In wireless networks, an exponential increase in demand for high computation potentiality has arisen due to continuous increases in mobile applications ^[2]. On the other hand, cloud computing is a proper platform designed to support offloading of computationfrom mobile devices ^[3]. Those data centers, being remote networks, increase latency and network delay, thus affecting the performance of IoV applications in an ITS ^[4]. For critical applications, a local approach to offloading is mandatory ^[5].

Along with rapid growth of the IoV and artificial intelligence (AI), vehicle networks have become smart unified networks. Vehicular edge computing enhances services by offering computational offloading services in the vicinity of vehicles. Intelligent Connected Vehicles (ICVs) connect themselves as well as linking to urban traffic networks to execute intelligent applications. These applications are generally delay-sensitive as well as compute-intensive such that the available resources are not capable of meeting service demands by all vehicles ^[6,7].

Offloading techniques are used in VEC settings to move resource-demanding tasks to neighboring servers to increase the effectiveness of vehicle services ^[8]. Offloading in VEC transfers resource-intensive programs that support local vehicular devices in order to decrease the workload, overhead, and expense of local execution. To satisfy computation offloading, both vehicular devices and VEC servers must operate offloading frameworks. Many technical publications examine this topic and suggest novel strategies for reaching the objectives in offloading criteria ^[9]. To the best of our knowledge, deep learning (DL) have not the subject of any study or review regarding offloading, despite its significance and the need for academics to work in the field ^[10]. Therefore, this survey examines recent research, and investigates various VEC paradigm strategies, systematically covering DL-based approaches.

The key contributions of this article are:

· reviewing survey studies on deep learning offload techniques for VEC, highlighting each one's benefits and drawbacks,

· investigating the most recent deep learning methods for offloading in VEC,

· giving a systematic evaluation of present methods, suggesting a detailed taxonomy, and

· Deliberatingfuture research issues to increase compu-tation offloading in VEC environments.

The remainder of this article is arranged as follows. Section 2 explores thetaxonomy of task offloading techniques. Section 3 presents optimization applied in task offloading. Different ML techniques for offloading tasks in VEC networks are explained in Section 4. Section 5 discusses various DL techniques for offloading tasks in VEC networks. A comparison and analysis are presented in Section 6. Finally, the conclusionis presented in Section 7.

2. Taxonomy of Task Offloading

On roads with increased traffic flow, the server’s calculation limitation threatens the distinction of the offloading facility ^[11]. By positioning mobile edge computing (MEC) servers at the network edge, the computational burden on ICVs can be significantly alleviated by offloading ^[12,13]. It is essential to design a suitable architecture that focuses on improving quality of service (QoS). By using mobile edge computing as a distributed model providing proficient resources [closer to vehicles, the response time in the network can be reduced to a greater extent ^[14].

A task offloading algorithm proficiently reduces the delay and resource consumption in multiple-user and server VEC environments ^[15]. The offloading algorithm finds the best position for task deployment and decides on the implementation order of the tasks at the server. A taxonomy of existing task offloading algorithms that have contributed to the literature in recent years is detailed in Fig. 1.

3. Optimization Techniques Applied in Task Offloading Strategies

This section provides a brief summary of optimization approaches used to address task offloading difficulties. Itoffers an overview of the fundamental optimization techniques, including a few recommended readings, a survey of approaches in the literature, and a flowchart outlining an application state for an optimization strategy.

For real-time, latency-sensitive applications, the task offloading issue in vehicle edge computing is crucial. While taking into account network delays, the processing period in fog nodes, network bandwidth, and current loadsonfog nodes, the data engendered by VEC should be balanced across the fog nodes. Therefore, taking into account the network properties and the current load on the fog nodes, the selected fog device for taskoffloading must fulfill the given QoSlimits, notably the response time. This is an NP-hard task. Because sensors and fog nodes is increasing, the problem gets exponentially harder. Therefore, employing conventional greedy search techniques is difficult. The presented approaches are metaheuristic algorithms employing ant colony optimization (ACO) to overcome this challenge. An ant's capacity to locate the quickest route from the colony to a food supply is the inspiration for the probabilistic metaheuristic technique known as ACO. Finding the shortest path is one of several optimization issues that are solved using the collective foraging behavior of living ants. An ant first departs from its colony and travels along a course chosen at random while looking for food sources nearby. Pheromones, a type of chemical signal, are released by ants to communicate with one another informally. The amount of pheromone released by the ant along the route back to the colony is related to the amount and worth of the food it has found. As a result, other ants will be more likely to follow trails with high pheromone concentrations.

In the end, every ant will take the quickest and safest route between the food source and the colony. Each of the$m$ants chooses path $n$ with the highest pheromone concentration from among all potential paths. The travelling salesman and task-scheduling problems are two NP-hard research topics for which ACO is successfully implemented as an optimization metaheuristic. Moreover, the ant colony optimization algorithm is applied to related cloud scheduling issues, such as scheduling VM in cloud resources, and scheduling tasks in VMs with the intention of load balancing and lowering the response time. The ACO algorithm is also employed for cloud-based edge computing task scheduling. ACO is employed to schedule tasks for fog computing with deadline awareness in a tiered edge computing architecture. The suggested algorithm concentrates on increasing a fog service provider's revenue while taking into account the task completion deadline restrictions for vehicle edge computing tasks. A workload task is offloaded via sensor $R_{i}$ with probability of success expressed in Eq. (1) when searching for a fog node,$fh_{i}$, that can decrease the response time:

where$\alpha $ and $\beta $ are heuristic constants; $\alpha \geq 0$ denotes heuristic parameter that regulates the pheromone quantity,$\beta \geq 1$ is the heuristic parameter that describes the virtual quality of task offloading; $\eta _{ij}(w)$ symbolizes the heuristic function that signifies task offloading quality. It is evaluated as follows:

where$load_{j}$ calculates the load on fog node $j$, and when$S_{ij}$increases, $load_{j}$decreases and $\eta _{ij}(w)$increases. Therefore, there is little chance that the scaling of sensor $i$ will be transferred to fog node$fh_{i}$. The amount of pheromone on the trail for task offloading left by ant $l$during iteration $(w)$is represented by${\tau }_{ij}^{k}(w+1)$, which is calculated with Eq. (3):

where $\Delta {\tau }_{ij}^{k}(w)=1/S_{ij}$, and $\rho $epitomizes the proportion of pheromone evaporation that simulates the pheromone evaporation effect at every single stage. Furthermore, if all the ants accomplish task offloading, the pheromone trail is globally updated, i.e., iteration is completed.

By making a continuous link between edgewith cloud, also shortening the effectual distance, fog computing gets around the restrictions of the cloud. However, fog computing has significant difficulties when offloading work for remote computation. Therefore, the key research focuses in fog computing is optimality of work offloading. The particle swarm optimization (PSO) algorithm ^[16] is utilized for task offloading in vehicular edge computing. Particle swarm optimization is called a metaheuristic optimization approach because it relies on a swarm size,$N_{swarm}$, which is a population of collaborating individuals. Each member of the swarm (each particle) represents a place in the search space that corresponds to a solution. The expression for a particle,$Q_{i}$,on D-dimensional vector is known as D-dimensional search space,$Q_{i}=\{q_{i1},q_{i2},\ldots q_{id}\}$. Each particle's search movement is determined by the global optimal discovered for all particles as well as the local optimum. Eqs. (4) and (5) illustrate how all elements of$Q_{i}$appraise their newest location, $A_{i}(w+1)$, and depend upon computing the velocity,$U_{i}(w+1)$:

where$i=1,2,\ldots ,N_{swarm}$, and $w=1,2,\ldots ,iter_{\max }$denotes the number of iterations;$\omega $implies preliminary velocity of inertia weight that balances exploration with exploitation. At the beginning of the search, ahuge inertia weight value is appliedif exploration is preferred, whereasa lesser value facilitates more exploitation. Learning coefficients are $D_{1}$ and$D_{2}$;$\alpha $and $\beta $are consistently supplied random counts in the range ^[0,1]. PSO was first suggested as a method for resolving issues in continuous domains. PSO needs to be adjusted to the domain, because the planning of mobile task offloading happensin a separate search space.

The position of particle $P_{ij}$exemplifies sensor node offloading by$S_{node}$to fog node$f_{node}$. Demonstrations of the task offloading sensor and the fog nodes are separate. The particle’s place and velocity are updated in all iterations by Eqs. (4) and (5). Nevertheless, the value of every particle position needsisconverted asdistinct numerical value using Eq. (6):

where$N_{node}$ is the fog node count, and $\left\lfloor | P_{ij}| \right\rfloor $ signifies maximization of$P_{ij}$exact value for particle$P_{j}$. This approach begins by setting more iterations, $N_{iter}$; the parameters updated by velocity are$\alpha $, $\beta $, and$\omega $, and the probability of mutation is $P_{mutation}\,.$ Additionally, the proposed particle swarm optimization algorithm sets the positions of particle $P_{i}$and velocity $U_{i}$ to random values. Every particle is evaluated with fitness function,which exemplifies the particle encryption, exposing the solution’s quality. The fitness is evaluated with Eq. (7):

The solution is considered better when the fitness function value is higher. Every particle’s local optimum is determined after computing its fitness. The global optimum is set at the swarm's highest fitness level. The algorithm repeats itself for the set number of iterations$N_{iter}$. Eqs. (4) and (5) are used to update each particle's velocity and location after each iteration, and Eq. (6) transforms each particle's position into a discrete number. The novel local and worldwide optima are continued via the PSO algorithm. Every single particle is then subjected to a random mutation. To create novel task offloading for particles, particle values for a pair of random nodes are switched throughout the mutation process.

Numerous tasks are sent to a cloud server because mobile devices in cloud environments have processing limits. In the 20 years following the introduction of the cloud paradigm, this has resulted in an increase in the effectiveness of mobile applications. However, because a cloud server is typically located far from mobile users, task offloading might not be an appropriate solution for mobile applications that are delay-sensitive. A joint optimization algorithm (JOA) ^[17] is applied to solve this issue. In this model, the edge server near the access point (AP) is not required to handle a vehicle's computing needs. In order to achieve system load balancing, this may lessen the weight in a few hot edge servers.

Additionally, the JOAaids in enhancing QoS and cutting down on queuing delays for computing tasks carried out on edge servers. The distance issue consequently introduces an additional communication lag. The task is deployed to an edge serverdenoted$j$; queuing delay is indicated by the notation ${W}_{j,k}^{Q}$calculated by adding the evaluated implementation time in the edge server for every task in $j$’s queue and is the task's authentic accomplishment time. The communication latency from AP $i$ to edge server$j$is Eq. (8):

where$dis_{i,j}$stands for the distance between edge server $j$ and AP $i$. If the vehicle's computing function is transferred to an edge server close to the associated AP, then${W}_{i,j}^{C}=0$.

An entire task is separated as task units that are offloaded to numerous edge servers in successive cells,since the data size of task $u_{l}$ is enormous, each cell's coverage is just marginally enough. A maximal count of tasks should be finished in each cell to meet the deadline for task completion. Based on accessible computing with network resources atthe cells, together with maximal count of TUs, the task could be completed in the local vehicle.Also needed maximal count ofTUs finished by edge servers:${N}_{k}^{s}$.For cells in which tasks can be finished in the vehicle,we assume that vehicle $k$is entering the coverage cell$L_{s}$, time the vehicle stays in the cell can be calculated with Eq. (9):

where${v}_{k}^{s}$denotes speed of vehicle $k$in cell. Depending upon the time the vehicle stays in the cell, calculating the maximalcount of TUs completed locally is done with Eq. (10):

A maximal count of task units processed denotes${N}_{off,k}^{max,s}$, is equal to the number of task units that the vehicle can offload onto an edge server. The length of time spent in the cell, the channel conditions there, and power of the edge server’s processor affects the value of${N}_{off,k}^{max,s}$. The rate for data transmission within the cell is represented by the medium rate of uplinks toall APs, as described in Eq. (11):

The average scaling computing capacity in the cell is estimated by Eq. (11), which is the average value for processing offloaded tasks.

4. Machine Learning Task Offloading Techniques for VEC Networks

Different machine learning methodsare utilized for optimal task offloading in VECnetworks. Recently, task offloading in VEC networks has been performed through various deep learning strategies. Some of them are discussed below.

To meet the demand for rapid offloading in vehicle networks, we propose an effective task offloading algorithm that depends upon the support vector machine (SVM) ^[18]. Through a weight allocation mechanism that takes into account the MEC servers' available resources, the algorithm can divide a large task into multiple smaller ones. Following that, SVMs are used to determine if every sub-task should be implemented locally or offloaded. If they are offloaded, sub-tasks are assigned when the vehicle passes MEC servers. Every server guarantees that the sub-task will be completed and returned on time. Along a straight road, $N$ road side units (RSUs) are connected to a MEC server. Vehicles access a road side unit only where they are situated, and each RSU covers a single region. The set of mobile edge computing servers or road side units is expressed in (12):

The set of MEC servers is expressed in (13):

The proposed offloading technique assumes moving vehicles segment tasks into minor tasks. The first sub-task is uploaded to$M_{1}$when the vehicle first enters region$R_{1}$, and $M_{1}$then completes it as the vehicle moves. Before leaving$R_{1}$, the vehicle can quickly receive the result. Then, the final decision function for task offloading is expressed in Eq. (14):

With respect to the resolution operation, Task $X$is categorized such that if$f\left(X\right)> 0$, the related$y=+1$, and the task is offloaded. If$f\left(X\right)< 0$, resultant$y=-1$, and the task is completed locally.Task offloading using the SVM maximizes the efficiency of the energy consumed by offloading decisions and from the allocated resource block count.

Unsupervised machine learning uses the K-means clustering algorithm ^[19],which is necessary to train the type of model used to categorize the tasks in VEC networks. The freshly arrived task is then assigned a set of tasks with comparable features. Here, three task features,which are seen in Eq. (15), are considered for task offloading and classification:

The tasks stochasticallyallocated forcomputing nodes to determine their features. Afterimplementing several tasks, the task administrator creates task lists of trained data with separate features and workload categories, likeOLTP, streaming, web serving,graphics. Euclidean distance is applied to choose the adjacent centroid deeming$C_{k}\left(K=1,2,3\right)$, offloading them to $t_{j}$ as expressed in Eq. (16):

Next, measuring the relation in terms of$d\left(t_{j}-c_{1}\right)$, $d\left(t_{j}-c_{2}\right)$, and $d\left(t_{j}-c_{3}\right)$,task offloading is categorized into clusters withthe least identical value. The task offloading system is inappropriate for highly incorporated or comparatively simple tasks that cannot be partitioned.

5. Deep Learning Task Offloading Techniques for VEC Networks

This section presents different deep learning methods utilized for optimal task offloading in VEC networks. Recently, task offloading has been conducted through various deep learning strategies as discussed below.

Deep Q-learning ^[20] empowers task offloading for vehicular edge computing in urban informatics. To proceed with optimal task offloading, theknown offloading tactic is initially denoted$\pi $, which is derived from a moderate system $b^{l}$action in state $S^{l}$ using the Q-function expressed in Eq. (17):

The value and the strategy iteration can be used to determine the most utility as well as the best task offloading tactic. Iterations can be changed using a Q-learning technique used in reinforcement learning. The Q-value functioningin learning procedure is adjusted in each iteration and is expressed in Eq. (18):

where$\alpha $indicates the learning rate. Q-function assessment basis, applying deep Q-learning attains optimum offloading strategies denoted$\pi ^{\ast }$. The neural network base approximates the Q-network, and the group of network parameters is denoted$\theta $. By utilizing Q-network, the Q-function in Eq. (17) is used to analyze Eq. (19):

Depending on$Q'$, the optimum task offloading method in every state acquired from the eventsresults in higher utility. Hence, deep Q-learning in terms of task offloading for VEC minimizes the cost of using computing resources with higher latency.

Convolutional neural networks (CNNs) ^[21] are used for task offloading strategy prediction in VEC. Atask offloading system with the help of deep learningpredicts the offloading result with a binary classifier (i.e., success/failure), and service latency prediction accounts for any backslide. Calculating the task offloading efficiency also creates a task offloading decision depending on predicted results, and the mode is determined along with the offloading data history. This method containsthreelayers: vehicle, edge, and cloud. The vehicles are in contact with road side units through wireless channels. Vehicles offload tasks to roadside units via vehicle-to-infrastructure (V2I) wireless communications with edge servers and cloud servers; secondly to a cloud server through the RSUs; and third,toa cellular BS. Here, the CNN is utilized for task offloading in the VEC network. The CNN primarily comprises three modules: convolution, pooling,recombination layers. The convolution collects local informationsthen creates features. The recombination produces newlyfeatures owingto a reduction in the number of features. Combining raw features to get new features increases the input feature space. In this strategy, every component is designated as seen below. After preprocessing a feature, every single input sample is converted from a $14\times 1$ matrix to a $16\times 40$matrix anywhere in which $40$is the implanting measurement chosen for the embedding process period, and two more 14 x 16 rows are then added to the raw features. A$16\times 40$matrix is sent to convolutional layer, thenthe feature map output becomes$\left(16,40,6\right)$.

The convolutional layer output is compressed by the pooling layer. To maintain the observation window maximal value, the suggested model uses max-pooling at thepooling layer. In general, max value frequently offers more details than the middle value. Mean-pooling typically reduces a feature map's data.

An$h\times 1$pooling window handles the convolution output. For feature consolidation, the width is fixed to 1 to confirm that the feature map could not subsample at width measurement. This hypothetical 1\textsuperscript{st}convolution output implies$C^{1}$, first pooling layer output implies$P^{1}$, which is expressed in Eq. (20):

Consider$x$,$y$indicates row and column feature map index. The second convolutional layer receives its input from the first pooling layer's output,as provided from Eq. (21):

where$P^{i}$indicates output of pooling layer $i$,$E^{i+1}$indicates convolutional layer$i+1$input. After convolution and pooling, the recombination layer is scheduled. Local feature crossings exist between the feature maps that travel via the convolutional and pooling layers. However, directly sending feature maps to a densely linked layer as the pooling layer's output results in global feature loss together withlessdata density. To overcome these, the pooling layer's output is combined again in the recombination layer by using an entirely allied neural network. The recombination layer feature output is new features created through feature crossover. The task offloading prediction model is updated to include both new features and raw features. New features are defined with (22):

where$i$indicates total count of convolution, pooling, recombination cycles. Afterwards, by using (23), the new features are added to base features:

Let$E$ denotes input, $E'^{T}$indicates the service delay, and$R^{T}$denotes the offloading result. The prediction model contains some hidden layers, then the 1\textsuperscript{st}hidden layer input for completely linked network is represented by $I_{1}$in (24):

where$Flatten$indicates original 2 dimensional feature, and$E$is the matrix converted to a vector. Output of hidden layer $i$is represented as $O_{i}$given that $I_{i}$ is the input of hidden layer$i$:

Here$W^{i}$indicates the weight matrix of hidden layer $i$, $B^{i}$ indicates hidden layer$i$ bias value,$\mathrm{Re}lu$ denotes activation function. A multi-input network design is used by the task offloading prediction model. There are two processing units in the network:(i) task offloading result classifier,(ii) service-delay regressor. Predicting a task offloading outcome is defined as binary categorization, hence, binary cross-entropy should be used as the classifier's loss function. Predicting task offloading service latency is the regressor’s challenge. MSE is regarded as the regressor's loss function. Theoutcome ofthe classifier is expressed in (26):

Here$I_{nh}$indicates final hidden layer activation result. Task offloading outcome prediction is a two-class sorting problem, and the classifier's activation function produces a task offloading success rate using the $Sigmoid$function. Finally, the suggested CNN-based task offloading strategy prediction in VEC maximizes the overall computation overhead under the weighted-sum of task completion time alongmonetary cost for scaling resources.

Deep Neural Network (DNN) ^[22] energy-efficient task offloading obtains consumer association schemes in vehicular edge computing schemes.The DNN input remains as large-scale disappearing modules and is represented in (27):

While the user’s data size is expressed in (28):

And the user latency requirement is expressed in (29):

where the parameters$\overline{h}$, $D$, and $T$ obtain the users’ association mode depending upon the nearest roadside unit’s association scheme. A one-step exploration is developed, which modifies one of the $K$user’s association schemes by maintaining other user associations. Every user accesses the other $M-1$RSUs in a one-step exploration, which is expressed in (30):

where$N_{one}$ indicates the one-step exploration, $K$denotes the association scheme created from random exploration, and$M$denotes the random selection of road side units.

The output layer acquires output values in the range ^[0,1] with the Sigmoid function. Subsequently, we acquire binary output values by choosing the road side unit along with a maximal output value for every user. The preprocessing data procedure is executed containing integration and normalization modes and is represented in Eq. (31):

where$x_{km}$ is obtained in variable units. Then, we normalize the input to measure input values $0$and $1$, which are computed using Eq. (32):

where the number of neurons on input and output layers are denotes$km$. Next is the training phase in which the classifier's activation function produces a task offloading process. Here, the task offloading scheme in VEC systems minimizes the vehicles’ total energy consumption and bit allocation. The nearby road side unit suggestion is determined; nevertheless, it avoids systematic solutions.

6. Comparison and Analysis

This section suggests that analysis of investigation papers depends on several criteria for multi-dimensional task offloading techniques in VEC networks,which are given in Table 1. Finally, the decision-making techniques in task offloading are compared in Table 2.

Fig. 2 shows the Summary of mathematical optimization task offloading algorithms.

7. Conclusion

In VEC networks, task offloading mechanisms minimize the costs of the system, which include communication cost and computation cost, when the task reaches the minimum allowable delay, and further constraints are gratified mainly by considering the characteristics of fast-moving vehicles. With some effort, this article presents a complete overview of existing work associatedwith task offloading in VEC. This article explored fog computing as well as task offloading procedures emulated through numerous task offloading factors leading the decision-making procedures. Here, numerous tasks offloaded with optimization and deep learning approaches are described. Furthermore, this article surveys deep learning strategies that are used for the task offloading process. Finally, this article proficiently deals with the identified gaps.