Task Allocation Methods and Optimization Techniques in Edge Computing: A Systematic Review of the Literature

Citation: Patsias, V.; Amanatidis, P.; Karampatzakis, D.; Lagkas, T.; Michalakopoulou, K.; Nikitas, A. Task Allocation Methods and Optimization Techniques in Edge Computing: A Systematic Review of the Literature. Future Internet 2023, 15, 254. https://doi.org/10.3390/ fi15080254 Academic Editor: Guan Gui Received: 8 July 2023 Revised: 20 July 2023 Accepted: 26 July 2023 Published: 28 July 2023 Copyright: © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). future internet Review Task Allocation Methods and Optimization Techniques in Edge Computing: A Systematic Review of the Literature Vasilios Patsias 1, Petros Amanatidis 1, Dimitris Karampatzakis 1,* , Thomas Lagkas 1, Kalliopi Michalakopoulou 2and Alexandros Nikitas 2 1Department of Computer Science, International Hellenic University, 65404 Kavala, Greece; [email protected] (V.P.); [email protected] (P.A.); [email protected] (T.L.) 2 Department of Logistics, Marketing, Hospitality and Analytics, Huddersfield Business School, University of Huddersfield, Huddersfield HD1 3DH, UK; [email protected] (K.M.); [email protected] (A.N.) *Correspondence: [email protected] Abstract: Task allocation in edge computing refers to the process of distributing tasks among the various nodes in an edge computing network. The main challenges in task allocation include determining the optimal location for each task based on the requirements such as processing power, storage, and network bandwidth, and adapting to the dynamic nature of the network. Different approaches for task allocation include centralized, decentralized, hybrid, and machine learning algorithms. Each approach has its strengths and weaknesses and the choice of approach will depend on the specific requirements of the application. In more detail, the selection of the most optimal task allocation methods depends on the edge computing architecture and configuration type, like mobile edge computing (MEC), cloud-edge, fog computing, peer-to-peer edge computing, etc. Thus, task allocation in edge computing is a complex, diverse, and challenging problem that requires a balance of trade-offs between multiple conflicting objectives such as energy efficiency, data privacy, security, latency, and quality of service (QoS). Recently, an increased number of research studies have emerged regarding the performance evaluation and optimization of task allocation on edge devices. While several survey articles have described the current state-of-the-art task allocation methods, this work focuses on comparing and contrasting different task allocation methods, optimization algorithms, as well as the network types that are most frequently used in edge computing systems. Keywords: task offloading; edge computing; task allocation; optimization algorithms 1. Introduction Edge computing refers to a type of computing that is decentralized and allows for data processing to occur closer to the data source, rather than relying on centralized data centers [ 1 ]. This approach is particularly advantageous for applications that demand a high bandwidth and low latency, such as autonomous vehicles, Internet of Things systems, and augmented reality applications at the edge of the network [ 2 ]. In edge computing, the distribution of tasks among the various nodes in a network is referred to as task allocation. Choosing the best place for each task is one of the primary difficulties of task allocation in edge computing. This is because the different tasks demanding varying amounts of computing power, storage, and network bandwidth. A work that needs realtime processing, for instance, could need to be assigned to a node with high computing power, while a task that needs a lot of storage would need to be assigned to a node with a lot of storage space. Additionally, tasks that are sensitive to network latency may need to be allocated to nodes that are close to the source of the data. This necessitates a trade-off between the node’s computing power, energy usage, and network delay. In Figure 1, we can see the architecture of and edge–fog–cloud computing system. Future Internet 2023,15, 254. https://doi.org/10.3390/fi15080254 https://www.mdpi.com/journal/futureinternet Future Internet 2023,15, 254 2 of 30 Figure 1. Edge, fog, and cloud computing architecture. Furthermore, task allocation in edge computing is dealing with the dynamic nature of the network. Edge computing networks are often composed of a large number of devices that can join or leave the network at any time. This means that the number of available resources and the location of those resources can change frequently. Task allocation algorithms must thus be able to adapt to these changes and decide in real time. Furthermore, several competing goals may need to be taken into account, including energy efficiency, data privacy and security, and QoS assurances. In edge computing, there are several methods for allocating tasks, each with its advantages and disadvantages. Using a central algorithm that makes decisions based on the network’s present state is a common strategy. This strategy calls for a central controller to keep an eye on the network and decide how to distribute the computational tasks. This strategy, however, may be susceptible to network outages and result in a decision-making bottleneck. Also, it may not be feasible to have a central controller in large-scale edge computing networks. Additionally, the use of a decentralized algorithm allows each node in the network to make decisions about task allocation independently. This strategy is predicated on the notion that every node has a local view of the network and can take decisions based on that perspective. This approach can be more robust to network failures and can lead to faster decision making. However, it may lead to sub-optimal solutions due to lack of global information [3–9]. The algorithms that adopt both centralized and decentralized techniques are called hybrid. These algorithms are designed to take advantage of the strengths of both centralized and decentralized approaches. For example, a hybrid algorithm might use a centralized algorithm to make high-level decisions about task allocation, while allowing each node to make low-level decisions based on its local view of the network. This approach can produce more effective solutions by balancing the trade-offs between centralized and decentralized methods. Utilizing machine learning methods to enhance work allocation is another current strategy. Algorithms that use machine learning can learn from past data and forecast how the network will behave in the future [ 10 ]. As a result, the algorithm may be able to allocate tasks more precisely. The network’s energy usage may also be optimized using machine learning methods. For instance, a machine learning algorithm may be taught to predict how much energy various nodes will need and make decisions about task allocation based on that prediction. This work examines the current methods for task allocation on edge devices. We evaluate the key techniques discussed in the literature to see how effective each of them is for edge computing and which would be most helpful for future research. We examine and evaluate the corpus of contemporary research on workload distribution on edge intelligence devices. Our work provides a thorough assessment of the main methods and algorithms used in edge computing for efficiently executing difficult workloads [11–79]. Future Internet 2023,15, 254 3 of 30 The main contribution of this paper is that it provides a review of the state-of-theart task allocation methods. More precisely, this paper only used research published in the last three years and it highlights the most frequently used task allocation methods and optimization algorithms in combination with the network types which are used in edge computing systems according to the application and the task allocation algorithm used. The purpose of this work was to show the task allocation and all the optimization algorithms and methods that have been, very recently, the most frequently used in order to advance the current understanding of choice-making when it comes to adopting the optimal task allocation method, algorithm, and networks according to the use case. All the task allocations methods reviewed in this paper tried to give a solution to the edge servers or the devices located in the edge computing layer by distributing their workloads to the end devices, fog layer devices, and cloud computing layer devices in order to provide optimizations according to the use cases of the applications. The document is divided into six sections. The systematic literature review (SLR) approach employed in this article is described in Section 2, including the definition of research questions (RQs). Next, in Section 3, we provide the list of findings in the literature. In Section 4, we provide a summary of the collected papers. In Section 5, we provide a critical discussion of our findings and identify future work challenges. We finally conclude this review in Section 6. 2. Methodology and Research Questions Conducting a thorough assessment of the existing literature is commonly achieved through a systematic literature review (SLR), which is recognized as one of the most widely used approaches [ 80 – 83 ]. This type of study requires a clearly defined process to accurately identify relevant research. A review should follow a specific protocol to effectively collect and assess previous works. The methodology which was followed in this paper incorporates four steps: first, recognize the research questions; second, explain the literature sources and the search string; third, select applicable studies; and fourth, assess the studies gathered and extract the necessary data before combining them. This review aims to identify the task allocation approaches on edge devices. Thus, the RQs defined are the following: •RQ1: What are the contemporary task allocation techniques in edge computing? •RQ2: What are the most frequently used task allocation optimization algorithms used in edge computing? •RQ3: What are the most used computer networks in edge computing? The following method was employed to create a search string clearly corresponding to our RQs. The key search terms first and foremost need to be RQs-specific. Then, further search terms are obtained utilizing previously examined documents. The next stage is to search for alternative spellings and synonyms for the main keywords. The final step is to combine the keywords to create the search string. 3. Research Process To better understand task allocation in edge computing, in the following review, we will examine the most recent papers that refer to the specific topic. To achieve this, we conducted an extensive search in the Scopus search engine [ 84 ] for papers relevant to “task allocation on edge computing”. Therefore, search terms such as “task allocation” and “edge computing” are used to search in the titles, abstracts, and keywords of publications. We narrowed the search to the years after 2020 to catch out on the latest technologies in the field we are examining. The initial search yielded back 218 documents. Consequently, using the following inclusion/exclusion criteria: • In cases where the study has a conference version and a journal version of a study, the journal version is retained while the conference version is discarded. Future Internet 2023,15, 254 4 of 30 • If there are multiple published versions of a study, only the latest or most recent version is retained. • If a study is available in multiple sources, only one version is included. Exclusion criteria were also used to omit research that did not meet the requirements for inclusion. The following are the specific exclusion criteria: • Papers not available on an open access basis or through our university access systems that are opening the vast majority of the reputable editor houses and journal titles were not included in the final literature corpus. • Only documents written in the English language were used. • Only papers related to computer science were included; works that were only peripheral to edge computing were excluded. • Papers dealing with very specific datasets were excluded. We focused on conceptual papers, case studies, methodological papers, and position papers. We do acknowledge that potentially missing a relatively minor number of papers that we excluded from our systematic review approach due to the unavailability of access to the sources, these were published, and may have a minimal impact on the completeness of the corpus and thus on our results. We would argue, however, that our Universities’ journal access agreements in Greece and the UK are very inclusive and only miss very specialized and small publishers. We anticipate that this had an extremely limited impact on our work. In this research, we used the following keywords (with logical connectives) and their combinations: edge computing, task allocation, IoT, energy utilization, mobile edge computing, computation offloading, integer programming, resource allocation, resource management, antennas, reinforcement learning, task analysis, fog computing, energy efficiency, green computing, task offloading, UAV, 5G mobile communication systems, cloud computing, data handling, job analysis, MEC, network architecture, cloud-computing, computing environments, deep learning, edge server, fog, learning algorithms, big data, computationintensive task, computational modeling, computing resource, energy consumption, learning systems, multi-agent systems, optimization, QoS, servers, task allocation algorithm, and total energy consumption. To answer RQ1, the task allocation methods are categorized according to the following methods: 1. Resource-aware task allocation: Allocating tasks depending on the resources that are currently accessible on edge devices, such as processing power, memory, and battery life. 2. Distributed task allocation: Allocating tasks across a network of edge devices to optimize the performance and reduce latency. 3. Dynamic task allocation: Adapting task allocation in real-time based on changes in device performance or network conditions. 4. Machine learning-based task allocation: Using machine learning algorithms to predict resource utilization and allocate tasks accordingly. 5. Energy-efficient task allocation: Allocating tasks in a way that minimizes energy consumption in edge devices. 6. Quality of service (QoS)-aware task allocation: Allocating tasks based on the required QoS levels for different tasks. 7. Collaborative task allocation: Allocating tasks across a network of edge devices by taking into account the collaboration between devices. 8. Context-aware task allocation: Allocating tasks based on the context of the edge devices, such as location, available resources, and user preferences. Table 1and Figure 2present a comparative analysis of those papers based on the classification and contextualization of their key results. Note that the columns in Table 1 are directly related to the answers addressing our RQs. Future Internet 2023,15, 254 5 of 30 Table 1. Summary of task allocation methods in edge computing. Proposed Task Allocation Optimization (RQ2) Applied Network Type (RQ3) Distributed task allocation [18] FL WHAB [22] OFB VANETs [24] GT, RL MEC [29] MVAA MEC [45] PSO IoT [50] GT IoT [55] Hybrid MEC [59] CW Blockchain [61] BSUB MEC [63] FL MEC [31] DLA IoT [33] RL EC [34] CFG C-RAN [35] AsP IoT [36] GSO MEC, Fog [65] AsP NDN-IoT [66] AB, GA Wireless, cellular, satellite [67] Heuristic MEC Collaborative task allocation [12] PFA EC [77] MINLP EC [17] ACO IoT [27] Heuristic IoT Context-aware task allocation [52] DFD DAG [69] GT IoT, 5G [78] LPA MEC Energy-efficient task allocation [68] MAPE-K IoT [11] Lyapunov MEC [19] Bi-Level MEC [25] DRL IoV [30] QT 5G [43] PSO IoT [47] ILP MEC [71] JTORA MEC [73] hybrid RF-FSO Industrial IoT Dynamic task allocation [13] MPSO VVECNs, VANETs [79] MH IoV [20] AA Fog [23] DP MEC [41] DRL IoT [44] ECTA EC [46] AA Fog [49] ILP [56] GA, GEN IoT [60] PSO EC [64] BPSO Future Internet 2023,15, 254 6 of 30 Table 1. Cont. Proposed Task Allocation Optimization (RQ2) Applied Network type (RQ3) Machine learning-based task allocation [28] DQN-D EC [54] IoT [38] MARL IoT [42] ACO IoT, 5G [51] FLOM-Opt [53] Q-Learning IoT, IoV [57] MA IoT [58] Heuristic, RL EC Quality of service task allocation [26] EC [32] MMAS EC [48] QT EC, IoT Resource-aware task allocation [76] MAPPO MEC [62] MDP EC [14] MAP VEC [75] JTORA VEC [72] JTORA MEC [74] JTORA MEC [15] DNF [16] MARL MEC [21] JTORA NOMA-MEC [37] EC [39] ELB [40] Knapsack MCS [70] EC, 5G The task allocation methods (RQ1) covered by the papers we examined are shown in Figure 3. Most of the papers use distributed task allocation (26.1%), followed by resourceaware (18.8%) task allocation. Dynamic task allocation (15.9%), energy-efficient (13.2%), and machine learning (11.6%) task allocation followed next. Collaborative (5.8%), quality-ofservice (4.3%), and context-aware (4.3%) task allocation are the methods that are mentioned less. It is also very important to note that, in many papers more than one task allocation methods was used and applied. Regarding the task allocation optimization algorithms (RQ2), as shown in Figure 4, the most frequently used is the swarm optimization in various forms (10.1%), various methods of reinforcement learning (10.1%). JTORA (7.2%), multi-agent methods (5.8%), the game theory (4.3%), the heuristic approach (4.3%), and the auction algorithm (4.3%). We also identified the use of many other task allocation optimization algorithms that were unique; these cases refer to 53.8% of the sample examined. We should note again that many authors in their works implemented more than one optimization algorithm to illustrate the effectiveness of each optimization algorithm. The communication network types (RQ3) are illustrated in Figure 5. Most papers applied IoT networks (27.6%), followed by mobile edge computing (MEC) networks (26.1%), EC networks (18.8%), vehicular networks (10.1%), and 5G networks (5.8%). The rest are illustrated as one category ’various kinds of networks’ (11.6%). Future Internet 2023,15, 254 7 of 30 Figure 2. The classification mind map from the analysis of papers in task allocation methods in edge computing. Future Internet 2023,15, 254 8 of 30 Figure 3. Percentages of the task allocation methods used. Figure 4. Percentages of allocation optimization methods used. Figure 5. Percentages of communication networks used in task allocation. Future Internet 2023,15, 254 9 of 30 4. Literature Review and Data Extraction This section presents a classification of the task allocation methods used in the recent literature. Each task allocation method in combination with the computer network type and the deep learning compression methods has its own strengths and it is used to address different objectives such as minimized computational power, energy consumption, and network delay. 4.1. Distributed Task Allocation Applying edge computing in high-altitude balloons (HABs), as per Sihua Wang et al. [18] , take into account a HAB network that supports MEC and allows users to periodically request computing tasks with different data volumes. The HABs must dynamically decide on the best job allocation, service sequence, and user association to provide their users with computing services. The objective of this combination of job allocation, service sequence, and user association was to reduce the weighted total of the time and energy used by all users. The authors propose a federated learning (FL) approach that uses support vector machines (SVMs) to predict user association in advance, addressing the optimization challenge in task allocation. By using this approach, the heterogeneous autonomous base stations (HABSs) can collaboratively train an SVM model that can accurately predict the best user association without needing to transmit large amounts of task data or previous user association results. The original, nonconvex issue is split into two sub-optimization problems—the task allocation optimization problem and the service sequence optimization problem—which are then addressed repeatedly based on the expected user association. Specifically, the authors developed a closed-form formula for the ideal service sequence given the job allocation vector. This optimization problem may be converted into a piecewise linear problem that can be addressed using linear programming if the best service sequence is known. Previous research showed [ 22 ] the occurrence of a multi-device game that offloads processing for several users. Computing efficiency may be significantly increased by using the various available intelligent devices that industrial vehicles can assign calculation assignments to perform the operation in parallel. Moreover, to determine the optimal offloading strategy, they formulated a sequential game of multi-user computation offloading, treating multiple industrial vehicles as multi-devices for handling multiple targets. They took into account that the system cost is primarily made up from the price of rental IDs, energy consumption, and execution time. To improve the chances of establishing wireless connections between industrial vehicles (IVs) and unmanned aerial vehicles (UAVs), a dynamic scheduling technique based on the density of small partitions was introduced. The current vehicle density in the partition determines the UAV’s residence time, which was scheduled by the software-defined network controller. They developed an algorithm for minimum incremental task allocation (MITA), a job distribution mechanism for industrial vehicles to distribute some of the computational duties among several IDs. MITA was determined as the best multi-target task allocation strategy and reduces the system cost associated with task execution. In order to reduce the user response time, balancing the load, and maximizing server resource usage, Zhenjiang Zhang et al. [ 24 ] developed a task offloading technique based on a multi-agent approach. Additionally, the authors addressed different problems that centralized scheduling would create and also presented the benefits of centralized training and dispersed operation. The authors used an experimental approach to confirm the viability and efficacy of their distributed task offloading algorithm based on multi-agent and load balancing (DTOMALB) method and the analysis of the simulation results. This paper [ 29 ] addressed the problem in MEC of indivisible task allocation with heterogeneous resources (TAHRC). In a scenario with diverse resources, the authors provided an accurate mathematical method for indivisible task allocation. Furthermore, an effective task allocation that takes into account various types of resources, as well as binary computation offloading, was presented. The authors defined TAHRC as an integer Future Internet 2023,15, 254 16 of 30 4.5. Dynamic Task Allocation Literature [ 13 ] suggested the idea of a “resource pool” made up of service vehicles that move together. They suggest a cooperative task scheduling strategy to reduce the task execution time, based on these idle resources that may be allocated in resource pools. They define the min–max issue with the allowable latency constraint for the job execution time optimization model. Additionally, they formulated problems that take node mobility into account. To find the optimum work allocation strategy and reduce the task execution time for the min–max issue, the authors adapted the max–min fairness algorithm and PSO algorithm, respectively, depending on whether to summon all service cars. The simulation results showed the success of the suggested strategies. Another effort made by Ammar Awad Mutlag et al. [ 20 ] applied edge computing in the healthcare system. The authors presented a critical healthcare task management (CHTM) paradigm for monitoring electrocardiograms (ECGs), in a fog-cloud computing topology. In more detail, this paper showed how fog computing might improve service delivery and network congestion, as well as the limits of cloud computing in terms of real-time healthcare applications. Fog computing does not have a distributed design, and its nodes are diverse and exclusive in how they share resources and space. In order to manage crucial activities, the suggested CHTM model intends to offer scheduling, resource sharing, interoperability, and dynamic task allocation. The model suggested a multi-agent system to completely control the network from the edge to the cloud, as well as a resource scheduling model for fog nodes. The outcomes of the simulation demonstrated how the suggested approach lowers the network utilization, reaction time, network latency, energy consumption, and instance cost. Finally, the authors discussed potential future study options as well as the comparisons of the suggested model with relevant work. Overall, the CHTM model has met its performance requirements and offers an effective resource scheduling plan for crucial healthcare activities between the fog node layer, edge layer, and cloud. Tristan Braud et al. [ 23 ] provided a model for a multi-server task allocation method to optimize allocation over various connections. A solution for mobile augmented reality (MAR) and the ways in which mobile devices are limited by the processing and latency requirements of MAR apps was provided. The authors illustrated the durability of the system in situations of network instability using simulations and real-world trials. The authors also go over 802.11ax and the forthcoming 5G technologies. A task dependency graph, optimization methods for related tasks, and a scheduling algorithm for work distribution over several wireless links to D2D, cloud, and edge servers were presented. The authors’ contributions included conducting multiple simulations and implementing a real-world multipath, multi-server mapping application. A recently published paper by Yan Chen et al. [ 41 ] addressed the issue of dynamic task allocation and service migration (DTASM) in edge-cloud IoT systems in order to improve the task allocation strategy and reduce the load transferred to the cloud server while fulfilling migration, latency, and computing capacity restrictions. The work suggested a deep reinforcement learning (DRL)-based solution. A cloud server was used to administer the whole system, together with a number of RAN nodes and IoT users running various sorts of apps. Extensive simulations were used to test the suggested technique, and the results demonstrate that it outperforms other benchmarks’ task allocation strategies. Investigating the DTASM problem in a heterogeneous edge-cloud IoT system that needs dynamic task allocation and offering a DRL-based solution to the problem of a large discrete action space are two contributions made by this study. Overall, the study examined the service latency restriction and the smooth service migration requirement, making it appropriate for edge-cloud systems with several IoT users. Moreover, Ping-Chun Huang et al. [ 44 ] addressed the problems of how edge computing can become a viable tool for reducing the communication lag. In order to balance the workload across edge servers and reduce the transmission distance, the authors suggested a load-balancing model that takes the computational load on edge servers as well Future Internet 2023,15, 254 17 of 30 as transmission distance into account. Edge server placement (ESP) in conjunction with the proposed algorithm performed better than conventional heuristics. Furthermore, the proposed method was more successful in integrating work allocation and server placement, according to their simulation findings. Bartosz Kopras et al. [ 46 ] analyzed task distribution in latency-constrained fog computing networks, where fog nodes (FNs) are closer to the network edge and have lesser computational capabilities than the cloud resources. An optimization issue to reduce task-related energy for transmission and processing with delay limitations was presented in this study. The optimization takes into account both the workload distribution across the nodes and the CPU frequency at each active node. Two useful algorithms termed energy-efficient resource allocation (EEFFRA) and low-complexity energy-efficient resource allocation (LC-EEFFRA) were developed after the issue was converted with successive convex approximation and decomposed using the primal and dual decomposition techniques. The utilization of EEFFRA/LC-EEFFRA greatly decreased the count of computational requests that did not meet the expected delay requirements. Additionally, the authors made use of dynamic voltage and frequency scaling (DVFS) to reduce energy usage while still meeting delay requirements. Modeling the energy consumption and delays associated with transmission and processing in the fog and cloud layers, as well as putting forth and resolving a challenging optimization problem, were the paper’s key contributions. A recent paper by Mingjin Zhang et al. [ 49 ] proposed an entirely novel edge-native task scheduling system (ENTS) that co-schedules networking and computing resources to improve the performance of edge-native applications. Although Kubernetes is now the de facto standard for container orchestration, it does not enable edge-native apps or take into consideration their unique throughput and latency needs. This paper includes the case study of a video analytics application to show how task distribution and data flow scheduling enable ENTS to maximize work throughput. To tackle the mixed-integer nonlinear problem, this paper offers two online methods. The research utilized metrics like task throughput and the average waiting time to compare the performance of ENTS to different baseline techniques on a real-world testbed. The findings demonstrated the necessity for specialized scheduling systems for strengthening edge-native applications by demonstrating considerable gains in work throughput and latency. This study also offered suggestions for potential future advancements, including the creation of more sophisticated algorithms for group job scheduling and the incorporation of software-defined networking into the network controller. Furthermore, Shida Lu et al. [ 56 ] suggested a network architecture of cloud edge fusion for the MEC system to enable thorough query and computing for real-time applications. By utilizing MEC’s short distance from the user equipment (UE), difficult computational tasks can be executed by UE and useless data can be removed before flowing to the cloud. A suggested distributed task scheduling technique was based on the network architecture to deal with complicated computational tasks; several MEC servers cooperate, work in parallel, and eventually accelerate the task execution response times. To demonstrate that this approach performs better, the authors ran many simulation trials. A task allocation optimization strategy for distributed edge networks to reduce energy usage while achieving a high quality-of-service was published by Philippe Buschmann et al. [ 60 ]. ILP, PSO, and DRL are the three optimization techniques that were compared in this paper. PSO outperforms ILP in smaller problems, but DRL is better suited to larger problems and has the lowest upper bound for the optimality gap of the three techniques. According to the study’s findings, the performance of the three techniques varies depending on the size of the job allocation problem in the edge networks. The PSO is more suited for smaller issue sizes, whereas DRL is better suited for bigger ones. According to the findings, the extended PSO algorithm is unreliable and useless for task allocations of greater than 20 and 60, respectively. The study highlighted the potential future research areas, such as investigating the trade-offs between speed and optimality in ILP problems that prioritize previously assigned tasks, and exploring the field of heterogeneous networks. Future Internet 2023,15, 254 18 of 30 Furthermore, Junling Yu [ 64 ] suggests a binary particle swarm optimization (BPSO)- based MEC application in the classroom assessment system for English education. It addresses the NP-hard problem of decoupling the channel resource allocation problem, proposes a multi-user and multi-MEC scenario based on mobile edge computing to maximize total revenue to finish the work, and formulates an objective function to reduce the task execution cost. This study uses big data allocation research from mobile edge computing resources for the application of the English teaching classroom assessment, accelerating the growth of classroom instruction. It offers recommendations on how to enhance and put into practice the English education assurance system. 4.6. Machine-Learning-Based Task Allocation A method for online scheduling optimization for requests based on DAGs in edge computing networks was proposed by Yaqiang Zhang et al. [ 28 ]. The process was characterized as MDP, where the best task allocation strategy was learned at each decision step via temporal-difference learning. Temporal restrictions between concurrent requests were satisfied while the system’s long-term latency and energy use were reduced. In comparison to state-of-the-art approaches, the suggested mechanism exhibits promising results in lowering long-term delay and energy usage. The offloading of complicated structured tasks and pertinent strategies for task scheduling in edge networks are also reviewed and discussed in this paper. The suggested approach attempts to ensure the quality of experience (QoE), particularly when requests are latency-sensitive and include a lot of network traffic. A work published by Zeina Houmani et al. [ 54 ] addressed the issues with managing deep learning applications over an edge-cloud architecture where data are evaluated and transported between resources at the network’s edge. Trade-offs between accuracy and latency are important for DL applications since they need outcomes close to real-time with a precision that the user has set as acceptable. When trade-off improvements are possible, the research recommended a data-driven scheduling technique and a data management strategy that lowers the resolution of incoming data. This architecture is for time-critical DL operations. The pipeline was deployed on distributed resources using the suggested system, which also has the ability to monitor each job separately to maintain the system performance. To effectively manage the full DL pipeline in practical deployments while resolving tradeoffs between quality of service (QoS) indicators, the paper emphasizes the need for resource and data management solutions. It also highlights related developments in edge computing, microservices, and edge-enhanced data analytic systems. The architecture of the system has three levels: data management, infrastructure, and workflow management. The developer’s requirements were met by the system’s timeliness and accuracy thanks to the data management level’s selection of the data quality distribution for data sources. The study defines the system’s objective and analytical models to evaluate the accuracy and end-to-end latency of K data sources. The study looked at object detection in a multi-user setting on Grid’5000 and found that it increased the average system makespan by up to 54.4% compared to a cloud-only setup. To improve this, future research should develop a resource allocation tool for deep learning workflows that considers load, tasks, and resource requirements to meet performance constraints. Another work [ 38 ] studied the topic of load balancing as well as task allocation in computational infrastructures for smart cities that include edge and cloud computing. The authors provided an RL method for task distribution that takes into account the differences between nodes and tasks, as well as their interdependencies and time delays. The authors showed a simulator to evaluate their method with existing allocation methods and introduce an abstract description of the computing infrastructure for smart cities. The suggested method handles dynamic changes well and evenly distributes the workload across computing components. The contribution of this study was to find ways to optimize resource usage for smart city applications. Future Internet 2023,15, 254 19 of 30 A new tool to guide the task controller in making decisions (i.e., the allocation process) was provided by Madalena Soula et al. [ 42 ]. Such decisions are dynamically influenced by the effectiveness of the edge computing nodes and the updated datasets. The authors assumed that TCs assign various jobs to various edge computing nodes simultaneously as they process the incoming tasks in batches. A clustering-based allocation method and a bio-inspired model (i.e., a modified version of the well-known PSO technique) were both proposed as alternative allocation techniques. The goal of both models was to identify the best allocations at a given moment in time. They relied on procedures that did not require any training. In order to show the cost for each allocation based on the demands of the jobs and the condition of the edge computing nodes, they also considered the cost of allocation as stated in their past work [ 85 ]. The cost of allocation is also used as an indication for calculating the ranks of edge computing nodes, resulting in a “rewarding” mechanism for resolving the issue at hand. A method for handling large-scale tasks in edge computing was presented by [ 51 ] using the federated learning-based optimization methodology (FLOM). To accomplish accurate task categorization, global load balancing, and lower task processing costs, FLOM combines FL and deep feature learning approaches. The suggested method incorporates an FL architecture that permits data exchange and model parameter updates across several edge computing and cloud centers without disclosing sensitive information to outsiders. To learn the deep features of task requests and hosts in the substrate network, the authors provide a deep network model. The experimental findings show that the FLOM technique outperforms other approaches and efficiently handles large-scale task categorization and allocation. In edge computing environments, the research emphasizes the necessity for effective task allocation methods that can identify independently jobs while achieving system load balancing. The suggested method has several potential applications in industries including intelligent manufacturing, intelligent IoT, and smart cities. Previous research [ 53 ] tried to apply edge computing in smart cities in order to implement a task distribution among the various nodes. The authors researched a variety of allocation techniques, from stochastic allocation to intuitive methods to integer programming-based mathematical optimization. Furthermore, a demonstration of how efficient job distribution in HE µ Cs offered appreciable gains in total performance. The authors also showed that it is completely possible to execute integer programming at the periphery with little overhead (less than 2% of the total makespan time in their tests in the worst case scenario) for the size of the scheduling they performed. Their research contribution was that they defined and provided a motivation for the idea of heterogeneous edge micro-clusters (HE µ Cs) and also that they showed the examples of common HE µ C compute platform tasks. This research demonstrated the effectiveness of using mathematical optimization as a method for batch job distribution in HEµCs. Ishihara et al. [ 57 ] proposed a task allocation method where all agents involved in executing tasks can express their preference for which tasks they want to work on, and the manager agent then allocates tasks based on these preferences and system performance requirements, without employing any extra knowledge of the agents’ capabilities or state. The authors also took into account the manager’s particularity, which means that each manager has a different set of customer requests based on the place and the time. Therefore, by balancing their workloads, agents must be able to dynamically select which manager to support with the specified tasks. The author suggests that, when making decisions about service choice, justice and societal considerations should be given priority over the concept of centralized control. This is because centralized control is not suitable for existing applications, where services are provided by multiple independent organizations. Lastly, Liu and Liu [ 58 ] proposed an approximate technique for task allocation that uses the least amount of overall energy while taking into account multi-task parameters in MEC. Taking into account the binary computation offloading mode and restricted frequency subchannels, the authors simulated the multi-task allocation problem, resulting in an integer programming problem that is severely NP-hard. It is the first of its type, and Future Internet 2023,15, 254 20 of 30 trials demonstrated that the proposed PTAS algorithm for multi-task allocation in MEC can locate nearly optimum solutions while striking a decent compromise between speed and quality. The study recognized the need to develop an effective solution to address the multi-task allocation problem. 4.7. QoS Task Allocation A recent article [ 26 ], provided a summary of the techniques, advantages, and limitations of the G-PATA approach. This methodology is unique in that it simultaneously optimizes conflicting objectives, namely the reduction in energy consumption at the rational edge and the minimization of processing time at the server, while also fulfilling specific privacy requirements. The authors discussed the challenge of task allocation in edge computing systems that take into account privacy concerns for delay-sensitive applications related to social sensing. They also formally outlined the problem’s goals after presenting the task model, underlying assumptions, and privacy model they utilized. Furthermore, Zhou Li et al. [ 32 ] presented edge computing into mobile crowd sensing (MCS) and suggested a three-tier architecture, in which they presented ideas for task distribution and user data input. The edge servers transmitted the job requirements to users during task allocation so that exact personal information is no longer required. Edge servers enable participants to contribute high-quality altered data without compromising their privacy. The authors outline the challenge of sensing cost reduction while maintaining the privacy and provide the data sensing mechanism with user privacy preserved (DS-UPP) method as a solution. In DS-UPP, they created a compressive sensing-based method to reduce the quantity of required sensing data and an LDP-based algorithm to safeguard the privacy of the participants. They also examined the DS-UPP that satisfies e-differential privacy. Taking into account the limitations posed by the privacy budget and task recovery error, the authors established both the minimum and maximum number of participants required to mathematically complete the task, along with the expected amount of data that each participant should contribute. Additionally, through the use of a simulator, they conducted in-depth simulations to assess the effectiveness of the DS-UPP. After applying the PrivKV algorithm for comparison, according to the experimental findings, DS-UPP may, on average, cut the cost of sensing by over 90% while still maintaining privacy and data quality standards. A method to lower the overall latency on an MEC platform by providing a work allocation mechanism was proposed by Katayama et al. [ 48 ]. Three different server types were available on this platform: a dedicated cloud server, a MEC server, and a shared MEC server. The authors initially calculated the processing time and transmission delay for different types of servers to determine the time taken for a task to be submitted and a response to be received. They used queuing theory to determine the transmission latency for the shared MEC server, with the bottleneck node being modeled as an M/M/1 queueing model. To minimize the overall latency for all tasks, they formulated an optimization problem for task allocation. Tasks may be properly distributed across the MEC servers and cloud servers by resolving this optimization challenge. Even with the use of a metaheuristic approach like the genetic algorithm, the computation time is still exceedingly long. As a result, the authors also suggested a heuristic technique to find the roughly ideal answer more quickly. A core algorithm and three supporting algorithms make up this heuristic algorithm’s four components. Tasks are split into two groups in this method, and task distribution is carried out for each group. The examination of the performance of their suggested heuristic algorithm in comparison to the results provided by the genetic algorithm and other techniques was provided. The task allocation methods are relatively simple and can be easily implemented on an MEC platform. However, there is a risk of falling into a local minimum when using the suggested heuristic approach. To address this issue, the authors utilized the random search approach (ARSET) and heuristic random optimization (HRO) to avoid getting trapped in local minimums. Future Internet 2023,15, 254 21 of 30 4.8. Resource-Aware Task Allocation The authors in [ 14 ] employed their proposed model to simulate a practical vehicular environment and investigate the problem in a large-scale network. A small part of the source task was processed at the vehicle, and the remaining task was offloaded at other vehicles and then at a vehicular edge computing server level. This lowers the cost of the entire system while also allowing for the utilization of ample vehicle resources and a reduction in the strain on the overworked VEC server. The MAP algorithm was used in the VEC scenario which enabled each vehicle to choose its adjacent cars depending on the finest resources available at the lowest cost. Additionally, they estimated the transmission rates for V2I and V2V communication while taking into account realistic assumptions. The determination of whether a job should be computed locally, on a nearby vehicle, or at the vehicle edge computer (VEC) was based on the percentage of the job and was conditional on two factors: the maximum allowable delay and the vehicle’s stay time. Finally, by contrasting it with other methods, they could assess how various variables and vehicular contexts affect their MAP task offloading approach. In another paper published by Chen et al. [ 15 ], the authors suggest a cooperative learning process that makes use of simulated and real-time captured data to decrease the quantity of data required to produce a trustworthy data-driven model. The authors employed the DCTA technique to distribute tasks in a data-driven system. Through both a trace-driven simulation and a brand-new, in-depth real-world AIOps case study that connects theory and practice with a novel architecture, a key component was designed within an AIOps system. The authors also assessed numerous distinct work allocation methodologies. At the end, the DCTA method significantly reduces the processing time by 3.24 times and saves 48.4% of the energy consumption compared to state-of-the-art approaches when addressing task importance for multi-task learning (MTL) using task allocation in time-constrained and integrated management (TATIM). An optimization problem that aims to minimize the total number of users and transmission delays using a multistack RL method was presented by Wang et al. [ 16 ]. The results of this study showed that each base station (BS) opted to allocate a greater number of downlink subcarriers and transmit power in the downlink direction to a user whose task needed to be processed by the MEC server, to reduce the maximum delay among all users. The preference of each BS was to give more uplink subcarriers and a greater uplink transmit power to a user who needs to locally complete a job. The authors presented a novel strategy for load balancing in edge computing by allocating tasks through intermediary nodes. Xue and An [ 21 ] analyzed a network situation with several mobile edge server nodes (MSNs) and a number of edge devices (EDs), each of which had an MEC server to offer wireless and computation resources. Their study concentrated on ED-specific tasks, which could be broken into any number of smaller portions for both local and remote computing needs. Data communication during task loading was achieved using NOMA technology to increase the resource usage and MEC performance. The authors took into account the limitations of wireless and communication resources, and they formulated a combined optimization problem of task loading and resource allocation to increase the system’s capacity for job processing. To address the defined MINLP problem with the characteristics of the objective function, the authors divided the original problem into two subproblems: the resource allocation (RA) problem and the task allocation (TA) problem. Then, they further divided the RA problem by allocating them as either computation and communication resources. Initially, the authors assumed that the power distribution of subchannels was equal when allocating communication resources. The researchers approached the subchannel allocation problem in the context of a two-sided matching process between mobile terminals (MTs) and subchannels. They proposed a sub-optimal algorithm with low complexity to allocate subchannels, and then treated the transmission power allocation as a convex optimization problem, using the Lagrange multiplier approach. A task allocation method based on resource allocation was used to overcome the task allocation issue. Future Internet 2023,15, 254 22 of 30 Computer simulation results demonstrated that the suggested task offloading and resource allocation technique enhanced the MEC system’s overall performance. Recently, Canete et al. [ 37 ] proposed a set of four modules that work together to support developers in adjusting the deployment process based on the specific requirements of an application and the capabilities of the infrastructure. These modules are designed to function independently while also complementing one another. The work first defined the FMs feature models (FMs) of the infrastructure and application. An FM depicted, based on characteristics, the relationship between the common and variable parts of a product family (either an application family or a system family). FMs were represented as a collection of features that were arranged hierarchically, with parent–child connections between the features and a set of constraints (known as cross-tree constraints) that show the links between the features. The infrastructure consisted of a range of different devices, each with various hardware and software attributes, such as the device type, computational capacity, peripherals, network capabilities, operating system, third-party libraries, and so on. These attributes were related to the types of services or tasks that could be performed on the execution platform, and they pertained to the hardware and software requirements of the applications. The use of feature modeling (FM) helped create the more accurate models of the infrastructure, which is often overlooked in the software product line (SPL) models. This approach allowed the mapping of application features from the application of FM to the software and hardware features of the infrastructure, even when different nodes had different capabilities. To simplify the configuration of the infrastructure’s feature model (FM) and manage the evolution of hardware and software characteristics separately, the authors divided the hardware and software features into two separate FMs. A load-balancing method for MEC servers was published by Chen et al. [ 39 ] which operates in ultra-dense networks based on load estimated. The suggested method accounts for the dynamic distribution of user equipment and ad hoc application activities, and it provides a novel idea of task unit load transfer overhead for load balancing that is both lowcomplexity and high-efficiency. In addition, the study investigated load estimations based on overlapping coverage and user load prediction, and it suggested a three-tiered computing network design made up of devices, edge nodes, and cloud servers that improves load balancing while taking energy use and latency into account. Simulated experiments were performed to verify the efficacy of the remedies that were suggested. Li et al. [ 40 ] integrated the edge computing with MCS and put up a three-tier architecture. The authors worked on creative task allocation and user data submission ideas. The edge servers transmit the job requirements to users during task allocation so that exact personal information is no longer required. Edge servers assist participants in submitting high-quality altered data that protect their privacy. They define the challenge of privacypreserving sensing cost reduction and provide a solution using DS-UPP. In DS-UPP, a compressive sensing-based method to reduce the quantity of required sensing data and an LDP-based algorithm to safeguard the privacy of the participants were created. Also, the authors examined the DS-UPP’s theoretical properties. It has been established that DS-UPP complies with the “differential privacy with the restrictions that privacy budget is” a recovery error. To solve this problem, the authors provided the mathematical lower bound and upper bound of the number of participants needed for the task achievement as well as the typical quantity of data that should be supplied by a participant. Lastly, they applied the algorithm PrivKV for comparison. According to their experimental findings, DS-UPP gained lower costs for sensing by roughly 90% while maintaining privacy and data quality standards. Moreover, Xuefeng et al. [ 62 ] addressed the drawbacks of the aforementioned current technologies. The paper created a platform for intelligent operation inspection that relies on a multi-agent system and is used for the live line measurement of substation equipment. The multi-agent system technology was used in the creation of the operational framework. The paper proposed integrating edge computing into the data processing stage of a multiagent system, which enables the intelligent operation inspection platform to overcome Future Internet 2023,15, 254 23 of 30 problems associated with high latency, network instability, and limited resources. This integration also leads to increased flexibility, reliability, and resource utilization. This design enhances the systematization and intelligence level of substation operation inspection by creating an intelligent operation inspection system for substation live inspection. This covers various aspects such as decision making, wireless communication, data processing, data sensing, and analysis. Lastly, Cumino and Sargento [ 70 ] discussed the deployment of UAVs in 5G and beyond MEC scenarios, along with how they may be used as mobile base stations and communication relays. This paper described the difficulties in implementing flying edge computing. The authors suggested a multi-tier architecture using UAVs with various mobility models, and assesses the effectiveness of UAVs as edge nodes. This study also described the benefits of a multi-tier viewpoint and talked about the topology and design of a smart city network. The article concludes by summarizing related research on UAV networks, job offloading, and edge computing models that might increase the available computational resources, and it highlights current research gaps and future directions. 5. Discussion Edge computing is a rapidly emerging and dynamically diversifying field that allows for data processing to occur closer to the source of data, rather than solely relying on centralized data centers. This review offers a thorough analysis of edge computing task allocation techniques. This draws attention to the difficulties in choosing the best place for tasks depending on resources such as the processing power, storage, and network bandwidth as well as in adjusting to the network’s dynamic nature. The research explores several task allocation strategies, including centralized, decentralized, hybrid, and machine learning algorithms, and assesses their efficiency in completing challenging tasks. After applying the inclusion and exclusion criteria, a total of 69 papers were included in our SLR. Based on the network types, deep learning compression techniques, and job allocation optimization algorithms utilized in edge computing, the research gives a comparative analysis of task allocation approaches. Overall, the paper highlights the difficulties and complexities of allocation in edge computing and underlines the emerging need to achieve a compromise between a number of competing goals, including energy economy, data privacy, security, latency, and QoS. The optimization techniques are the main source of challenge regarding task distribution. Dynamic scaling, resource-aware task distribution, resistive PSO, and meta-heuristic optimization approaches are some of the typical optimization strategies. The selection of an optimization approach must take into account the particular needs of the application and strike a balance between a number of competing goals, including energy efficiency, data privacy, security, latency, and QoS. This review assesses the efficacy of each method that is suggested for further study, emphasizing the difficulty of job distribution in edge computing and the need to compromise between conflicting objectives. These techniques are applied in various networks such as IoT networks, MEC networks, fog networks, vehicular networks, 5G networks, and distributed networks. Deep learning compression methods play a decisive role in choosing the appropriate technique for task allocation in edge computing. General deep learning compression, conventional neural networks, knowledge distillation, auto-encoders, and pruning are among the deep learning compression techniques that merit additional study. These techniques were applied in a number of suggested task optimization algorithms in many network types, including satellite, wireless, cellular, MEC, and the IoT [ 86 – 90 ]. The development of more effective and efficient task allocation mechanisms for edge computing may be influenced by further investigation into these compression techniques. The main contribution of our work is providing a solid answer to each of the research questions we made in Section 2. Regarding RQ1, the most contemporary task allocation techniques in edge computing are resource-aware task allocation, distributed, dynamic, machine learning-based, energy-efficient, quality-of-service-aware, collabora- Future Internet 2023,15, 254 24 of 30 tive, and Contex-aware allocation techniques which are also shown in Figure 3. As far as answering RQ2, the most used optimization techniques are the swarm, reinforcement learning, JTORA, multi-agent, game theory, heuristic approach, and last auction-based algorithms, which are also shown in Figure 4. When it comes to answering RQ3, the most used computer networks in edge computing are the MEC, the EC, vehicular, and lastly, the 5G. After answering the RQs, it is clear that every task allocation technique, optimization algorithm, and network has its own strengths and must be carefully chosen according to the application. For example, applications such as edge-based video analytics use the resource-aware task allocation method; the distributed method is used in edge computing for IoT systems; and dynamic traffic management systems are using the dynamic task allocation method; predictive maintenance systems are based on machine learning task allocation methods; the energy-aware methods are usually used in IoT systems; and finally, the QoS-aware allocation method is used in real-time video streaming applications. The optimization algorithms and network types are depending on the application and the topology that needs addressing. Future research [ 91 – 95 ] in the area of task allocation in edge computing must deal with the difficulties of adjusting to the dynamic nature of each network and establishing a balance of trade-offs between several competing objectives such as energy efficiency, data privacy, security, latency, and QoS. In addition, it is necessary to create new techniques and algorithms that can manage the growing complexity of edge computing networks. Further study is required on hybrid strategies that mix centralized and decentralized methods to maximize work distribution while taking into account the special features of edge computing settings. Also, further research is needed in the field of task allocation optimization using machine learning methods. Lastly, research must continue to concentrate on enhancing the task allocation’s efficacy and efficiency so that edge computing systems can manage workloads that are getting more complicated. 6. Conclusions The challenge of task allocation in edge computing is complex and demanding. It calls for a thorough analysis of the needs of each application. Multiple approaches for distributing tasks on edge intelligence devices, including centralized, decentralized, hybrid, and machine learning algorithms, were explored and assessed in this systematic review of task allocation methods in edge computing. Finding the optimum location for each task based on processing power, storage, and network bandwidth needs, as well as being able to adjust to the network’s dynamic nature, are the key issues in task allocation. This review investigated and assessed studies for many competing objectives, including data privacy, security, energy efficiency, QoS, IoT, task analysis, fog computing, and computationintensive tasks, among other topics. We thoroughly covered all the 69 articles that were identified by our Scopus-based keyword combination search. We also note that the number of publications on the topic has seen an upward trend over the years, indicating the importance of task allocation in edge computing. A comprehensive understanding of task allocation approaches in edge computing was attained via the analysis of these papers. More specifically, we classified and contextualized the different types of task allocation approaches that are used, also including the network types and the deep learning compression methods for optimal task offloading. The suggested approaches include dynamic scaling for effective resource allocation in fog computing as well as resource-aware task allocation based on the priority queue and available computing resources. Additionally, simulations and field testing were used in each study to assess the efficacy of the recommended approaches; something that we reported on. Overall, the analysis comes to the conclusion that task allocation in edge computing is a difficult but necessary problem that the requires careful consideration of the application’s specific needs. Research studies on alternative job distribution strategies, competing goals, and methods such as simulation and field testing might shed light on the most effective Future Internet 2023,15, 254 25 of 30 workload distribution on edge intelligence devices. Our review provides robust evidence of the growing significance of task allocation in edge computing for IoT devices, industrial applications (under the umbrella of Industry 5.0), driverless cars, UAVs, and augmented reality, among other applications. Author Contributions: Conceptualization, V.P., P.A., T.L. and D.K.; methodology, V.P., P.A., A.N. and D.K.; data curation, V.P. and K.M.; writing—original draft preparation, V.P., P.A. and D.K.; writing—review and editing, T.L., K.M. and A.N.; visualization, V.P., P.A. and K.M.; supervision, D.K.; project administration, D.K.; funding acquisition, T.L. All authors have read and agreed to the published version of the manuscript. Funding: This work received funding from the European Union’s Horizon Europe Research and Innovation Programme under Grant Agreement No. 101070181. This paper only reflects the authors’ views, and the Commission is not responsible for any use that may be made of the information it contains. Data Availability Statement: Not applicable. Conflicts of Interest: The authors declare no conflict of interest. Abbreviations The following abbreviations are used in this manuscript: AA Autonomous-Agents AB Auction-Based ACO Ant Colony Optimization AP Access Point ARIMA Autoregressive Integrated Moving Average ARSU Aerial-Roadside Units AsP Assignment Problem AVA Application Variability Adaptor BPSO Binary Particle Swarm Optimization BSUB Block Successive Upper Bound CFG Coalition Formation Game CRAN Cloud Radio Access Network CTA Collaborative Task Allocation CW Crowdsourcing Workflows DAG Directed Acyclic Graph DFD Data-Flow-Driven DLA Distributed Learning Automata DNF Deep Network Flow DP Dynamic Programming DQN Deep Q-Learning Network DQN-D Deep Q-Learning Network with Double Q-Learning DRL Deep Reinforcement Learning DS-UPP Data Sensing Mechanism with User Privacy Preserved DTASM Dynamic Task Allocation and Service Migration DTOMALB Distributed Task Offloading Algorithm Based on Multi-Agent and Load Balancing DVFS Dynamic Voltage and Frequency Scaling EC Edge Computing ECLAM Energy and Latency Minimizer ECTA Edge Computing Task Allocation EDAF Edge-Deployment Alternatives Finder EEFFRA Energy-Efficient Resource Allocation ELB Enhanced Load-Balancing ET Energy Transmitter FAP Fog Access Points FL Federated Learning