Adaptive Task Scheduling in Edge-Fog-Cloud with Network Failure Resilience

Adaptive Task Scheduling in Edge-Fog-Cloud with Network Failure Resilience Abstract—Task scheduling across the edge–fog–cloud continuum faces significant challenges from dynamic network conditions, including bandwidth fluctuations, intermittent connectivity, and variable latency. This paper presents an enhanced comparative study of machine learning schedulers with quality-of-service thresholds versus traditional heuristic approaches under both stable and failure-prone network scenarios. Using the EdgeSimPy simulation framework integrated with the real-world AI4Mobile iV2I+ dataset, we evaluate three ML models in two configurations, compared against four heuristic schedulers across 1000 tasks per experiment. The enhanced implementation integrates federated learning with encryption-based privacy preservation, multi-tenant scheduling with differentiated service policies, realtime constraint monitoring, and mobility-aware node management. Results demonstrate that threshold-constrained ML schedulers achieve optimal balance across latency, energy consumption, and cost metrics, reducing energy consumption by up to 44% versus ML-only approaches and 50% versus heuristics. Under failure-prone conditions with 10% probability of datarate degradation, threshold-constrained ML maintains robust performance with 95.7–97.2% success rates, demonstrating superior adaptability compared to pure ML approaches. Index Terms—Edge Computing, Fog Computing, Task Scheduling, Machine Learning, QoS-aware Scheduling, Dynamic Networks, Federated Learning, Multi-tenancy I. INTRODUCTION The proliferation of latency-sensitive applications and Internet of Things (IoT) devices has created unprecedented demand for efficient computational resource management. The edge–fog–cloud continuum offers a hierarchical computing paradigm that places computation closer to data sources while leveraging centralized cloud resources for intensive processing [1]. However, this distributed architecture introduces complex scheduling challenges, particularly under dynamic network conditions where network jitter can significantly degrade application performance. In vehicular and wireless time-sensitive systems, including emerging 5G and 6G networks, communication jitter and latency can vary drastically depending on network load, significantly degrading deterministic performance in dense environments [2]. Such variability underscores the importance of end-to-end synchronization and intelligent scheduling across heterogeneous layers. Mobile IoT environments experience significant network fluctuations that traditional static scheduling approaches fail to address adequately. These static methods lead to suboptimal resource utilization, increased latency, and energy inefficiency under real-world dynamics. This motivates the need for adaptive scheduling strategies that can respond intelligently to changing conditions while balancing multiple optimization objectives, including latency, energy consumption, and operational cost [1], [2]. This paper makes several important contributions to edge–fog–cloud task scheduling: 1) Integrated federated learning with secure distributed model aggregation using encryption-based data protection to ensure privacy preservation. 2) Implemented multi-tenant scheduling enabling multiple tenants with differentiated service policies, where Tenant 1 processed tasks with 989 edge, 3,583 fog, and 5,275 cloud allocations, while Tenant 2 handled tasks with 969 edge, 3,578 fog, and 5,264 cloud allocations across both scenarios. 3) Introduced mobility-aware nodes with dynamic location and velocity tracking, achieving a distance variance metric of 21.54 for realistic mobile conditions. 4) Added real-time constraint monitoring using frame-rate based validation, detecting 4,487 timing violations. 5) Enhanced failure modeling through improved datarate threshold validation to prevent invalid task scheduling. The remainder of this paper is organized as follows. Section II reviews background and related work on edge–fog–cloud task scheduling. Section III presents the experimental methodology, including simulator setup, dataset preparation, and evaluation metrics. Section IV details the enhanced scheduling framework and its adaptive components, while Section V discusses results and comparative analysis under stable and failure-prone network conditions. Finally, Section VI concludes the paper and outlines future research directions. II. BACKGROUND AND RELATED WORK The edge–fog–cloud continuum represents a three-tier architecture: edge nodes provide ultra-low latency (2–10 ms) with limited resources, fog nodes offer intermediate capabilities (10–30 ms latency), and cloud datacenters deliver massive computational power at higher latency (50–100 ms). This hierarchical structure enables flexible task placement based on application requirements and network conditions [3], [4]. Recent advancements emphasize resilience in this continuum, with architectural approaches classified into capacity, trustworthiness, and efficiency to handle failures in cloud, fog, and edge systems [5]. Task scheduling in distributed systems aims to optimize multiple objectives: minimizing response time, balancing load, reducing energy consumption, and controlling operational cost. Traditional and modern approaches include: Heuristic Methods: Round Robin [6] provides simple load distribution but ignores resource heterogeneity. Least Loaded scheduling [7] targets underutilized servers but may increase network latency. Greedy algorithms optimize local decisions without global awareness. These methods offer low computational overhead but struggle with dynamic conditions and lack adaptability to network variations. Machine Learning Methods: Recent work has explored ML for adaptive scheduling. Lyapunov optimization-based peer offloading [8] addresses energy constraints and dynamic uncertainties but lacks multi-tenancy support. Clustering-based methods [9] using K-means with Random Forest demonstrate improved allocation but introduce high overhead unsuitable for real-time edge scenarios. Deep Learning Approaches: Recurrent Neural Networks and Convolutional Neural Networks [10] excel at pattern recognition but face transfer learning challenges when adapting from stable cloud environments to volatile edge networks. Federated Deep Q-Networks [11] distribute learning across nodes but depend heavily on pre-trained models that may not generalize to dynamic conditions. Transformer-based approaches [12] show promise for sequence modeling but require extensive computational resources. Federated Learning in Edge Systems: Recent advances in federated learning [13], [14] enable privacy-preserving distributed model training across edge devices. However, most implementations focus on model training rather than realtime scheduling decisions and lack integration with traditional scheduling algorithms. Pandora [15] provides container-based Compute-as-aService (CaaS) orchestration for edge computing. Key features include dynamic container migration, QoS-aware placement policies, and resource elasticity. We integrate Pandora’s orchestration concepts into our EdgeSimPy implementation through ContainerImage and ContainerLayer abstractions, enabling realistic container-based task modeling with placement constraints and migration capabilities. Existing work [8]– [11] primarily evaluates schedulers under stable network conditions typical of controlled datacenter environments. Few studies comprehensively compare ML and heuristic approaches under realistic failure-prone conditions with adaptive threshold mechanisms. Current research lacks a comprehensive evaluation of machine learning-based schedulers with QoS constraints under network failures, as well as the integration of federated learning mechanisms with realtime scheduling decisions. Moreover, there is limited comparative analysis between threshold-constrained ML models and conventional pure ML approaches. Existing studies also overlook mobility-aware scheduling in dynamic edge environments and fail to address multi-tenant scheduling with differentiated service policies. Table I summarizes limitations in current research and highlights our contributions. TABLE I COMPARISON OF EXISTING TASK SCHEDULING APPROACHES Ref Approach Dynamic Key Limitations [8] Lyapunov Offloading ✓Energy limits; No multi-tenancy [9] K-means + RF ✓High overhead; No multi-tenancy [6] Round Robin ×Static; Poor adaptability [10] RNN/CNN ✓Transfer issues; High latency [11] Federated DQN ✓Data dependency; No heuristics [7] Genetic Algorithm ×Manual config; Not real-time [13] Federated Learning ✓Focus on training only Ours ML + FL + Thresholds ✓Multi-tenant; Mobility-aware III. EXPERIMENTAL METHODOLOGY This section outlines the experimental design, including system architecture, simulation framework, dataset preprocessing, failure modeling, and evaluation metrics. A. System Architecture and Workflow Figure 1 illustrates the overall workflow of the EnhancedEdgeSimPyIntegratedSimulator, which integrates simulation, learning, scheduling, and performance evaluation within a unified edge–fog–cloud framework. The process begins with the EdgeSimPy simulator, which models the distributed computing environment comprising heterogeneous edge, fog, and cloud servers connected through dynamic network links. This simulation layer captures mobility patterns, varying network conditions, and task characteristics to enable realistic experimentation. User requirements are incorporated to define simulation parameters and initiate system execution. Within this environment, the framework establishes a foundation for adaptive resource orchestration under diverse and fluctuating workloads. At the core of the system, the joint module performs model training, federated aggregation, and multi-tenant scheduling to ensure secure, intelligent, and fair task distribution. Machine learning models are trained on the AI4Mobile dataset to predict optimal execution layers, while federated learning protects data privacy through encrypted parameter sharing. The multi-tenant scheduler allows multiple service policies to coexist efficiently, adapting decisions based on latency, energy, and cost objectives. The trained models are deployed across the edge–fog–cloud layers through adaptive scheduling, followed by continuous performance analysis that evaluates system efficiency and feeds insights back into the optimization loop. B. EdgeSimPy Simulation Framework and Infrastructure Setup EdgeSimPy provides a cost-effective simulation environment for edge–fog–cloud systems, offering 90-95% cost reduction compared to real-world testbeds [16]. Key advantages include reproducible experiments with controlled network conditions, scalable testing, safe failure injection, and comprehensive metrics collection without instrumentation overhead. The simulator models realistic network behavior with configurable latency, bandwidth, and failure patterns. The simulator is configured with realistic three-tier infrastructure: 4 edge servers (16,000 CPU units, 32 GB memory, 200 W power, $0.50/hour), 2 fog servers (64,000 CPU units, 128 GB memory, 500 W power, $1.00/hour), and 1 cloud server (200,000 CPU units, 512 GB memory, 2000 W power, $5.00/hour). Network topology includes edge-edge links (2 ms latency, 10 Gbps), edge-fog links (10-12 ms latency, 1 Gbps), and fog-cloud links (50-55 ms latency, 500 Mbps). Random jitter of ±20% is applied to all latencies to simulate network variability. C. Dataset Preprocessing and Task Generation The AI4Mobile iV2I+ dataset captures an autonomous cleaning robot connected to a 4G campus network over 16 hours of indoor/outdoor movement [17]. We extract multiple data sources including cell metrics (RSRP, RSRQ, RSSI, SINR), network performance (datarate, jitter, ping latency), and mobility data (GPS coordinates, speed, position). Preprocessing involves concatenating five data sources to minimum common length, converting all columns to numeric format, and removing NaN values. To simulate real-world sensor uncertainty, we inject 10% Gaussian noise using x = x + N(0, 0.1 · x) for all features except datarate. Adaptive thresholds are computed from the datarate distribution to enable dynamic QoS adaptation. Each task is derived from dataset rows with computational demands mapped from network and mobility characteristics. CPU demand is computed as max(1,⌊datarate 106⌋), while memory demand is max(100,⌊speed ×100⌋). User coordinates and velocity are updated per task based on ROS position and twist measurements, enabling realistic mobility simulation. D. Enhanced Failure Models We evaluate two network scenarios to assess scheduler robustness. The stable scenario has zero failure probability representing ideal network conditions. The failure-prone scenario introduces 10% probability of datarate reduction per task, simulating realistic network degradation. Tasks with reduced datarate below 5 Mbps are explicitly rejected with latency penalty, preventing invalid scheduling attempts. This enhanced failure handling addresses previous limitations where failed tasks were improperly accounted for in performance metrics. E. Evaluation Metrics Comprehensive evaluation includes seven key metrics: mean latency across all tasks including penalties, latency standard deviation measuring variability, total energy consumption computed from server power and utilization, operational cost based on per-server rates, layer distribution showing edge/fog/cloud/failed percentages, success rate measuring successfully scheduled tasks, and degradation rate tracking datarate reduction events. These metrics provide holistic assessment of scheduler performance across multiple optimization objectives. Fig. 1. Workflow of EnhancedEdgeSimPyIntegratedSimulator. IV. ENHANCED SCHEDULING FRAMEWORK This section introduces the proposed enhanced scheduling framework that integrates federated learning, adaptive thresholds, mobility awareness, and real-time monitoring for robust edge–fog–cloud task allocation. A. Problem Formulation We consider a three-tier computing infrastructure with distinct resource characteristics. The edge layer E= {e1, e2, ..., ene}provides limited computational resources with power consumption Pe≈200 W. The fog layer F= {f1, f2, ..., fnf}offers moderate resources with Cf≈4Ce, Mf≈4Me, and Pf≈500 W. The cloud layer C= {c1, c2, ..., cnc}delivers massive capacity with Cc≈12Ce, Mc≈16Me, and Pc≈2000 W. Tasks arrive as stream T={t1, t2, ..., tn}, each characterized by CPU demand, memory demand, network features (datarate, SINR, ping, jitter, RSRP, RSSI), user position (xi, yi), and velocity (vx, vy). The scheduling problem assigns each task tito server sj∈ E∪F∪Cto minimize: min  α·¯ L+β·X j Pj·Uj+γ·X j Costj  subject to QoS constraints: cpui≤Cavail j,memi≤Mavail j, and Li< Lthreshold for latency-critical tasks. We model stable networks with zero failures and failure-prone networks with 10% datarate degradation probability, rejecting tasks below 5 Mbps threshold with explicit latency penalties. B. System Architecture and Components Our framework integrates seven components: Dataset Integration Module for AI4Mobile data preprocessing with noise injection, ML Training Pipeline with federated learning aggregation, Threshold Manager for dynamic QoS updates every 200 tasks, Scheduler Engine with seven strategies, Federated Learner using Fernet encryption with <1ms overhead, MultiTenant Manager supporting differentiated policies (high/low priority) with performance isolation, and Real-Time Monitor tracking 4,487 constraint violations for frame-rate validation. Model aggregation uses federated averaging: θglobal = 1 NPN i=1 decrypt(θencrypted i), ensuring no raw data sharing between nodes. C. Mobility and Adaptive Mechanisms MobileAwareNode tracks location (x, y), velocity (vx, vy), and mobility history with distance variance of 21.54, enabling accurate latency estimation and handover prediction from ROS position data. Adaptive thresholds are computed from datarate quantiles: T33 = max(Q0.33(DR), DRmin)and T66 = max(Q0.66(DR),1.5·T33), where Qp(DR)is the p-th percentile and DRmin = 5 Mbps, enabling dynamic adaptation to network conditions. D. ML Scheduler with Thresholds The ML scheduler leverages federated learning for privacypreserving task placement. Algorithm 1 presents the workflow: it first extracts and encrypts network features (datarate, SINR, ping, speed, RSRP, RSSI, jitter) using federated learning encryption. Tasks with ping <20 ms are directly assigned to edge servers for ultra-low latency. For other tasks, the ML model predicts the optimal layer (Edge/Fog/Cloud) based on decrypted features. Within the predicted layer, the scheduler selects the server with minimum average utilization over the last 5 tasks: s∗= arg mins∈candidates 1 5P5 k=1 Ut−k s, subject to resource constraints Cavail s≥cputand Mavail s≥ memt. If no suitable server exists, the algorithm performs a global search across all layers as fallback, ensuring guaranteed task placement. Algorithm 1 Enhanced ML with Thresholds Scheduler Require: Task t, ML model M, thresholds {T33, T66} Ensure: Assigned server s∗ 1: Extract features: f← [DR, SINR, ping, speed, RSRP, RSSI, jitter] 2: fenc ←FederatedLearner.encrypt(f) 3: fdec ←FederatedLearner.decrypt(fenc) 4: if t.ping < 20 then 5: candidates ←Edge servers 6: else 7: layer ←M.predict(fdec) 8: candidates ←Select layer (Edge/Fog/Cloud) 9: end if 10: s∗←arg mins∈candidates1 5P5 k=1Ut−k s 11: subject to Cavail s≥cput∧Mavail s≥memt 12: if s∗=None then 13: Fallback to global search 14: end if 15: Update s∗.avail cpu,s∗.avail memory 16: return s∗ V. RESULTS AND ANALYSIS This section presents and analyzes the experimental results under both failure-prone and stable network scenarios, evaluating scheduler performance across latency, energy, and cost metrics. A. Experiment 1: Failure-Prone Network Table II presents results under 10% datarate degradation probability. Threshold-constrained ML schedulers achieve lowest energy consumption (18.99–21.64 kW), representing 36.2–43.8% reduction versus ML-only (29.89–35.03 kW) and 38.1–49.9% reduction versus heuristics (30.62–37.95 kW). The degradation rate ranges from 9.0–11.5% across all schedulers, confirming realistic failure injection. Failed task percentages are maintained at 2.8–4.3%, demonstrating effective failure handling with explicit rejection of tasks below datarate thresholds. Threshold-constrained ML with SVC achieves the best energy efficiency (18.99 kW) with 96.5% success rate and moderate latency (450.32 ms). ML-only schedulers exhibit strong cloud bias (59.7–65.5%) resulting in significantly higher energy consumption. Heuristic schedulers achieve lower latency (402.46–533.32 ms) but at substantial energy cost. The Random scheduler demonstrates surprisingly competitive performance with 96.7% success rate and 30.68 kW energy consumption. Figure 2 illustrates the mean latency comparison across all schedulers under failure-prone network conditions. The visualization clearly demonstrates that threshold-constrained ML schedulers maintain balanced latency performance while heuristic approaches like Round Robin achieve minimal latency at the cost of other optimization objectives. Fig. 2. Mean latency comparison across all schedulers under failure-prone network conditions showing threshold-constrained ML maintains balanced performance. B. Experiment 2: Stable Network Table III shows results under stable conditions with zero failure probability. Threshold-constrained ML maintains energy efficiency (20.69–21.64 kW) with 100% success rate and moderate latency (104.09–106.83 ms). The energy consump- TABLE II PERFORMANCE UNDER FAILURE-PRONE NETWORK (10% DATARATE DROP) Scheduler Edge % Fog % Cloud % Failed % Lat (ms) Lat Std Energy (kW) Cost ($) Succ % ML+Th (RF) 31.9 18.3 46.7 3.1 412.15 1717.87 19.76 0.74 96.9 ML+Th (GB) 33.6 16.3 46.6 3.5 450.97 1821.49 20.16 0.74 96.5 ML+Th (SVC) 30.9 20.5 45.1 3.5 450.32 1821.55 18.99 0.73 96.5 ML-Only (RF) 7.5 27.8 61.9 2.8 414.13 1629.37 31.47 0.95 97.2 ML-Only (GB) 9.4 27.7 59.7 3.2 450.12 1738.78 29.89 0.92 96.8 ML-Only (SVC) 0.0 30.7 65.5 3.8 522.11 1885.76 35.03 1.00 96.2 Random 15.6 53.7 27.4 3.3 411.72 1773.42 30.68 0.55 96.7 Round Robin 14.5 54.9 27.4 3.2 402.46 1747.16 30.62 0.55 96.8 Least Loaded 23.7 26.6 45.4 4.3 533.32 2009.27 36.58 0.74 95.7 Greedy 11.3 48.0 37.2 3.5 448.04 1821.47 37.95 0.67 96.5 tion is remarkably consistent across all three ML models with thresholds, demonstrating robustness of the approach. Heuristics achieve best latency (84.19–109.43 ms) but consume 49.9–96.8% more energy (32.23–40.76 kW). MLonly schedulers exhibit strong cloud bias (63.3–70.2% cloud allocation) resulting in 58.7–87.6% higher energy consumption compared to threshold-constrained variants. The Round Robin scheduler achieves low latency (86.00 ms) but at significant energy cost (32.23 kW). All schedulers maintain 100% success rate under stable conditions, validating the effectiveness of the failure detection mechanisms. Figure 3 presents the total energy consumption comparison for both stable and failure-prone scenarios. The visualization clearly shows that threshold-constrained ML schedulers consistently achieve the lowest energy consumption across both network conditions, with ML+Th (SVC) and ML+Th (GB) demonstrating superior energy efficiency. C. Comparative Analysis Threshold-constrained ML demonstrates optimal balance across objectives with 30.9–34.6% edge, 16.3–20.5% fog, and 45.1–49.2% cloud allocation. This balanced distribution contrasts sharply with ML-only approaches that heavily favor cloud resources (59.7–70.2%), and heuristics that either over-utilize fog (53.7–57.0% for Random/Round Robin) or distribute inefficiently. Energy consumption analysis reveals striking differences between approaches. Under failure-prone conditions, threshold-constrained ML consumes 18.99–21.64 kW compared to 29.89–35.03 kW for ML-only (36.2–43.8% reduction) and 30.62–37.95 kW for heuristics (38.1–49.9% reduction). Under stable conditions, the energy advantage persists with threshold-constrained ML using 20.69–21.64 kW versus 32.78–38.80 kW for ML-only and 32.23–40.76 kW for heuristics. The ML+Th (SVC) configuration achieves the lowest energy consumption in failure-prone scenarios (18.99 kW), while ML+Th (GB) excels in stable conditions (20.69 kW), as illustrated in Figure 3. Latency performance shows expected trade-offs, as visualized in Figure 2. Heuristic schedulers achieve lowest latency in stable conditions (84.19– 109.43 ms) but this advantage comes at substantial energy cost. Threshold-constrained ML maintains competitive latency (104.09–106.83 ms stable, 412.15–450.97 ms failure-prone) while optimizing energy efficiency. ML-only approaches suffer Fig. 3. Total energy consumption comparison across all schedulers for both stable and failure-prone network scenarios, demonstrating superior energy efficiency of threshold-constrained ML approaches. from both higher latency (136.98–150.76 ms stable, 414.13– 522.11 ms failure-prone) and energy consumption due to excessive cloud utilization. Success rates remain high across all schedulers (95.7– 97.2%) under failure-prone conditions, demonstrating robust performance. The enhanced failure detection mechanism successfully limits failed task percentages to 2.8–4.3%, significantly improved from previous implementations. Degradation rates ranging from 9.0–11.5% confirm realistic failure injection matching the 10% probability configuration. Cost analysis reveals threshold-constrained ML achieves lowest operational costs ($0.76–0.78 stable, $0.73–0.74 failure-prone) compared to ML-only ($0.97–1.06 stable, $0.92–1.00 failure-prone) and most heuristics ($0.56–0.78). While Round Robin and Random TABLE III PERFORMANCE UNDER STABLE NETWORK CONDITIONS Scheduler Edge % Fog % Cloud % Failed % Lat (ms) Lat Std Energy (kW) Cost ($) Succ % ML+Th (RF) 33.0 17.8 49.2 0.0 106.83 86.76 21.64 0.78 100.0 ML+Th (GB) 34.6 17.7 47.7 0.0 104.09 86.99 20.69 0.76 100.0 ML+Th (SVC) 32.1 19.2 48.7 0.0 106.66 86.42 21.15 0.77 100.0 ML-Only (RF) 7.3 27.7 65.0 0.0 140.06 72.64 34.65 0.99 100.0 ML-Only (GB) 9.0 27.7 63.3 0.0 136.98 74.29 32.78 0.97 100.0 ML-Only (SVC) 0.1 29.7 70.2 0.0 150.76 64.81 38.80 1.06 100.0 Random 15.3 57.0 27.7 0.0 84.19 67.80 32.46 0.56 100.0 Round Robin 14.8 56.4 28.8 0.0 86.00 68.31 32.23 0.58 100.0 Least Loaded 23.1 29.0 47.9 0.0 109.43 81.19 39.51 0.78 100.0 Greedy 11.3 48.1 40.6 0.0 104.53 73.78 40.76 0.71 100.0 achieve lower costs ($0.55–0.58), their energy inefficiency limits applicability to sustainable edge deployments. Federated learning overhead remains minimal with encryption operations adding less than 1 ms per task. Multi-tenant scheduling successfully distributed 20,000 total tasks across two tenants with differentiated policies: Tenant1 (989 edge, 3,583 fog, 5,275 cloud, 185 failed) and Tenant2 (969 edge, 3,578 fog, 5,264 cloud, 157 failed). Real-time monitoring detected 4,487 violations, providing foundation for adaptive rate control. Mobility awareness with 21.54 distance variance enabled accurate latency estimation for mobile scenarios. VI. CONCLUSION This study demonstrates that combining machine learning with adaptive QoS thresholds provides an optimal balance for edge–fog–cloud task scheduling in energy-constrained mobile IoT environments. Threshold-constrained ML schedulers achieve superior energy efficiency while maintaining competitive latency and robust performance under network failures, reducing energy consumption by up to 44% compared to MLonly approaches and 50% compared to heuristics. The integration of federated learning with minimal overhead (<1 ms per task), multi-tenant scheduling, and mobility awareness demonstrates practical feasibility for real-world deployments. While heuristic approaches offer simplicity and lower latency, their significantly higher energy consumption limits their applicability to sustainable edge systems. The comprehensive evaluation across ten scheduler configurations and two network scenarios validates that adaptive QoS-aware scheduling with federated learning represents a promising direction for next-generation edge–fog–cloud systems supporting diverse IoT applications. Improvements stem from adaptive thresholding, QoS-driven placement, and mobility-aware scheduling, ensuring stable and efficient operation. REFERENCES [1] C. Dobre, V. Suciu, A. Nita, and A. Marinescu, “A survey on iot–edge–cloud continuum systems: Status, challenges, use cases, and open issues,” Future Internet, vol. 15, no. 12, p. 383, 2023. [Online]. Available: https://www.mdpi.com/1999-5903/15/12/383 [2] J. Haxhibeqiri, X. Jiao, E. Municio, J. M. Marquez-Barja, I. Moerman, and J. Hoebeke, “Bringing time-sensitive networking to wireless professional private networks,” Wireless Personal Communications, vol. 122, no. 4, pp. 3171–3194, 2021. [Online]. Available: https: //link.springer.com/article/10.1007/s11277-021-09056-0 [3] W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, “Edge computing: Vision and challenges,” IEEE Internet of Things Journal, vol. 3, no. 5, pp. 637–646, 2016. [4] A. Yousefpour, C. Fung, T. Nguyen, K. Kadiyala, F. Jalali, A. Niakanlahiji, J. Kong, and J. P. Jue, “All one needs to know about fog computing and related edge computing paradigms: A complete survey,” Journal of Systems Architecture, vol. 98, pp. 289–330, 2019. [5] M. Ali, S. Anjum, S. U. Khan, and A. Y. Zomaya, “Architectural resilience in cloud, fog and edge systems: A survey,” IEEE Access, vol. 8, pp. 89 291–89 310, 2020. [6] S. Guo, J. Liu, Y. Yang, B. Xiao, and Z. Li, “Energy-efficient dynamic computation offloading and cooperative task scheduling in mobile cloud computing,” IEEE Transactions on Mobile Computing, vol. 18, no. 2, pp. 319–333, 2019. [7] Y. Mao, J. Zhang, and K. B. Letaief, “Dynamic computation offloading for mobile-edge computing with energy harvesting devices,” IEEE Journal on Selected Areas in Communications, vol. 34, no. 12, pp. 3590– 3605, 2016. [8] L. Chen, S. Zhou, and J. Xu, “Computation peer offloading for energyconstrained mobile edge computing in small-cell networks,” IEEE/ACM Transactions on Networking, vol. 26, no. 4, pp. 1619–1632, 2018. [9] M. T. Thai, N. H. Nam, and M. T. Hoang, “Resource management in mobile edge computing: A comprehensive survey,” ACM Computing Surveys, vol. 56, no. 4, pp. 1–38, 2023. [10] Y. Li, M. Chen, Z. Liu, and Y. Zhang, “Deep reinforcement learning for task offloading in mobile edge computing systems,” IEEE Internet of Things Journal, vol. 9, no. 2, pp. 1517–1530, 2022. [11] P. Zhao, H. Jiang, J. Liu, and F. R. Yu, “Federated deep reinforcement learning for resource allocation in heterogeneous networks,” IEEE Transactions on Wireless Communications, vol. 21, no. 6, pp. 3724– 3737, 2022. [12] W. Zhang, Y. Xie, and F. Wang, “Transformer-based resource allocation for mobile edge computing,” IEEE Internet of Things Journal, vol. 10, no. 8, pp. 6895–6908, 2023. [13] Q. Yang, Y. Liu, T. Chen, and Y. Tong, “Federated machine learning: Concept and applications,” ACM Transactions on Intelligent Systems and Technology, vol. 10, no. 2, pp. 12:1–12:19, 2019. [14] P. Kairouz, H. B. McMahan, B. Avent, A. Bellet, M. Bennis et al., “Advances and open problems in federated learning,” Foundations and Trends in Machine Learning, vol. 14, no. 1-2, pp. 1–210, 2021. [15] C. Santos, A. M. Rego, A. Passito, R. L. Oliveira, D. Moura, and A. E. Schaeffer-Filho, “Pandora: A container-based edge cloud paas architecture based on raspberry pi clusters,” in 2016 4th International Conference on Future Internet of Things and Cloud Workshops (FiCloudW). IEEE, 2016, pp. 117–122. [16] P. S. Souza, T. Ferreto, and R. N. Calheiros, “Edgesimpy: Python-based modeling and simulation of edge computing resource management policies,” Future Generation Computer Systems, vol. 148, pp. 446– 459, 2023. [Online]. Available: https://www.sciencedirect.com/article/ pii/S0167739X23002340 [17] R. Hernang´ omez and et al., “Toward an ai-enabled connected industry: Agv communication and sensor measurement datasets,” IEEE Communications Magazine, vol. 62, no. 4, pp. 90–95, 2024.