Figure 1.
Figure 2.
Figure 3.
Figure 4.
Figure 5.
Figure 6.
Figure 7.
Figure 8.
Performance Metrics
| Metric | Description |
|---|---|
| Task Completion Time (TCT) | Average time required to complete all scheduled tasks. |
| Energy Consumption | Total energy utilized across edge, fog, and cloud layers. |
| Load Balance Rate (LBR) | Standard deviation of the task load distributed across all nodes. Lower is better. |
| Task Success Ratio (TSR) | Percentage of tasks completed successfully within their specified deadlines. |
| Scheduling Overhead | Time and resource consumption incurred by the scheduling mechanism itself. |
| Model Convergence | Convergence behavior and stability of federated reinforcement learning agents over training rounds. |
Comparative Summary of Related Work
| Work | Approach / Technique | Key Contributions | Limitations / Gaps |
|---|---|---|---|
| [11] Ghanavati et al., 2020 | Ant-mating optimization for fog task scheduling | Energy-aware heuristic improves fog-level efficiency | Limited adaptability; no learning; not suitable for dynamic workloads |
| [12] Su et al., 2021 | Secure federated learning for smart grid | Privacy-preserving edge–cloud FL; secure aggregation | Not designed for online task scheduling; lacks real-time decision-making |
| [13] Baghban et al., 2022 | Actor–critic RL in federated edge computing | Improved IoT service provisioning; distributed RL training | No multi-objective modeling; no local queue optimization |
| [14] Soula et al., 2022 | Machine learning + bioinspired scheduling | Intelligent task allocation at edge; hybrid heuristics | Scalability concerns; lacks global coordination |
| [15] Ramezani et al., 2023 | RL-based scheduling in edge–fog–cloud | Multi-objective load balancing with RL | No federated coordination; centralized RL limits scalability |
| [16] Kar et al., 2023 | Survey on ML + optimization for cloud–edge–fog | Comprehensive taxonomy of offloading strategies | Highlights need for FL, multi-objective coordination, and dynamic adaptivity |
| [17] Kim et al., 2023 | Federated reinforcement learning for dynamic scheduling | Collaborative policy learning; improved coordination | No coupling between global policy and local task reordering |
| [18] Wu et al., 2024 | FRL for vehicular edge computing | Privacy-preserving scheduling for large AI models | High communication cost; no local execution optimization |
| [19] Shen et al., 2025 | RL-based scheduling in end–edge–cloud | RL for heterogeneous resource environments | Centralized data dependence; limited SLA/tardiness modeling |
| [20] Shidik et al., 2025 | Unsupervised cluster Q-learning | Energy minimization in federated edge cloud | Weak SLA handling; limited multiobjective scope |
| [21] Lilhore et al., 2025 | Hybrid cloud–edge deep learning | Scalable resource optimization for IoT | Does not integrate global offloading + local queue optimization |
Summary of Simulation Environment
| Category | Configuration Details |
|---|---|
| Task Characteristics | 50–200 heterogeneous tasks per run; varied CPU, memory, and latency requirements |
| Edge Nodes | 10 nodes with low capacity and limited energy |
| Fog Nodes | 5 intermediate nodes with moderate processing and energy |
| Cloud Data Centers | 3 high-performance data centers with abundant energy and low resource constraints |
| System Heterogeneity | Varying processing speed, power consumption, and network latency across nodes |
| Workload Types | Real-time loT streams, batch tasks, latency-sensitive applications |
Sensitivity Analysis of Key Hyperparameters
| Parameter Group | Hyperparameters Tested | Sensitivity Observation |
|---|---|---|
| FRL Module | Learning rate, discount factor, exploration rate | Minor performance variations; stable convergence within broad ranges |
| A-ABC Module | Colony size, limit value, employed/onlooker ratio | Larger colonies improve quality but increase overhead; performance remains stable across typical ranges |
| Federated Learning | Aggregation frequency, participation rate, non-IID severity | Hybrid model remains robust; slight slowdown under extreme non-IID settings but no convergence failures |
Ablation Study of System Components
| Configuration | Strengths | Limitations | Overall Outcome |
|---|---|---|---|
| FRL-Only |
|
|
|
| ABC-Only |
|
|
|
| Hybrid FRL + A-ABC (Proposed) |
|
|
|