Federated deep reinforcement learning enabled hierarchical Edge–Fog–Cloud architecture for intelligent task offloading in 6G networks

Latency sensitive, computation intensive and mobility aware applications in Edge Fog Cloud environments have increased the demand to develop intelligent offloading task mechanism that can dynamically scale to changing network conditions whilst remaining scalable, energy efficient, and preserving data privacy. Traditional, heuristic and centralized based learning offloading methods frequently have problems in accommodating heterogeneous workloads, non- stationary environments and privacy limitations associated with the next generation distributed computer system. In order to overcome these drawbacks, the present paper will suggest a Federated Deep Q-Learning (FDQL)-based task offloading framework, which incorporates deep reinforcement learning and federated learning to support adaptive, decentralized and privacy-conscious decision-making across hierarchical Edge Fog Cloud architectures. The framework proposed solves task offloading as a Markov Decision Process, with the execution decisions being trained based on the joint consideration of the latency, bandwidth availability, queue length, computational load, and energy state, as well as user mobility, without sharing raw data during federated model aggregation. In comparison to the current CNN-, LSTM-, SVM-, and rule-based methods, which use fixed threshold values or rely on centralized training, the FDQL architecture allows collaborative learning between distributed edge nodes, enhancing generalization as well as resilience as network conditions evolve. Large-scale experimental analysis is performed using a trace-driven simulation based on a publicly available task offloading dataset of tasks and the performance is evaluated based on the latency, energy consumption, task success rate, robustness analysis, and computational efficiency. Experimental findings indicate that the proposed FDQL framework demonstrates improved performance under distributed and resource-constrained environments compared to baseline approaches since shorter latency, increased energy efficiency, and more predictable execution-layer selection are achieved. The significance of federated learning, mobility awareness, and bandwidth-aware optimization in the stability of the performance is also confirmed by ablation studies. In order to achieve a better level of transparency and trustworthiness, SLA-based confusion matrix analysis and ROC analysis are performed as well as SHAP-based explainability analysis, which proves that the decisions made by FDQL are based on physically interesting, as well as SLA-relevant features, like latency, bandwidth, and resource use. All in all, the designed FDQL framework is a successful, interpretable, and scalable approach to intelligent task offloading, so it would fit perfectly into the implementation of the 6G-enabled application, such as smart cities, industrial internet of things, and autonomous systems in the future.