Scalable hybrid framework for real time and non real time task scheduling in fog computing using federated reinforcement learning and PSO GA

Abstract

Abstract Fog computing offers a decentralized paradigm to address the low-latency and energy-efficiency requirements of emerging IoT applications. However, the heterogeneity of edge nodes, the dynamic nature of workloads, and the dual need for both real-time and non-real-time scheduling introduce significant challenges in task allocation. This paper presents FRAHTOS, a Federated Reinforcement Learning and Hybrid Optimization Scheduling framework, to address these issues. FRAHTOS integrates Markov Decision Process (MDP) modeling, Federated Reinforcement Learning (FRL) for real-time tasks, and a PSO-GA hybrid optimization algorithm for non-real-time scheduling. Feature preprocessing and dimensionality reduction are performed using Adaptive Variational Autoencoders (VAE), followed by clustering with GMM and DBSCAN, and lightweight labeling using decision trees. The framework further enhances system responsiveness with EDF scheduling and VARIMA-based load forecasting. Simulation results using iFogSim demonstrate 85–95% utility, 86–96% task completion, and 3-5.5 ms latency, outperforming conventional methods. Additionally, the system sustains energy consumption between 50 and 80 mJ, suitable for battery-constrained nodes. FRAHTOS delivers a robust, scalable, and adaptive solution for intelligent IoT task scheduling. Future work includes validation on real-world data and integration with advanced federated simulation platforms.