HGWO-RL: hybrid grey wolf optimization with Q-learning refinement for energy-aware multi-objective task scheduling in heterogeneous fog–cloud networks

Abstract

Energy-aware task scheduling in fog–cloud IoT networks must balance monetary cost, deadlines and energy. We propose hybrid grey wolf optimizer-reinforcement learning (HGWO-RL), which combines a grey wolf optimizer (GWO) with a tabular Q-learning agent that selects move operators to refine the three elite solutions. The model includes release times, FIFO node queues, a utilization-dependent power model and the objective F = Ctotal + λEtotal, where λ is an energy price; the problem is proved NP-hard. Seven algorithms were compared under an equal budget of 3000 evaluations, 30 paired runs and 23 configurations. HGWO-RL achieved the lowest objective in 22 configurations; relative to Min-CCV it reduced the objective by up to 3.64% and energy by up to 18.1%, and improved deadline satisfaction. From random initialisation it outperformed GA, PSO, GWO and ACO by 35–94%. HGWO-RL obtained the best Friedman mean rank (1.04), and Holm-corrected Wilcoxon tests found it significantly better than every baseline, although the Nemenyi test did not separate it from PSO and GA. The ablation study shows that elite local refinement provides most of the gain; Q-learning operator selection adds a significant 15.6% improvement over random selection under cold start, but no significant improvement under warm start.