Multi-objective Reinforcement Learning Algorithm for Computing Offloading of Task-Dependent Workflows in 5G enabled Smart Grids

Computational offloading is considered a promising emerging paradigm for addressing the limited resources of edge devices in expanding power grids. However, with the advancement of intelligent technologies such as digitalized power grids, applications often consist of several interdependent subtasks, forming interconnected automated workflows. This paper focuses on the computational offloading technique within task-dependent workflows. It proposes a multi-objective optimization problem for offloading, considering both time and energy consumption. The model takes into account the constraints of task duration, communication capacity, and computational capacity. Additionally, a predictive-guided a predictive-guided multi-objective reinforcement learning algorithm based on Pareto optimization (MORLBP) is introduced. This algorithm combines the principles of multi-objective optimization, Pareto optimality theory, and deep reinforcement learning. It utilizes the quality of the Pareto front as a metric and is compared against NSGA-II and MOPSO algorithms. The proposed algorithm’s effectiveness and advancement are validated through simulations, demonstrating its efficiency and innovation in tackling the multi-objective offloading problem within task-dependent workflows.