Multi-user motion state task offloading strategy for load balancing in mobile edge computing networks

Abstract

In mobile edge computing (MEC) networks, users can offload computational tasks from their devices to nearby mobile edge servers, reducing their computational loads and improving user experience quality. However, users exhibit various movement patterns with inherent random mobility in practice. Additionally, data that needs processing arrives randomly over continuous periods. To stabilize data and energy consumption in complex real-world environments and maximize the network system’s data processing capacity, we propose a User Trajectory Prediction-Lyapunov-guided Deep Reinforcement Learning (UTP-LyDRL) algorithm. This algorithm first predicts the movement trajectories of mobile users (MUs) using a Mobility-aware Offloading (MO) mechanism. It then formulates the problem of both MUs and fixed users (FUs) as a Mixed Integer Nonlinear Programming (MINLP) problem. Through Lyapunov optimization, the multi-stage MINLP problem is decomposed into deterministic MINLP sub-problems for each time frame, ensuring long-term constraint satisfaction. Subsequently, combining model-free training with DRL, the algorithm addresses the binary offloading of FUs across sequential time frames and overall system resource allocation. Simulation results indicate that the proposed UTP-LyDRL algorithm optimizes computational performance and ensures the stability of all data and energy queues within the system.