IUS-MAAC: Multi-Agent Attention-Critic for Task Offloading in Integrated User-Server Mobile Edge Computing

In recent years, the rapid development of computation-intensive mobile applications has posed significant challenges to the computational capabilities of user devices. In response to these challenges, Mobile Edge Computing (MEC), as an emerging technology, effectively alleviates the computational pressure on user devices. Task offloading, as the core mechanism of MEC, has become a focal point in current research. Deep Reinforcement Learning (DRL) has demonstrated its unique advantages in handling task offloading. However, current DRL-based task offloading methods primarily focus on the user level, considering factors such as limited battery life and computing capacity, thereby neglecting server performance. These methods often assume that servers have sufficient computational power and storage resources. Additionally, traditional DRL performs well in handling discrete or continuous actions but still faces challenges when dealing with hybrid action spaces. Based on these, the paper proposes a novel Integrated User-Server Multi-Agent Attention-Critic (IUS-MAAC) algorithm for MEC task offloading. This takes into account various constraints during the task offloading process and demonstrates better performance in handling hybrid spaces compared to existing Multi-Agent Deep Deterministic Policy Gradient (MADDPG) and heuristic algorithms. Based on our verification, the results show that the proposed algorithm has better convergence and improves the overall reward value by approximately 9.3% compared to other state-of-the-art algorithms.