Multi-agent deep reinforcement learning for computation offloading in cooperative edge network

Mobile Edge Computing (MEC) has emerged as an effective paradigm for reducing latency and enhancing computational efficiency. However, the rapid proliferation of edge servers and user devices has significantly increased the complexity of task processing and resource management. Traditional task offloading approaches often rely on centralized decision-making, resulting in high computational complexity and time costs. To address these challenges, this paper introduces a dynamic collaborative framework involving multiple users and edge servers. We formulate the problem of resource allocation and task offloading as a multi-objective Markov Decision Process (MDP) with a mixed action space. To solve this, we propose a novel algorithm called Multi-Agent Mobile Edge Computing (MA-MEC), which leverages multi-agent reinforcement learning. In MA-MEC, each mobile edge server (MES) operates as an independent learning agent. Through centralized training and decentralized execution, these agents collaborate to develop efficient task offloading strategies in complex and dynamic edge environments. Simulation results demonstrate the effectiveness of our approach. MES agents learn to execute tasks more efficiently, increasing the number of processed tasks by 12.5 \(\%\) $\%$ , while task offloading rates rise by 17 \(\%\) $\%$ , and time costs are reduced by 53 \(\%\) $\%$ compared to baseline methods. The proposed method shows significant advantages, especially in resource-constrained scenarios.