Adaptive Computation Offloading Based on Enhanced Multi-Agent Deep Reinforcement Learning

Mobile edge computing (MEC) is a promising enabler for computing-intensive or latency-sensitive services. To enhance the quality of experience (QoE), this paper proposes an adaptive computation offloading approach that minimizes the normalized sum of latency and energy consumption of all computing-intensive users based on MEC offloading, device-to-device offloading (D2D offloading), and local computing strategies. An attention-enhanced multi-agent deep deterministic policy gradient (AT-MADDPG) algorithm is proposed to figure out the optimal collaborative users, MEC servers, and their corresponding offloading ratios. Simulation results show that AT-MADDPG achieves convergence before the 80th training episode, with a 68% improvement compared to MADDPG's convergence at the 250th training episode. As the number of users increases, AT-MADDPG performs even more prominently and outperforms MADDPG by 12% and 38% in terms of latency and energy consumption when the number of users reaches 40. Compared with simply offloading to D2D or MEC, the adaptive offloading approach improves the weighted performance of latency and energy consumption by 33% and 42%, respectively.