Deep Reinforcement Learning based Task Offloading for MEC with User Mobility

Mobile edge computing (MEC) enables various mobile devices (MDs) to offload computation-intensive tasks to nearby edge servers (ESs), which has the potential to improve energy efficiency and reduce task completion time. However, designing efficient task offloading strategies for multiple MDs and multiple ESs is non-trivial due to the mobilities of MDs, and the hybrid decisions for how many tasks and which ES to be offloaded. To deal with these challenges, we model the task offloading problem based on the deep reinforcement learning model. We propose a mobility-aware deep deterministic policy gradient (MO-DDPG) algorithm for task offloading. The MO-DDPG algorithm is based on the DDPG algorithm, however with a novel mobility-aware experience replay buffer, which assigns higher priorities to recent experiences, and samples hybrid experiences during training. Simulation results demonstrate that our MO-DDPG algorithm outperforms the baselines in terms of cost and convergence speed.