Deep Reinforcement Learning offloading for Computation Cost Minimization of Wireless Powered Mobile Edge Computing
Hang Yu, Jun Yang · 2024
The development of the Internet of Things has increased the demand for real-time communication and computing in user equipments (UEs), which is a challenge for the limited battery capacity and computing power of UEs. Mobile edge computing (MEC) is a solution for UEs to achieve low delay and low energy consumption by offloading compution-intensive tasks to edge servers. The combination of wireless power transfer (WPT) and MEC further extends the battery lifetime of UEs. Conventional offloading methods are inadequate in adapting to environmental variations and achieving the long-term performance of the wireless powered MEC system with random task requirements and wireless channel states of the UEs. We introduce a deep reinforcement learning framework that takes into account the wireless powered MEC system's long-term average computing cost. We utilize the deep deterministic policy gradient approach to effectively acquire knowledge on optimal task schedule and runtime of the edge server. Additionally, we optimize energy transmission power, offloading time, offloading ratio, and local computation time simultaneously to assess the performance of actions. To enhance exploration of actions, we introduce the differential evolution algorithm which expands and variegate the action generated by the policy network while also accelerating convergence. The simulation outcomes indicate that our proposed scheme outperforms four benchmark strategies in terms of reduced computing costs and accelerated convergence rate.