Online Partial Computation Offloading Optimization in Wireless Powered Mobile Edge Computing Network

Computation offloading is one of the most crucial issues of wireless powered mobile edge computing (WP-MEC) network, which decides how to offload the tasks of wireless devices (WDs) to the access point (AP) for further intensive computation. However, in practical application, efficient computation offloading and resource scheduling are challenging with online factors, i.e., the time-varying channel conditions and random task arrivals. Compared with binary offloading decisions, partial offloading decisions can be flexibly adjusted according to the characteristics of the task. The proliferation of feasible partial offloading schemes also brings challenges to the optimization of computational offloading policies. To resolve the above problems, this paper proposes an online partial computation offloading scheme based on Lyapunov optimization and deep reinforcement learning (LyDROP) in a WP-MEC network, where the wireless power transfer (WPT) duration time, partial offloading decisions and transmission time allocation are jointly optimized. The objective is to maximize the weighted sum computation rate of all wireless devices while reducing the data queue backlog as much as possible. Firstly, Lyapunov optimization is used to decouple the stochastic online problem into a series of per-frame deterministic problems. Subsequently, combination of model-based optimization and model-free machine learning algorithms is adopted to perform extended simulation experiments. The performance of LyDROP is validated utilizing a real-world dataset of Melbourne in Australia. Finally, in comparison with several baseline or state-of-the-art algorithms, including Lyapunov-optimization-based coordinate decent (LyCD), Lyapunov-optimization-based randomly partial computation offloading (LyRPO) and Myopic algorithm, the superiority of LyDROP algorithm is demonstrated.