Computation for Reinforcement Learning at the Mobile Edge Network

Abstract

Wireless Mobile Edge Computing (MEC) is one of several promising models emerging in recent years. A wireless powered MEC network is researched in this paper. The proposed online offload framework based on deep reinforcement learning (DROO) draws lessons from the previous offload experience. As such, it fulfills the need to solve complex problems like MIP. The computational complexity will not surge with the increase of network size. DROO decomposes the original optimization problem into a secondary problem from design to unloading decision and a resource allocation sub-problem. It is suitable for the space without interruption, and there is no need to discretize the channel gain, so as to avoid the disaster caused by the problem of dimension. By studying the simulation results, the DROO algorithm is found that it can achieve almost perfect performance by designing and calculating the method at the current stage, but it also needs to reduce CPU execution latency by at least one order of magnitude, so as to ensure that the realtime system optimization is fading. It is really feasible to use MEC networks when wirelessly powered in the environment.