Multi-access edge computing (MEC) has attracted lots of research interests from both academic and industrial fields, which aims to extend cloud service to the network edge to reduce network traffic and service latency. One key decision-making problem in MEC systems is task offloading which aims to decide which tasks should be migrated to the MEC server to minimize the running cost. In this article, I am going to talk about how to use learning-based methods to address the above problem.

Policy gradient theorem is the basis of many deep reinforcement learning methods (e.g., PPO, TRPO, and SAC). Meanwhile, Variational Auto-encoder (VAE) is one of the popular generative methods in machine learning. Since the target problems of these approaches share the same formulation, we can use policy gradient as an alternative training method for VAE.