A Deep Reinforcement Learning Approach for Computation Offloading at the 5G Network Edge

Multi-access edge computing (MEC) is envisioned as a key enabler for fifth generation (5G) networks, bringing new resources into the computing continuum spanning from user devices to the centralized cloud. The advantages introduced by MEC play an important role in achieving the targeted 5G key performance indicators (KPIs), but at the same time its introduction raises new questions in terms of how frequently data should be offloaded from the end user device to the edge for processing since the optical front-, mid- and backhaul are constantly employed in the data transfer and MEC capabilities are limited compared to cloud data centers. In this context, a deep reinforcement learning (DRL) computation offloading solution is proposed which aims to provide optimal binary offloading decisions (local or MEC task execution). The DRL agent is set to evaluate the trade-off between factors such as computation energy consumption, communication delay via time-varying wireless channels and optical connections between 5G base stations (gNodeBs) and multiple MEC nodes. In order to preserve user data privacy, we propose an alternative training method for the DRL algorithm, i.e., a federated learning framework, as opposed to the centralized method where training data is collected in a predefined storage location.