Binary Offloading in Multi-Access Edge Computing Systems: Deep Reinforcement Learning Approach

In Multi-Access Edge Computing Systems, A UE provide tasks to edge nodes, these edge nodes may get over-whelmed in edge computing, resulting in processing lag or task dropouts. This is an issue because, particularly in a decentralized system, it might be challenging for each device to determine whether to offload duties and to which node. With the variable load dynamics at edge nodes in mind, this work tackles the problem of offloading tasks to reduce costs. It presents a deep reinforcement learning technique allowing UE to independently decide which tasks to offload without having to know what other UE have decided or what task models they need to know. To improve this we incorporate Gated Recurrent Unit(GRU), dueling and double DQN techniques, leading to simulation results of almost 400 ms average delay and average energy usage of 220 Joules and the runtime taking 9.45 minutes to train.