Adaptive Task Offloading in Coded Edge Computing: A Deep Reinforcement Learning Approach

Abstract

In this letter, we consider a Coded Edge Computing (CEC) network in which a client encodes its computation subtasks using the Maximum Distance Separable (MDS) code before offloading them to helpers. The CEC network is heterogeneous in which the helpers are different in computing capacity, wireless communication stability, and computing price. Thus, the client needs to determine a desirable size of MDS-coded subtasks and selects proper helpers such that the computation latency is within the deadline and the incentive cost is minimal. This problem is challenging since the helpers are generally dynamic and random in the computing, communication, and computing price. We thus propose to adopt a Deep Reinforcement Learning (DRL) algorithm that allows the client to learn and find optimal decisions without any prior knowledge of network environments. The experiment results reveal that the proposed algorithm outperforms the standard Q-learning and baseline algorithms in both terms of computation latency and incentive cost.