Computation Offloading with Reinforcement Learning for Improving QoS in Edge Computing Environments

A computation offloading scheme based on edge collaboration was proposed for smoothly handling high delay-sensitive tasks in the Internet of Things (IoT). However, it shows poor service time and load balancing due to resource limitations and the number of processed task restrictions of edge servers. To solve this problem, edge collaboration using Reinforcement Learning (RL) has been proposed. RL requires a lot of exploration to maximize the cumulative reward. In a distributed environment, RL has a problem that the agent does not have enough experience due to the data sparsity for learning. Also, training variation between agents is high in a distributed environment. In this paper, we propose computation offloading with reinforcement learning for improving Quality of Service (QoS) in edge computing environments. We select offloading target based on RL for minimizing the service time and maximizing the load balance. The proposed scheme determines the priority of experience and shares the experience to improve reward. The priority of experience is calculated by Temporal Difference (TD) error and a state value. Also, the sharing of experience is proceeded on the basis of the policy gradient of agents. Experimental results show that the proposed scheme achieves a better QoS through high reward compared to the existing schemes.