RL-Based Load Balancing for Edge Server Based Vehicular Offloading System

Intelligent vehicular applications are currently under active research. However, there are limitations in meeting the service requirements with the vehicle's limited resources. Therefore, the methodology of offloading tasks to edge servers is gaining attention. However, edge servers are resource-constrained compared to centralized servers, and this can lead to overload and localized offloading failures with high traffic or when a large number of high-computation tasks are requested. In this paper, we propose an algorithm designed to address the problem of load imbalance among edge servers in a vehicular offloading system. The proposed algorithm aims to maximize the utilization of the limited edge servers' resources to increase the offloading success rate. We have formulated an optimization problem to form a cluster among edge servers and minimize the deviation of available resources within the cluster. To facilitate optimal load distribution in dynamic environments, we propose an algorithm that utilizes a reinforcement learning model TD3 (Twin Delay DDPG). Through experiments, we demonstrate that the proposed algorithm effectively alleviates load imbalance, particularly in environments with a growing number of vehicles or an increasing diversity of tasks. The algorithm achieves a high offloading success rate and confirms the scalability improvement resulting from load balancing.