Distributed Internet of Things Load Balancing using Deep Reinforcement Learning

Abstract

The current emerging fifth-generation (5G) system has a significant impact on the usage of Internet of Things (IoT) devices. Load balancing plays an important role in the distributed system, as it is directly associated with the performance of the whole system. Therefore, in this paper, the author recommends a load-balancing architecture for distributed IoT systems based on deep reinforcement learning (DRL), which is capable of dealing with dynamic and large-scale network situations. Specifically, the author recommends implementing a two-layer load balancing architecture. The top layer uses the long short term memory (LSTM) based Dueling Double Deep Q-Learning Network (D3QN) model for clustering the IoT devices. The bottom layer uses the plain DRL with more than one behavior policy on joint exploration. The key objective of this paper is to improve load balancing by using the technique mentioned above, as the data source and the infrastructure of the distributed IoT system can be dynamic. Experiments are conducted by using real-world datasets for evaluating the implementation. The outcome shows that the implementation of DRL on load balancing has indeed achieved a significant improvement in the performance of the distributed IoT system compared to other simplified DRL models and static clustering methods.