Variational Autoencoders Integrated with Reinforcement Learning for Autonomous IoT Resource Allocation and Load Balancing Optimization

Abstract

The current paper introduced a new framework combining Variational Autoencoders (VAE) with Deep Reinforcement Learning (DRL) in autonomous maximum utility resource allocation and load balancing in Internet of Things (IoT) setting. The proposed VAE-RL model overcame the difficulties of dynamic network resource management of heterogeneous IoT networks by learning latent representations of network conditions and using an actor critic architecture in optimal decisions making. We deployed and tested the framework in three different scenarios of IoT: infrastructure in a smart city, manufacturing with industrial IoTs, and edge computer setup. The results of experimental work proved our methodology showed a 34.7 percent reduction in response time, 28.3 percent better resource consumption and 41.2 percent cut in the load imbalance than the baseline methods. The framework demonstrated suitable performance in changing network conditions and scaled well to grassing that comprised of 5000 IoT devices.