Intelligent and Autonomous Edge Slicing for IoT Systems

Edge intelligence is rapidly emerging as a pivotal platform for supporting future IoT networks. The integration of artificial intelligence and machine learning (AI/ML) with edge computing furnishes a new era in which edge systems can learn environment dynamics and optimize resource autoscaling policies. However, the heterogeneity of IoT networks, characterized by diverse applications and requirements, necessitates edge systems with advanced intelligence to tailor resource autoscaling policies to specific environments. Despite ongoing research in edge intelligence, most studies have focused on a single environment or service type, thereby limiting their applicability to real-world scenarios with varied IoT services. To address this limitation, we propose the intelligent and autonomous edge slicing (IAES) system, a novel approach designed to recognize diverse IoT environments and implement per-slice resource allocation policies. IAES leverages deep reinforcement learning (DRL), namely, dueling double deep Q-networks (D3QN), to optimize resource autoscaling across distinct IoT environments, such as smart cities, eHealth, and smart factories. Additionally, IAES incorporates an intelligent environment classification component that utilizes joint traffic prediction and classification models. Several AI algorithms such as long short-term memory (LSTM), convolutional neural networks (CNN), and multilayer perceptron networks (MLP), are evaluated for their efficacy in predicting IoT environments. Simulation experiments demonstrate that the IAES system achieves a 50-60% reduction in system costs compared to both rule-based commercial autoscaling employed in Kubernetes systems and an intelligent prediction-based autoscaling algorithm.