Data Plane Failures and Recovery Techniques in SDN-IoT Networks Based on Reinforcement Learning

Abstract

Node density, limited resources, and dynamic link states in SDN-based Internet of Things (IoT) networks make data-path management difficult. In this framework, we rely on Support Vector Regression (SVR) to predict link reliability and then we find optimal paths to send data in Internet of things (IoT) using reinforcement learning (RL)-based routing. Link reliability data (link downtime, link uptime, and link failure reasons) are the network data utilized to train the SVR model [44]. The predictions from these models help the RL agent to create the state space as link reliability, switch load, bandwidth, delay, and packet loss and the action space as paths with reliability ℜ≥ 0.6. The reward function combines these aspects together to choose low-cost and high-confidence paths. To demonstrate the ability of the SVR in predicting link reliability, the SVR model prediction accuracy reaches an 85.3% accuracy by performing simulations on four real network topologies (Abilene, USNet, OS3E and DFN) [19]. The paths selected by the RL agent with the least amount of delay, more bandwidth, and lower packet loss after 1,000 episodes. Compared with RSIR and Sway, our proposed framework achieves a better balance between QoS and throughput and performs more stable under increased network complexity. The analysis illustrates that SVR-RL integration not only improves link failure recovery but also intelligent routing in SDN-IoT networks, resulting in enhanced reliability and service quality..