Opt DR-GNN: optimization enabled reliability prediction and hybrid deep reinforcement -graph neural network for dynamic restoration of WDM networks

Abstract

The ability of an optical network to resist and recover from failures and disruptions reflects its overall effectiveness. Many practitioners and researchers are striving to realize maximum survivability of optical network systems. This paper developed a hybrid Deep Learning (DL) approach for dynamic restoration in Wavelength Division Multiplexing (WDM) networks. The process begins with a WDM network, where the shortest communication path is determined using Dijkstra’s algorithm. Following this, link stability maintenance occurs to guarantee that the communication between the network nodes remains intact and reliable throughout a given period of time. If no failure occurs, the network continues to function normally. However, when a link failure is detected, the system triggers the dynamic restoration phase. When a failure is detected in the WDM network, the proposed Hybrid Deep Reinforcement-Graph Neural Network (Hybrid DR-GNN) predicts optimal recovery paths by combining the learning power of Deep Reinforcement Learning (DRL) and Graph Neural Networks (GNNs) to predict an optimal restoration path based on the current network state. In order to predict the reliability, Double Exponential Sea Horse Optimization Algorithm (DESHO) is proposed, which is an integration of Double Exponential Smoothing and Sea Horse Optimization Algorithm, for selecting the most efficient recovery strategy. Finally, the overall availability and approval of the WDM network are monitored during restoration for a proper restoration determination and system integrity. The experimental results say that the DESHOA-based Hybrid DR-GNN achieved a blocking probability of 0.0061, availability of 0.993, and service provisioning time of 3.698.