Comprehensive analysis of network reliability prediction for multistate networks with machine learning approaches

Abstract

A multistate network (MSN) is a widely used model for constructing real-world systems. Evaluating the network reliability of an MSN is a crucial task. While existing studies have developed analytical algorithms for network reliability evaluation, these methods often prove inefficient for the complex MSN due to the NP-hard property. To address the challenges of large-scale networks, the study conducts a comprehensive investigation of various machine learning (ML) approaches to predict network reliability, aiming to assess their effectiveness and efficiency. Some commonly applied and easy-to-use ML approaches are explored: linear regression, regression trees, support vector machines (SVM), ensemble learning methods (i.e., LSBoost and Bagging), Gaussian process regression (GPR), and deep neural networks (DNN). Domain knowledge of MSN is used to transform necessary information and exact reliability into formats suitable for these approaches. A practical computer network serves as a benchmark to evaluate their performance, assessed using the coefficient of determination (R²), root-mean-square error (RMSE), mean absolute error (MAE), training time, and prediction efficiency. Prediction efficiency evaluates the speed of making predictions, while R², RMSE, and MAE assess prediction accuracy. Both are critical for real-world network reliability applications. Experimental results demonstrate that GPR and DNN achieve the best performances, with a tri-layered DNN exhibiting the most efficient prediction performance. Hyperparameter optimization for the tri-layered DNN is conducted to give the final recommendation of these ML models.