Enhancing a multilayer perceptron model for multi-state network reliability evaluation via an arc-wise architecture

Abstract Many real-world systems, including supply chains and manufacturing, have been modeled as multi-state networks (MSNs) to assess network reliability performance. This paper introduces an arc-wise architecture designed to enhance the efficiency of a multilayer perceptron (MLP) model for MSN reliability estimation. By representing each arc with a dedicated subnetwork and incorporating system-level requirements directly into the input layer, the proposed MLP effectively maps complex dependencies between arc capacities and system-level reliability. Integrated with Bayesian optimization for hyperparameter tuning, the proposed approach is validated through comprehensive numerical experiments. The results demonstrate high estimation accuracy and efficiency, highlighting the efficacy of deep learning for reliability analysis in complex, large-scale MSNs.