Investigation of the Effect of Artificial Neural Network Performance Parameters and Training Dataset on the Probability Estimate Capacity in Structural Reliability Problems

Abstract

This study highlights two of the important details of the implementation of artificial neural networks to the structural reliability problems by pointing out the effect of training dataset, and the relationship between the performance parameters (coefficient of determination of train, validation, and test sets) of a network and its probability estimation capacity when it is used as a surrogate model in structural reliability problems. Four numerical examples are covered regarding these key aspects including one that is derived from a real-life reinforced concrete structure. Results have shown that the dataset can affect the probability estimation capacity for complex problems. Furthermore, it is also observed that having a neural network with good performance parameters does not mean that the network always has good probability estimation capacity. However, in order to have a network that can be used for probability estimate purposes, its performance parameters must be at a satisfactory level.