Finite-element model updating is a numerical tool to improve the performance of numerical models when they are used to predict the behavior of civil engineering structures. For this purpose, some physical parameters of these models are tuned to better characterize real structural behavior. The experimental and numerical modal properties of the structure (natural frequencies and associated vibration modes) are normally considered for this comparison process. Regardless of the considered method (frequentist or stochastic) to perform the updating process, the main limitation of this technique is the high simulation time required to solve the updating problem. As the bottleneck of the process is the time required to compute a numerical modal analysis, the number of evaluations of the finite-element model must be reduced. For this purpose, the use of a surrogate model can be considered. Thus, a low fidelity model of the structure is generated based on a preliminary high fidelity finite-element model. Among the different methods, the use of an artificial neural network model is commonly considered. Despite its good results, its main limitation is the optimal selection of the hyperparameters of this surrogate model. Bayesian optimization is commonly used to cope with this issue. Thus, the performance of using Bayesian optimization for the hyperparameter tuning of neural networks when they are used for the surrogate-assisted model updating of civil engineering structures is analyzed in detail herein. For this purpose, the model updating of a benchmark structure, a steel cantilever beam has been considered.