Due to its technical and economic advantages, the use of explosives in underground rock excavation is widely adopted. However, some safety and, especially, environmental issues arise when using this technique, mainly concerning ground vibrations induced by blasts. Thus, to minimize dynamic environmental impacts, prediction of blast-induced vibrations is imperative. In the last few years, artificial neural networks (ANNs) have been applied to model blast-induced ground vibrations. Nevertheless, ANN’s architecture, mainly the number of neurons in the hidden layer, has been selected manually concerning ANN’s performance parameters. To avoid over-fitting and reduce model’s complexity, this paper presents a bilevel optimization of ANN architecture, considering two transfer functions, based on the maximization of quality of the adjustment and model’s complexity, this last one as a penalty criterion. An ANN approach based on this bilevel optimization was successfully applied on a database of 1188 samples obtained from underground blast-induced ground vibration monitoring in a granitic rock mass. A residual analysis of the best-fitted model is performed to ensure the quality of the adjustment. It is demonstrated that the determined ANN model offers much higher generalization ability than the traditional prediction models usually used for blast-induced ground vibration amplitude predictions and other ANN architectures applied to ground vibration prediction.