In the present work, neural networks were used for modelling average roughness Ra as a function of process parameters: grain size, density of abrasive, pressure of honing stones on the workpiece’s surface, linear speed and tangential speed. For doing this, first experimental semifinish honing tests were performed. Then results were used for selecting best configuration of the neural network, taking into account either one or two hidden layers. In addition, neural models were compared to regression models.