In a previous paper, artificial neural networks were employed for modelling average roughness Ra in rough honing processes as a function of process variables, namely grain size, density, linear speed, tangential speed and pressure, by means of the direct problem. In addition, neural network model was compared to statistical models for modelling roughness. In the present paper the inverse problem was studied and analyzed by means of neural networks, in which given a certain average roughness Ra value, the model predicts process variables to be employed. This is not possible with statistical models. Two different approaches were considered: use of a single network or use of five networks.