This paper explores three different strategies for the inversion of spectral lines (and their Stokes profiles) using artificial neural networks. It is shown that a straightforward approach in which the network is trained with synthetic spectra from a simplified model leads to considerable errors in the inversion of real observations. This problem can be overcome in at least two different ways that are studied here in detail. The first method makes use of an additional preprocessing autoassociative neural network to project the observed profile into the theoretical model subspace. The second method considers a suitable regularization of the neural network used for the inversion. These new techniques are shown to be robust and reliable when applied to the inversion of both synthetic and observed data.