Three different approaches far modelling a semi-batch polymerization reactor using artificial neural networks (ANN) have been investigated. Based on the characteristics of the semi-batch reactor a multi-stage strategy is recommended. It divides the whole reaction process into two periods, semi-batch and batch, and further divides the semi-batch part into two sub-periods that are before and after the maximum temperature is reached. Different ANN architectures are used to model the three parts separately. The results demonstrate that the multi-stage approach proposed can be used to estimate difficult-to-measure polymer variables with acceptable accuracy for semi-batch processes. Concentrations of the monomer and the initiator in the reactor are estimated from reactor temperature, feed temperature and the concentration of the initiator in the feed.