In this contribution genetic programming (GP) is used to evolve dynamic process models. An innovative feature of the GP algorithm is its ability to automatically discover the appropriate time history of model terms required to build an accurate model. Two case studies are used to compare the performance of the GP algorithm with that of filter-based neural networks (FBNNs). Although the models generated using GP have comparable prediction performance to the FBNN models, a disadvantage is that they required greater computational effort to develop. However, we show that a major benefit of the GP approach is that additional model performance criteria can be included during the model development process. The parallel nature of GP means that it can evolve a set of candidate solutions with varying levels of performance in each objective. Although any combination of model performance criteria could be used as objectives within a multi-objective GP (MOGP) framework, the correlation tests outlined by Billings and Voon (Int. J. Control 44 (1986) 235) were used in this work. (C) 2003 Published by Elsevier Ltd.