The dynamic behaviour of some parts of a process can be relatively easily used as part of a process model (the initial process model). Small perturbations are applied to the process input, and the initial process model can be updated by a learning procedure to give an updated process model, which should accurately match the process.