System identification via genetic programming

Compared to bit string coded GAs, GP is a more powerful system identification tool. However, there is no guarantee that GP will produce an exact solution. Perhaps that is the price associated with the increased flexibility of the paradigm. Bancroft (1995) reported that, failures to discover the correct model structure of time series were attributed to high levels of coloured noise and the presence of cross-product terms. Nevertheless, GP could usually evolve models that provide good output estimates, even when there is redundant data. This is confirmed by applications to a non-linear simulation of a reactor and pilot scale fermenter. Similar observations have been made in applications of GP to model industrial processes.