A comparison of neural networks, non-linear biased regression and a genetic algorithm for dynamic model identification

A variety of non-linear modeling techniques were applied to a single input/single output dynamic model identification problem. Results of the tests show that the prediction error of an artificial neural network with direct linear feed through terms is nearly as good or better than the other methods when tested on new data. However, non-linear models with nearly equal and occasionally better performance can be developed (including the selection of the model form and order) with a genetic algorithm (GA) in far less computer time. The GA derived models have the additional advantage of being more parsimonious and can be reparameterized, if need be, extremely rapidly. The non-linear biased regression techniques tested typically had larger, though possibly acceptable, prediction errors. These model structures offer the advantage of low computational requirements and reproducibility, i.e. the same model is produced each time for a given data set.