Abstract
While standardisation of variables is a common practice for many machine learning algorithms, it is rarely seen in the literature on genetic programming for symbolic regression. This paper compares the predictive performance of unscaled and standardised genetic programming, using artificial datasets and benchmark problems. Linear scaling is also applied to genetic programming for these problems. We show that unscaled genetic programming provides worse predictive performance than genetic programming augmented by linear scaling and/or standardisation as it is highly sensitive to the magnitude and range of explanatory or response variables. While linear scaling does provide better predictive performance on the simple artificial datasets, we attribute much of its success to an implicit standardisation within the predictive model. For benchmark problems, the combination of linear scaling and standardisation provides greater stability than only applying linear scaling to genetic programming. Also, for many of the simple artificial datasets, unscaled genetic programming produces larger individuals, which is undesirable in the search for parsimonious models.