Cautionary note on the use of genetic programming in statistical downscaling

Abstract

The selection of inputs (predictors) to downscaling models is an important task in any statistical downscaling exercise. The selection of an appropriate set of predictors to a downscaling model enhances its generalization skills as such set of predictors can reliably explain the catchment-scale hydroclimatic variable (predictand). Among the predictor selection procedures seen in the literature, the use of genetic programming (GP) can be regarded as a unique approach as it not only selects a set of predictors influential on the predictand but also simultaneously determines a linear or nonlinear regression relationship between the predictors and the predictand. In this short communication, the details of an investigation on the assessment of effectiveness of GP in identifying a unique optimum set of predictors influential on the predictand and its ability to generate a unique optimum predictor–predictand relationship are presented. In this investigation, downscaling models were evolved for relatively wet and dry precipitation stations pertaining to two study areas using two different sets of reanalysis data for each calendar month maintaining the same GP attributes. It was found that irrespective of the climate regime (i.e., wet and dry) and reanalysis data set used, the probability of identification of a unique optimum set of predictors influential on precipitation by GP is quite low. Therefore, it can be argued that the use of GP for the selection of a unique optimum set of predictors influential on a predictand is not effective. However, when run repetitively, GP algorithm selected certain predictors more frequently than others. Also, when run repetitively, the structure of the predictor–predictand relationships evolved by GP varied from one run to another, indicating that the physical interpretation of the predictor–predictand relationships evolved by GP in a downscaling exercise can be unreliable.