The optimization of designs of electro-technical systems often is based on expensive calculations (i.e., simulation models). To cut down the cost, surrogate models are constructed from and then used in place of simulations. Addressing the surrogate modelling by using the standard subset selection techniques (implemented in popular statistical software packages) requires user to predefine the full set of basis functions (mostly by setting the maximal order of the full model). However, in many cases the necessary set of basis functions is unknown and needs to be guessed, resulting in a non-trivial (and long) trial and error process. We consider a different approach – letting the regression model building method itself construct the basis functions necessary for model building. We also demonstrate the method in an empirical application.