Surrogate modelling techniques are widely used for design evaluation and optimization in many manufacturing industries. Typically, in surrogate modelling polynomial regression models of second-order are used. To model higher order variations, the order of polynomials must be raised. However, in order to reduce the problem of overfitting, the polynomial model must contain only the most desirable basis functions. This problem is commonly addressed by using the standard subset selection techniques implemented also in popular statistical software packages requiring user to predefine a number of parameters, e.g., the maximal order of the model (basis functions of which are further used in subset selection), the significance threshold for basis function inclusion/deletion a.o. However, in many cases the necessary values of the parameters are unknown and need to be guessed, resulting in a non-trivial (and potentially long) trial and error process. We consider a different approach than the subset selection – letting the regression model induction method itself construct the basis functions necessary for creating the model without restricting oneself to the basis functions of a predefined full model. The needed basis functions are adaptively constructed using heuristic search.