Abstract
Approximation of costly objective functions by surrogate models is an increasingly popular method in many engineering optimization tasks. Surrogate models can substantially decrease the number of expensive experiments or simulations needed to achieve an optimal or near-optimal solution.
In this paper, a novel surrogate model is presented. Compared to the most of the surrogate models reported in the literature, it has an advantage of explicitly dealing with mixed continuous and discrete variables.
The model use radial basis function networks for continuous and clustering and a generalized linear model for the discrete covariates. The applicability of the model is shown on a benchmark problem, and the model's regression performance is further measured on a dataset from a real-world application.