Structure selection for nonlinear models with mixed discrete and continuous inputs: a comparative study

A comparison of two methods for selecting inputs in nonlinear models with mixed discrete (categorical) and continuous variables is presented. Both methods assume that an initial superset of potential regressors is given along with a set of data. In the first approach, the relevant inputs are selected by a model-free search algorithm using fuzzy clustering to quantize continuous data into subsets. The second approach employs regression trees as an induction algorithm ´wrapped´ within a search method. The results obtained for two simulation examples and one real-world data set show that the fuzzy clustering-based method performs more consistently in selecting the model structure. Moreover, this method is much faster then the wrapper approach.