Input Selection for Local Model Approaches

This chapter discusses the topic of input selection in the context of local model networks. It utilizes a key advantage of this model architecture analyzed in Chap. 14 for the issue of selecting the relevant inputs to the model. Due to the curse of dimensionality, it is obvious that it can be beneficial to discard inputs with no or little relevance for the overall model performance and robustness. For local model networks, this can be done individually for the premises and the consequents. This is a major advantage compared to other model architectures and will be detailed in this chapter. Various strategies to exploit it are proposed and applied to toy and real-world examples. In addition, the visualization method “partial dependence plot” is introduced, which represents an excellent tool helping to visualize the main characteristics of high-dimensional mappings and should become more familiar in general and particularly in engineering. Also, the general issue of input relevance is addressed, which also can help in a data mining context.