A new approach to experimental design for function approximation and classification: The twilight method

Modeling real-world systems plays an essential role in system analysis, and contributes to a better understanding of their behavior and performance. Classification, optimization, controls, and pattern recognition problems heavily rely on modeling techniques. From a particular viewpoint, models could be categorized into three classes: white box, black box, and gray box models. The present study focuses on black box modeling. The satisfactory performance of a black box model depends on its structure and data used for calibration of the model. Although the number of data points is an important factor for improving the richness of the dataset, there are limitations on increasing the number of data points in real problems. For instance, gathering data from many real-life systems (e.g. industrial ones) imposes spending a huge amount of time and money. In this study, we discuss a method which yields richer datasets for a known number of data, in comparison to some other conventional experimental design methods. In the proposed algorithm, after extracting some data points by the factorial design method, the remaining data points are extracted based on the analysis of the available data and the characteristics of the model. The results illustrate the superior efficiency of the proposed method.