Constrained Design of Experiments for Data-Driven Models
Fabian Schneider, Max Schüssler, Ralph J. Hellmig, Oliver Nelles · 2022
The quality of data-driven models depends significantly on the data distribution in the input space. In this work, design of experiments (DoE) methods for constrained input spaces are discussed. An approach based on an Latin hypercube (LH) design is introduced to deal with strongly constrained input spaces. For the unconstrained case, where the input space is a hypercube, different design of experiments methods have been developed. The dominating state-of-the-art methods are Sobol sequences and Latin hypercubes. Instead of optimizing complete LH designs, the proposed strategy is to incrementally construct an LH design. Every new sample is selected by a distance-based metric. The presented method is then applied in two test cases and compared to a method based on a Sobol sequence. Here, an initial design is created by a Sobol sequence and every sample is removed that violates the constraints. The design qualities are measured by the resulting model accuracies of data-driven models. A function generator is applied to create synthetic data sets to train and evaluate local linear model networks.