Navigating kernel selections in kernel-based methods: The issues and possible solutions
Kehinde Sikirulai Oyetunde, Rhea Patricia Liem · AIAA SCITECH 2022 Forum · 2022
View Video Presentation: https://doi.org/10.2514/6.2022-0507.vid Kernel-based methods are employed in complex cases where the patterns are perceived to be nonlinear. These methods capture the nonlinearity in data and project such data into spaces where linear methods can be used. Kriging is one of the commonly used kernel-based methods. In kriging, kernel selection is a non-trivial issue. Poor selection of kernels could lead to inaccurate approximations especially in complex problems. Hence, it is important to select an appropriate kernel for specific applications. In this paper, we investigate the intuitions behind the behaviours of four main kernels that are commonly used within the engineering community. We then propose two possible methods to automate kernel selections and complete model training for different tasks. The results from benchmarking the proposed methods with both algebraic and real-world problems are presented. The models trained with the proposed approaches are shown to perform better than other models in most of the benchmark cases. An instance is seen in the significant improvement of the model accuracy on the axial transonic problem with NASA rotor 37 as its datum shape. The overall results are promising and demonstrate the potentials of these methods to automate kernel selections in kernel-based methods such as kriging.