On the Investigation of the Kernels in Gaussian Process Regression for the RCS Prediction of a Conducting Object

The Gaussian Progress Regression (GPR) is one of the most popular techniques in supervised machine learning regression algorithms. Recently it has been started to use for the prediction of the RCS of the conducting objects since it can help to decrease the number of computation points or widen the sample size of the frequency interval under investigation. It is known that the kernel of the GPR has a significant impact on the error behaviour of the algorithm. It has been shown in the literature that composing a priori knowledge of the physical solution of the problem with the kernel may improve the accuracy of the algorithm. In this study, we perform some numerical research on the effect of the kernels on the accuracy of the predicted radar cross section (RCS) of the NASA almond calculated by the classical method of moments (MoM). In numerical results, we show that different kernels result in different root mean square error (RMSE) and in contrast to the published results the best model consists of the product of non-periodic kernels. We also show that with this model the data regarding some frequencies that are not included in the training dataset can be predicted very accurately thanks to the GPR being an effective method in sparse regression.