Gaussian Process Regression With Maximizing the Composite Conditional Likelihood

Gaussian process regression (GPR) has an outstanding nonlinear fitting ability, and its uncertainty predictions can deliver the confidence level of the estimations, which is well adapted to deal with complex industrial processes. However, disturbances and noises in outputs might lead to mispredictions for new samples. In this article, a method using the modified likelihood is proposed to deal with the output corrupted by noises, which aims to achieve a more stable and reliable generative model. Furthermore, the proposed method is applied to a simulation experiment and an actual hydrocracking process to model the relationship between the input variables and the light ends, and the experimental results demonstrate the efficiency of the proposed method.