Surrogate assisted Cooperative Differential Evolution using Hyperparameter Transfer Gaussian Process Regression

In order to reduce the computational complexity of Gaussian process regression models when solving the sub-problem in large-scale optimization, this study considers the historical Gaussian process regression models generated during the sub-problem optimization iteration as the source models, and treats the Gaussian process regression model that needs to be established in the current optimization iteration as the target model. The similarity between the source and target models is measured according to their sample point distributions, and the hyperparameters of the source models are transferred to the hyperparameters of the target model, aiming to avoid the computational burden of training the hyperparameters of the target model. The proposed GPR model is adopted in the surrogate assisted evolutionary algorithm as the sub-optimizer to solve expensive large-scale optimization problems. Results from 12 constrained benchmark functions with dimensions up to 1000 dimensions demonstrate that the proposed hyperparameter transfer method can significantly reduce the running time of the algorithm without significantly compromising the quality of the optimization solutions. With appropriate parameter settings, the proposed hyperparameter transfer method can be effectively applied in cooperative coevolutionary algorithms assisted by surrogate models.