MODELING AND OPTIMIZING MULTI-STAGE DESIGN WITH GAUSSIAN PROCESS BASED ON SURROGATE MODEL CHAIN

Continued progress in the surrogate-model-based evaluation for the single-stage has been explored, but multistage has higher dimension and uncertainty. High dimension and low overall data of multi-stage leads to low accuracy of prediction, and cannot characterize the uncertainty of the final prediction performance. We propose a Gaussian Process-based surrogate model chain (GP-SMC) to evaluate the performance of multi-stage. Also, we combine the GP-SMC with the quasi-Newton method (L-BFGS-B), make full use of the gradient information of the GP-SMC to get an optimization solution rapidly. The MAE (Mean Absolute Error) and MRE (Mean Relative Error) and STD (standard deviation) of GP-SMC's predicted value are 10% of the prediction of a single surrogate model, which achieves a significant improvement in prediction accuracy and a significant reduction in uncertainty. Compared with the original optimization results, the average performance is improved by 21.05%. Based on the optimal solution and GP-SMC, the confidence interval of the final performance under the optimal solution is obtained, and the confidence level is 99%. The truth probability of GP-SMC is 91.25% in the test dataset, which is higher than single GP’s 85% truth probability. The technology is used in the case of Hot Rod Rolling, and can also be applied to complex product design with multi-stage.