A Study on Optimal Retraining of Artificial Neural Network-aided Aimed Multilevel Sampling Optimization Method

This paper discusses a recently proposed artificial neural network-aided aimed multilevel sampling optimization method. The method is particularly effective in the optimization of engineering problems where the emphasis is on minimizing the number of evaluations of the objective function, such as structural damage detection using FEM model updating. The proposed method uses sequential sampling to train a machine learning model, which is used to identify the optimal sample. Depending on the complexity of the analyzed problem, a certain number of samples are required to train the model so that the optimization process does not get stuck in local minima. The improvement of the proposed method lies in finding a suitable structure of the training dataset. The emphasis is on maintaining a balance between the complexity of the training data and thus the generality of the surrogate model on the one hand, and the localization of the training data due to fast targeting on the other hand. A suitably adjusted training set should contain a combination of a certain amount of general data together with localized data. This ratio changes gradually during the optimization process. A study of improved method is performed on the problem of bridge structure damage detection.