LAUSR

Constrained optimization problems (COPs) with multiple computational expensive constraints are commonly encountered in simulation-based engineering designs. During the optimization process, the feasibility analysis of the intermediate solutions depends on the computational simulations will be computationally prohibitive. To relieve this computational burden, an active-learning probabilistic neural network (AL-PNN) classification modeling approach is proposed to build a classifier for quickly analyzing the feasibility of the intermediate solutions. In the proposed AL-PNN approach, an interesting region tracking strategy is developed to locate the regions that may contain part of the constraint boundary. The judge rule of the interesting region is based on whether the predicted class labels of pseudo points are different in subregions, which is generated by dividing the design space with the K-means cluster algorithm. Once the interesting region is located, the newly infill sample point used to update the PNN classification model can be obtained by a distance screening criterion. Seven numerical cases and the design of the rocket interstage section are used to demonstrate the performances of the proposed approach. The results illustrate that the proposed AL-PNN approach can provide more accurate classification results than the compared four state-of-the-art algorithms.