Influence of Selection Criterion on the RBF Topology Selection for Crashworthiness Optimization

Meta-models are frequently used to offset high computational cost of crashworthiness optimization problems. Radial basis function based meta-models are gaining popularity among various meta-modeling techniques due to their ability to approximate non-linear responses with relatively low fitting cost. However, the performance of RBF networks is very sensitive to the choice of topology. In this paper, the influence of three selection criteria namely, PRESS, pointwise PRESS error ratio, and estimated variance of noise, over network topology is studied. The results are demonstrated for a few analytical functions and a crashworthiness simulation of a full NHTSA vehicle problem. The results showed that the PRESS-based method was the most reliable method to select network topology.