Global Approximation: Performance Comparison of Different Methods, With an Application to Road Vehicle System Engineering

Abstract

Abstract Optimisation of complex mechanical systems has often to be performed by resorting to global approximation. In usual global approximation practice, the original mathematical model is substituted by another mathematical model which gives approximately the same relationships between design variables and performance indexes. This is made to ensure much faster simulations which are of crucial importance to find optimal solutions. In this paper the performances of four global approximation methods (Neural Networks, Kriging, Quadratic Approximation, Linear Interpolation) are compared, with reference to an actual optimal design problem. The performances of a road vehicle suspension system are optimised by varying the system’s design variables. The Pareto-optimal set is derived symbolically. The performances of the different approximation methods taken into consideration are assessed by comparing the numerical- and the analytical-Pareto-optimal results. It is found that Neural Networks obtain the best accuracy.