Accelerating the convergence of evolutionary algorithms by fitness landscape approximation

Abstract

A new algorithm is presented for accelerating the convergence of evolutionary optimization methods through a reduction in the number of fitness function calls. Such a reduction is obtained by 1) creating an approximate model of the fitness landscape using kriging interpolation, and 2) using this model instead of the original fitness function for evaluating some of the next generations. The main interest of the presented approach lies in problems for which the computational costs associated with fitness function evaluation is very high, such as in the case of most engineering design problems. Numerical results presented for a test case show that the reconstruction algorithm can effectively reduces the number of fitness function calls for simple problems as well as for difficult multidimensional ones.