Comparative Evaluations of Evolutionary Computation with Elite Obtained in Reduced Dimensional Spaces

We propose an elite synthesis optimization strategy for accelerating evolutionary computation (EC) searches using elites obtained from a lower dimensional space. The method projects individuals onto n one-dimensional spaces corresponding to each of the n searching parameter axes, approximates each landscape using Lagrange polynomial interpolation or power function least squares approximation, finds the best coordinate for the approximated shape, obtains the elite by combining the best n found coordinates, and uses the elite for the next generation of the EC. The advantage of this method is that the elite may be easily obtained thanks to their projection onto each one-dimensional space and that there is a higher possibility that the elite will be located near the global optimum. We conduct experimental tests to compare our proposed approaches with previous acceleration approaches using differential evolution and ten benchmark functions. The results demonstrate that the proposed method accelerates EC convergence significantly, especially in early generations.