Case studies in applying fitness distributions in evolutionaryalgorithms. II. Comparing the improvements from crossover and Gaussianmutation on simple neural networks

摘要

Previous efforts in applying fitness distributions of Gaussianmutation for optimizing simple neural networks in the XOR problem areextended by conducting a similar analysis for three types of crossoveroperators. One-point, two-point and uniform crossover are applied to thebest-evolved neural networks at each generation in an evolutionarytrial. The maximum expected improvement under Gaussian mutation with asingle fixed standard deviation is then compared to that which can beobtained using crossover. The results indicate that the benefits of eachtype of crossover varies as a function of the generation number.Furthermore, these fitness profiles are notably similar (i.e., there islittle functional difference between the various crossover operators).This does not support a building block hypothesis for explaining thegains that can be made via recombination. The results indicate caseswhere mutation alone can outperform recombination and vice versa