Reducing the sampling variance when searching for robust solutions
Jürgen Branke · 2001
For real world problems it is often not sufficient to find solutions of high quality, but the solutions should also be robust. By robust it is meant that possible deviations from the solution should be tolerated, still yielding a good expected performance. One way to reach this goal is to evaluate each individual several times under a number of different scenarios, taking the average performance as fitness. But although this method is effective, it requires significant computational power. In this paper, we continue some previous work aimed at minimizing the search effort while still providing the desired robustness. In particular, we examine the effectiveness of de-randomizing the sampling mechanism using variance reduction methods, and the question whether the same scenarios should be used for all individuals in the population or not. As will be shown, a significant performance gain can be obtained by taking these ideas into account, without any additional computational cost.