Non-stationary Subtasks Can Improve Diversity in Stationary Tasks

Low diversity in a genetic algorithm (GA) can cause the search to become stagnant upon reaching a local optimum. To some extent, non-stationary tasks avoid this problem, which would be a desirable feature of GA for stationary tasks as well. With this in mind, we show that several methods of introducing artificial non-stationary elements help to promote diversity in a GA while working on an inherently stationary task. By analyzing online and offline diversity and fitness measures, we show that it is possible to improve overall performance through this technique, and that some measures intuitively related to performance are misleading.