Thresholding-a selection operator for noisy ES

The starting point for the analysis and experiments presented in this paper is a simplified elevator control problem, called `S-ring´. As in many other real-world optimization problems, the exact fitness function evaluation is disturbed by noise. Evolution strategies (ES) can generally cope with noisy fitness function values. It has been proposed that the `plus´-strategy can find better solutions by keeping over-valued function values, thus preventing inferior offspring with fitness inflated by noise from being accepted. The `plus´-strategy builds an implicit barrier around the current best population. We propose to make this barrier building process explicit and to employ a threshold value τ to be used in a selection operator for noisy fitness functions. `Thresholding´ accepts a new individual if its apparent fitness is better than that of the parent by at least the margin τ. First analytical investigations and empirical results from tests on the sphere-model and `S-ring´ are presented