Fitness noise and localization errors of the optimum in general quadratic fitness models

Evolutionary algorithms are generally believed to perform well in the presence of noise. Thus, they are often used for optimization in noisy environments. It comes as a surprise that hardly more than a handful of studies has dealt with the question of just how well they are doing and what can be done to improve their performance. The present paper presents empirical results regarding the behavior of genetic algorithms and evolution strategies in the presence of fitness noise for a range of objective functions. Bounds for the mean residual location error and the stationary fitness error or optimization of general quadratic fitness models with evolution strategies are derived and compared with measurements.