Applying evolutionary algorithms to problems with noisy, time-consuming fitness functions

For many "real world" applications of evolutionary computation, the fitness function is obscured by random noise. This interferes with the evaluation and selection process and adversely affects the performance of the algorithm. We present a study of noise compensation techniques designed to better counteract the negative effects of noise. We introduce algorithms that vary the number of samples used per candidate based on the amount of noise present at that point in the search space. Results show that these algorithms are significantly better than the traditional technique used by the optimisation community and that noise compensation is indeed a difficult task that warrants further investigation.