On the behaviour of evolutionary strategies for problems with varying noise strength

For many real-world applications of evolutionary computation, the fitness function is obscured by random noise which may vary throughout the search space. Previously, we presented algorithms that were significantly better than naive resampling, but found (perhaps counter-intuitively) that for some problems it is better to use a higher resampling rate where the noise strength is lower and vice versa. This paper analyses why this is the case, and explores how the evolutionary process works differently on these problems. We show why it is often the case that using a high resampling rate in high noise regions is ineffective and it is instead better to use these samples in low noise regions. We conclude that when applying a basic evolutionary strategy to this class of problems, it is only better to use higher resampling rates where the noise strength is higher if it is too difficult to reach a good solution without searching in or through the high noise regions.