On the Prediction of the Solution Quality in Noisy Optimization

Noise is a common problem encountered in real-world optimization. Although it is folklore that evolution strategies perform well in the presence of noise, even their performance is degraded. One effect on which we will focus in this paper is the reaching of a steady state that deviates from the actual optimal solution. The quality gain is a local progress measure, describing the expected one-generation change of the fitness of the population. It can be used to derive evolution criteria and steady state conditions which can be utilized as a starting point to determine the final fitness error, i.e. the expected difference between the actual optimal fitness value and that of the steady state. We will demonstrate the approach by determining the final solution quality for two fitness functions.