Genetic algorithms with noisy fitness

The convergence properties of genetic algorithms with noisy fitness information are studied here. In the proposed scheme, hypothesis testing methods are used to compare sample fitness values. The “best” individual of each generation is kept and a greater-than-zero mutation rate is used so that every individual will be generated with positive probability in each generation. The convergence criterion is different from the frequently-used uniform population criterion; instead, the sequence of the “best” individual in each generation is considered, and the algorithm is regarded as convergent if the sequence of the “best” individuals converges with probability one to a point with optimal average fitness.