Category: Genetic Algorithms Title: Randomness and GA Performance, Revisited

Previous studies by the authors have indicated that pseudo-random number generator (PRNG) quality has little e ect on the performance of a simple genetic algorithm (GA). In this paper we examine this subject further, in the context of what we call the \granularity hypothesis. We detail a set of PRNG quality tests tailored speci cally to the uses of randomness in a simple GA. We explain the application of detailed statistical analysis to the results of nearly ten-thousand individual GA runs, for large and small populations, over an eleven function GA test suite. We conclude that, although there is no evidence to support the notion that higher quality PRNGs cause better GA performance than lessor quality PRNGs, there is statistical evidence that certain PRNGs can provide improved GA performance.