The worth of the uniform uniform crossover

Abstract:  Abstract Since its popularization by Syswerda (1989), uniform crossover (UC) has become perhaps the most widely used crossover operator, despite the fact that arguments backing low-disruption operators are as old as the genetic algorithm (GA) itself. We question the usual arguments given in favor of UC, and argue that on problems with even a mild degree of difficulty, disruption is indeed harmful. We conduct experiments on a function which is a simple concatenation of several copies of a non-separable function, which makes it difficult for a simple hill climber. We find that in this setup, UC is outperformed by the standard 2-point crossover fitted with inversion. According to our experiments, on problems where applying a GA makes sense, UC appears to perform worse than the classic 2-point crossover fitted with inversion, and essentially the same as a simple copy-and-mutate operator