Spurious Correlations and Premature Convergence in Genetic Algorithms

Abstract What distinguishes genetic algorithms (GA) from other search methods is their inherent exploitive sampling ability known as implicit parallelism. We argue, however, that this exploitive behavior makes GAs sensitive to spurious correlations between schemata that contribute to performance and schemata that are parasitic. If not combatted, this can lead to premature convergence. Among crossover operators, some are more disruptive than others, and traditional arguments have held that less disruption is better for implicit parallelism. To explore this issue we examine the behavior of two crossover operators, two-point and uniform crossover, on a problem contrived to contain specific spurious correlations. The more disruptive operator, uniform crossover, is more effective at combatting the spurious correlations at the expense of also more disruption of the effective schemata. Elitist selection procedures are shown to be able to ameliorate this somewhat, suggesting that research into ways of dealing with the effects of the inevitable sampling errors may lead to generally more robust algorithms.