Replacement Strategies in Steady State Genetic Algorithms: Dynamic Environments

Recent years have seen increasing numbers of applications of Evolutionary Algorithms to non-stationary environments such as on-line process control. Studies have indicated that Genetic Algorithms using "Steady State" models demonstrate a greater ability to track moving optima than those using "Generational" models, however implementing the former requires an additional choice of which members of the current population should be replaced by new offspring.In this paper a number of selection and replacement strategies are compared for use in Steady State Genetic Algorithms working as function optimisers in dynamic environments. In addition to an algorithm with fixed mutation rates, the strategies are also compared in algorithms employing Cobb\u27s Hypermutation method for tracking environmental changes. On-line and off-line metrics are used for comparison, which correspond to different types of real-world applications.In both cases it is shown that algorithms employing some kind of elitism outperform those that do not, which is related to previous studies on stationary environments. An investigation is made of various methods of implementing elitism, including an implicit method, "conservative" selection. It is shown that the latter, in addition to being computationally simpler, produces significantly better results on the problems used, and reasons are given for this behaviour