The Influence of Population and Memory Sizes on the Evolutionary Algorithm's Performance for Dynamic Environments

Authors

Abstract

Usually Evolutionary Algorithms keep the size of the population fixed. Nevertheless, in Evolutionary Algorithms dealing with stationary problems some work has been done involving the idea of adapting the population\'s size along generations. In the context of dynamic environments, less attention has been devoted to the choice of this parameter. In fact, approaches based on the idea of dividing the main population into two, one that evolves as usual and the other one that plays the role of memory of past good solutions, choose off-line the size of these two populations. Usually memory size is chosen as a small percentage of population size, but this decision can be a strong weakness in algorithms dealing with dynamic environments. Recent work which makes possible changing the size of the population and memory during a run proved that the performance of evolutionary algorithm can be considerably improved. In this work we do an experimental study about the importance of this parameter for the algorithm\'s performance. In a first set of experiments the size of each component was kept constant but the relative proportion was changed; in a second set of runs we used an algorithm where the size of the two populations could change and compare it with the other fixed size schemes. Results show that tuning the population and memory sizes is not an easy task and the impact of that choice in the algorithm\'s efficacy is significant. Using an algorithm that dynamically adjusts the population and memory sizes outperforms the previous approach.

Keywords

Evolutionary Algorithms, Dynamic Environments, Populationâ??s size, Memoryâ??s size, Control Parameters

TechReport Number

TR 2008/002