Statistical inference on relative performance of genetic algroithms: first results on the significance of knowledge, direction, and representation in three small network structures
Greg J. Langevin · 1997
Genetic Alogorithms are black box function optimizers. What goes on in the black box is presently unknown. Theory and design principles are sorely missing in this field, but most agree the algorithms follow the principles of natural selection. This study attempts to understand what goes on in the black box beyond a simple understanding of its mechanics. The approach is to design a hybrid and experimental GA whose sole purpose to report what it is doing, and where it is in the search space. A central question in GA research is the consistency of GA operators given small changes in a search (problem) space. To determine consistency, three small Shortest Route networks are processed with two sets of domain knowledge, two directions, and two binary (Arc and Intron) representations. The operators did perform consistently, (that is, there are no confounding effects in the data that destroy the ability to interpret main effects), so performance data is subjected to the Median Test, a non-parametric statistical technique. There is a statistically significant difference in GA performance given different sets of domain knowledge. GA performance is marginally affected by direction and the Intron representation outperformed the Arc representation in eleven of twelve trials with four trials being statistically significant at alpha =.05.