Significance relations for the benchmarking of meta-heuristic algorithms

The experimental analysis of meta-heuristic algorithm performance is usually based on comparing average performance metric values over a set of algorithm instances. When algorithms getting tight in performance gains, the additional consideration of significance of a metric improvement comes into play. However, from this moment the comparison changes from an absolute to a relative mode. Here the implications of this paradigm shift are investigated. Significance relations are formally established. Based on this, a trade-off between increasing cycle-freeness of the relation and small maximum sets can be identified, allowing for the selection of a proper significance level and resulting ranking of a set of algorithms. The procedure is exemplified on the CEC'05 benchmark of real parameter single objective optimization problems. The significance relation here is based on awarding ranking points for relative performance gains, similar to the Borda count voting method or the Wilcoxon signed rank test. In the particular CEC'05 case, five ranks for algorithm performance can be clearly identified.