Performance Analysis of Hybrid Sampling and Meta-heuristics

This paper investigates the effect of hybridising sampling algorithms with population-based meta-heuristics. Recent literature has shown that alternatives to the traditionally used pseudo-random number generators to generate the initial population of meta-heuristics can improve performance. However, most studies focus on sample sizes that are limited to the size of the initial populations. In contrast, this paper studies the effect of extended random initialisation, which uses relatively large samples and then initialises the meta-heuristics from the points in the sample with the best-found fitness values. A portfolio of three meta-heuristics, four sampling algorithms and three different sampling budgets are analysed from the fixed budget perspective on the BBOB benchmark suite. Statistical analysis of the results shows that the hybrid algorithms converge to better solutions than their non-hybrid counterparts. The results further indicate that large sample sizes can be used to generate landscape analysis features, ensuring reliable approximations of the investigated functions' properties without lessening the meta-heuristics' performance.