Population Diversity in Genetic Algorithms via Clustering

Master's research (Applied Evolutionary Computation, PPGCC/UFMA) on maintaining population diversity in Genetic Algorithms. The idea is to periodically cluster the population, take the fittest individual from each cluster, and use them to replace the worst ones - forcing representatives of distinct search-space regions to survive. Three variants were compared against a standard GA (K-Means, hierarchical with Ward linkage, and DBSCAN), across 30 independent runs per combination. On the multimodal Schwefel function, clustering cuts mean fitness from 237.2 to 134.8 (-43%), and Tukey's test confirms all three variants significantly outperform the standard GA - yet are statistically indistinguishable from each other, meaning the gain comes from clustering itself rather than any specific method. The work then evolves into a BAT-PSO hybrid with probabilistic operator alternation and into constraint handling via dynamic penalty, where the result is categorical: on a quadratic problem with equality constraints, the 7 algorithms without explicit handling find no feasible solution at all, while the Penalized GA converges to 49.3.