Evaluation of optimization metaheuristics in clustering

Abstract

We have evaluated five metaheuristics of combinatorial optimization applied in clustering by partitions: simulated annealing, tabu search, genetic algorithm, ant colonies and particle swarms, using data tables generated randomly according to some defined parameters. Those techniques were compared to classical methods (k-means and Ward's agglomerative clustering). Sixteen tables were generated (four controlled factors, with two levels each) with normally distributed variables and, for each one, the experiment was repeated 100 times in a multistart procedure. The within-class inertia was used as the criterion to compare the classifications obtained. Best results were obtained for ant colonies, simulated annealing and the genetic algorithm.