Framsticks

Evolutionary algorithms are among the most successful metaheuristics for hard optimization problems. Nonetheless, there is still much room for improvement of their effectiveness, especially in the multimodal problems, where the algorithms are prone to falling into unsatisfactory local optima. One of the solutions to this problem may be to encourage a broader exploration of the solution space. Motivated by this premise, we compare the evolutionary algorithm without niching, with niching, the novelty search, and the two-criteria optimization (NSGA-II) where the criteria of fitness and diversity are not aggregated. We investigate these methods in the context of automated design of three-dimensional structures, which is one of the hardest optimization problems, often characterized by a rugged fitness landscape arising from a complex genotype to phenotype mapping. In the experiments we optimize 3D structures towards two different goals, height and velocity, using two genetic encodings and three distance measures: two phenetic ones and a genetic one. We demonstrate how different distance measures and diversity promotion mechanisms influence the fitness of the obtained solutions.