Framsticks

Blindly chasing after fitness is not the best strategy for optimization of hard problems, as it usually leads to premature convergence and getting stuck in low-quality local optima. Several techniques such as niching or quality-diversity algorithms have been established that aim to alleviate the selective pressure present in evolutionary algorithms and to allow for greater exploration. Yet another group of methods which can be used for that purpose are fitness diversity methods. In this work we compare the standard single-population evolution against three fitness diversity methods: fitness uniform selection scheme (FUSS), fitness uniform deletion scheme (FUDS), and convection selection (ConvSel). We compare these methods on both mathematical and evolutionary design benchmarks over multiple parametrizations. We find that given the same computation time, fitness diversity methods regularly surpass the performance of the standard single-population evolutionary algorithm.