Evolution is a family of obsolete black-box [[optimization algorithm]]s. It is a population-based method: many samples are drawn from an initial distribution, and merged, mutated or removed depending on their [[fitness metric]]. Its primary advantage is that it continues to work when the fitness metric is a pathological (expensive-to-compute, discontinuous, non-differentiable, etc) function, unlike [[gradient descent]], which has replaced it in most practical applications.