Answer: it doesn't. Genetic algorithms and the like are pretty much all terrible. They're ways of approximating your gradient, and fall to the curse of dimensi

Answer: it doesn't. Genetic algorithms and the like are pretty much all terrible. They're ways of approximating your gradient, and fall to the curse of dimensionality. The only reason Uber and open.ai published their papers on evolutionary strategies (something pretty different from what people think of as genetic algorithms) is that current policy gradient methods are really bad as well, allowing what is effectively random search to do well. It's kinda like how Bayesian hyperparameter optimization is pretty terrible and 2x random search almost always beats it easily.