Since learning about statistical modeling, particularly with regard to Bayesian techniques, my interest in optimization has vanished. In my opinion, concentrate on more promising fields.
When there are no good differentiable surrogate losses available and the loss landscape is not differentiable, they can be employed. Simulated annealing is also possible. Nonetheless, simulated annealing and genetic algorithms explore in different ways, so certain problems or parameterizations will do better with one than the other.