1. Because no evolutionary method is guaranteed to find a optimum. And

2. Because there are convergent methods that are much more reliable - and they also find an optimum.

Clear-cut, wouldn't you say?

The aim of hyperparameters optimization in machine learning is to find the hyperparameters of a given machine learning algorithm that return the best performance as measured on a validation set. To achieve this we generally use Bayesian Optimization, Grid Search, Random Search etc. Evolutionary and gradient search methods could also be used. But selection of hyperparameters optimization techniques depend on your requirements.

In general Bayesian optimization yields better performance than random and grid search. But it is time consuming, it could take a day on large dataset. Bayesian optimization is an informed learning approach which learns from past which eventually ends up giving good results, however random and grid search are noninformed learning which spends most of the learning time at bad hyperparameters.

Evolutionary algorithms are population based learning approach. Like Bayesian model it is also informed learning but it is more time consuming than Bayesian optimization due to involvement of multiple steps and at each step training and validation are involved to calculate the objective. For this reason it is still not preferred approach for hyperparameters tuning. Also, evolutionary algorithms require proper tuning of their own parameters. For instance in PSO proper selection of population size, inertia factor, velocity etc. are needed for better convergence.

However, the evolutionary algorithms should be better suited for parallelization?

Dear Roman Liessner Indeed, one population in evolutionary algorithm is independent from another population hence well suited for parallelization. However, Bayesian and some other optimizations can also easily parallelize.