The LUS method is generally recommended as the overlaid meta-optimizer. The tutorial source code contains suggestions for experimental settings that have been found to work well. It is best if you perform meta-optimization regarding the problems you are ultimately going to use the optimization method for. However, if your fitness function is very expensive to evaluate, then you may try and resort to using benchmark problems as a temporary replacement when meta-optimizing the behavioral parameters of your optimizer—provided you use multiple benchmark problems and the optimization settings are the same as those used in a real-world application. In other words, you should use benchmark problems of similar dimensionality and with a similar number of optimization iterations to what you would use for the actual problem you will ultimately optimize.