I have used a bunch of methods in consulting work. I used Noesis Optimus to drive multi-disciplinary optimization in the late 90's and early 00's.
The approach is very powerful. One automates the design chain, then selects the execution method which can be optimization, DoE or simple Table Lookup picking the best available option.
For what it is worth. My take on the matter is that there is no such thing as a best method very much for the same reason as there are no best pants. It is all about fit and perhaps, also - beauty.
A, say, 3LFF DoE may work beautifully for one problem and completely fail for the next. Some methods may be more robust in arriving at an end result than other, e.g. Genetic Algorithms (GA) seem to be quite robust but a Response Surface from a DoE is hard to beat as one tend to want something else once an optimum is found while a Gradient optimizer cannot be beat when the optimization target is clear cut and the design space is smooth.
To demonstrate the above, see the enclosed paper where a train wheel was optimized with respect to four continuous geometric variables and the outputs stress, radiated sound power and weight are considered. Stress needed to stay below a certain value and, hence, was a simple on/off selector. The end result showed there to be a relation between noise and mass which produced some dB/kg curves in which any point on the curve(s) is optimum. For obvious reasons, a GA would have great trouble in identifying such a relation as beauty (noise, weight or a ratio thereof) then lies in the eye of the beholder.
My simplistic take was to move from simple to advanced and use GA as a last resort when all else failed. That said, starting with a Taguchi DoE screening tends to focus the oprimization by removing second order design variables. A DoE may work for the local design space around the optimum identified by the GA.
So, all in all - for the general situation - you need all of them. The best mix tends to be problem dependent. The best end result tends to be arrived at with the mix that allows the most design iterations in a given time period (as a consultant, you always work against the clock). For a selected problem, there will be a best method choice - but then only as long as the design space does not greatly change .
I highly recommend the approach - it is a game changer as it makes the analyst very much more productive - through automation as well as moving faster in on target by use of the appropriate tool. So much so that most people do not believe me when I tell them about it.
Some more ramblings of mine on the subject can be found here. There is a story in there where not even delivering the 100 best designs was of any practical use. So, how well you do things - sometimes, it does not help as it cannot be produced. https://qringtech.com/learnmore/why-simulate-measure-correlate-automate/
Hope this helps
Claes
Article Multi-disciplinary optimization of railway wheels
As you are into optimization. Perhaps this shape optimization method can be of interest? Article 2005-01-2342 NVH optimization of truck cab floor panel embos...