# How do I choose which design of experiments method to use?

I'm setting up a computer simulation (which I know changes the design of experiments methods) and I can't really determine how to choose which design of experiments (DOE) method to use. I'm a little overwhelmed with all the choices and options and I am a bit unclear on which ones are best for which purposes.

I have 17 inputs and maybe 5 outputs. I say maybe because I'm not sure what will be interesting at the end. Ideally, I would like to develop a model (response surface) where given the inputs, it predicts the outputs.

But maybe some inputs aren't needed. Or maybe some outputs are not useful. Some of the outputs are independent of others so maybe a single 17 input, 5 output model wouldn't be good and it needs to be split into a few models where some inputs matter and some outputs are coupled. Maybe some inputs aren't important, or maybe some inputs are related -- for example, the ratio of two inputs might determine the response.

Ultimately this will probably be a multi-step process where I need to determine the sensitivity to the inputs for each output, or determine the important relationships between inputs (products, ratios, sums, etc) and where the data can be used to generate one or more models of the responses.

As always, I need to minimize the number of simulations. I'm using the Dakota package and from my reading of the user manual, Monte Carlo methods or Orthogonal Array - Latin Hypercube Sampling may work the best. But what makes one method better than the other for this kind of experiment? Will I be able to determine important inputs, relationships between inputs, and develop models using either one?

• Is the simulation deterministic or stochastic? Commented May 21, 2014 at 3:13
• @neverKnowsBest Deterministic. It's a structural mechanics simulation. So no need to repeat points. Commented May 21, 2014 at 3:29
• I've seen people use JMP to fit gaussian process models to deterministic computer simulations. JMP has an article introduction to space filling designs that may help (latin hypercube sampling is a type of space-filling design). I'm not well acquainted with the nuances there. Usually there's a parameter in the kernel for each factor and when that parameter is small then they throw it out. They also might fit a separate GP model for each output. Commented May 21, 2014 at 3:41
• With 17 inputs it is going to take hundreds of thousand of experiments to attempt a model. My recommendation is to perform a N=20 Plackett-Burman design in order to help identify the most important factors for your 17 inputs and reduce down the number of possible factors. Commented Sep 8, 2023 at 22:56

## 1 Answer

Here are some similar questions with answers, maybe a duplicate:

User @neverKnowsBest have this answer in a comment:

I've seen people use JMP to fit gaussian process models to deterministic computer simulations. JMP had an article introduction to space filling designs that may help (link not longer working) (latin hypercube sampling is a type of space-filling design). I'm not well acquainted with the nuances there. Usually there's a parameter in the kernel for each factor and when that parameter is small then they throw it out. They also might fit a separate GP model for each output.