Is SVEM appropriate for a 6 factor (continuous) optimization study?

[#1]

In last discovery summit there was a great presentation where SVEM was succesfully used for process modeling with a limited #runs; would this also work for buiding a 6 factor RSM starting from a 17-run DSD? In another presentation a 31-run OMARS was proposed for this 6-facor RSM, is SVEM a good alternative? 

[#2 · Accepted solution]

Hi Frank,

I think you are asking if a 17-run DSD modelled with SVEM is a good alternative to a 31-run OMARS in order to estimate the 6-factor RSM model. 

I can say with certainty that the 31-run design will give you more information. SVEM is very useful. But it is not magic. More runs will always be more information, SVEM or no SVEM.

You can certainly use SVEM with the 6-factor RSM model on a 6-factor, 17-run DSD. And that might enable you to meet your objectives.

I can't possibly say whether your case requires the 31-run OMARS or if the 17-run design would give you a good enough model for your objectives. It depends on too many other things. A lot of these will be unknown.

Beware that the 6-factor RSM model has 28 parameters (I think!). That is a lot to estimate with 17 runs! How good your model is will depend on to what extent effect sparsity is true in your case.

And there are correlations between the 2nd-order effects in the 6-factor, 17-ruin DSD (and in the 31-run OMARS). So that will limit the precision of estimating these effects.

I suggest that you read the paper by Lemkus et al if you want to really understand SVEM, its benefits, and limitations.

I would also suggest that you don't throw away the idea of sequential DOE just because SVEM is exciting. You could run the 17-run DSD. Model it with SVEM and other approaches. If that gives you what you need, then great. If it does not give you what you need, then augment.

I hope that helps,

Phil

[#3]

Hi Phil, thanks for input! We now have quite some DOE choices and indeed, as you claim, it depends on a number of criteria to make the rignt choice just like it was in the past for chosing the right fractional classical design. Experienced users can quickly make the right selection however for non-experienced users it's not easy to chose between categories i.e. Classical - DSD - OMARS - SVEM - Optimal... A flowchart for giving a direction to the right choice woud be great and maybe a topic to present on a next summit, what do you think?