nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Definitive Screening Design for optimization

Ella

Hi Community members,

I searched about definitive screening design and read a lot about this topic in JMP.

I have learned that Definitive Screening design is used to determine effective main, second order power and 2 factor interactions in the model.

I have 10 parameters and my mail goal is to make optimization. At the fist step using RSM or custom design with model including main effects+second order power+2 factor interactions effects is very costly. That's why i think that at the first stage, i have to find the significant parameters in the model. So i should apply definite screening design. After i ll find the effective parameters by using fit definitive screening platform, i will construct a OLS model with effective parameters. When i found the OLS model, do i use this model safely for optimization?

If you think that yes i can use it for optimization, why the aim of DSD is to determine the effective factors, it is also good for optimization, right?

Kind regards.

Accepted answers

All comments

P_Bartell

There is no way to answer your 'safely' question with any degree of confidence before seeing any data, using your inherent process knowledge, and understanding the system under study. Optimization type problems can have many varying pathways to ultimate problem resolution. In general I like your thinking wrt to starting with a definitive screening design. My suggestion is to go through that phase of the problem solving process. Pause and reflect on what you've learned to that point. Let those reflections guide your next steps. Maybe share your DSD experiment, with an explanation of the problem and system, the data, your analysis, and your conclusions here...and some of us may be able to offer our thoughts.

Ella

Thanks Bartell for your suggestion.

one experiment takes 3 hours. DSD offers min Numbers of run 28 for 10 factors which is also costly. That means that I have no chance to make wrong decision.

As I stated below, I am near to DSD acc to my search, but my other opinion is that choosing d or I optimal custom design with main and 2order powers. Then acc to the results, I can augment it with 2 factor interraction.

But i am not sure which is more effective way. I am near to DSD but have doubts because I haven’t use it before and it is not suggested to use when there is more than n/2 active factors and I don’t know whether there is more than n/2 active factors or not.

Kind regards.

statman

Ella, Echoing @P_Bartell , I have some additional thoughts regarding your questions, but do not have any idea what your situation is (and most, if not all advice would be situation dependent). First question: What do you mean by optimization? This is always an interesting question as it is counter to the idea of continuous improvement.

Is your situation multivariate? What are your response variables? Do they clearly describe the phenomena of interest? It is challenging to optimize all Y's simultaneously. Usually improvement requires compromise.

You have identified 10 design factors, what about the rest of the variables (noise or nuisance variables)? It doesn't do any good to "map the base of the mountain". None of the strategies you allude to contain strategies to handle noise (ambient conditions, raw material variability, measurement uncertainty, etc.).

It is premature to generalize about the most efficient or effective approach given what you have told us about your situation. It seems to make sense (in many situations) to identify a reasonable first order model, determine if this model repeats over changing conditions and then augment the space. Having a reasonable model does not mean it has to be complex. In fact, you should strive for a model that is simple and is robust to noise (perhaps this is what optimization means).

Ella

Hi statman,
I am trying to find optimum factor levels to reach our responses’ targets which are minimize “sound power “while stating “flow rate in between some range.
I have 10 continuous factor, all of which are controllable. All of the factors effect responds are these 10 factors, there is no noise or nuisance factors.
Since I have engineering knowledge that factors probably have curvature effect on responds, I planned to start with model having main and 2nd power effect.

By the way, to select whether to start the study with custom design or DSD, i made both type of designs for 10 factors and simulate the responds and try to estimate OLS models.I did it 3 times and all the time ii observed that starting with using custom design predicts the same model better. That’s why I choose custom design to start.

Kind regards

Mark_Bailey

I only want to add that the DSD is intended for a screening situation. It relies on common, key principles of screening, not optimization. One of these key principles is 'sparsity of effects.' The designs are meant to be highly economical in return for the ability to determine significance of a few of the factors, not optimization.

So, if you expect most of the 10 factors to be significant, the DSD is not a good choice. It provides insufficient data to find the active factors. On the other hand, if this study is a true screening case (less than half of the 10 are ultimately significant), it is a good choice.

Ella

Is not it effective when active Numbers of factors is less than n/2? Or is it when number of factors/2?

ı of not knot whether i mis sthg or not but i observed in my simulation study that DSD design and fit DSD platform can not predict all of the active factors. May be i have to increase the number of experiments but my final goal is to maske optimization, i felt my self safer starting with custom design.

I hope I will still feel safer after experiments are done

kind regards

P_Bartell

@Ella please take note of @Mark_Bailey 's reply. You state in one of your replies that you KNOW all 10 factors have an effect on the response. This is not a screening situation. I would not be considering a definitive screening design but using optimal DOE tactics.

Ella

Hi @P_Bartell and @Mark_Bailey ,

I need to correct that I have 10 factors but do not know exactly all of the have an significant effect on response.

Sure, if i knew it, i agree that DSD is not suitable for this case.

To decide which design is suitable for me, I constructed a case study:

I assumed that my real response model is that: Y= X1+ 2*X2 + 3*X3^2 + 4*X4*X5

First I used DSD, and run 25 experiments, and simulate the rexponse acc to the above function.

Then I use Fit definitive screening with unclick the heredity options.

Combined model parameters are X1,X2 and X7.

At the second step, i used Custome Design for 10 factors with model main effects and 2nd powers and again resimulated the response acc to the above function. (number of runs=28)

Then i constructed a OLS model. It predicts effective parameters: X1,X2,X3^2 and X10.

Acc to my case study Custom design gave me closer results. But i am still not sure whether i miss sthg in DSD. Because when we look at both designs' purpose of use, DSD seems more approprriate for early stages of the experiments. But acc. to case study, even in early stages, Custome design works well. If i miss stgh in DSD, i am glad to learn it.

(by the why, when i augmented Custom Design by including 2 factor interraction effects with 27 more runs, OLS exactly find the assumed model.)

If you were me, what design do you choose?

Kind regards.

Definitive Screening Design _case study.jmp

Augmented Design.jmp

Custom Design.jmp