cancel
Showing results for 
Show  only  | Search instead for 
Did you mean: 
Browse apps to extend the software in the new JMP Marketplace
Choose Language Hide Translation Bar
MetaLizard62080
Level III

DOE - Augment Design to Focus on More Narrow Range

Hi,

 

Question: Is there a way, to add runs to a DOE (Augment Design), which adds a focused range of a factor. This way you can evaluate a wider range, and also focus on the expected optimal range based on prior knowledge.

 

Ex. 25 runs with a factor(A) of 0-100 + 5 runs with factor(A) of 0-30

 

Background:

I have created a definitive screening design which ranges my factors very wide. I have prior knowledge (Both in-house, and through literature) showing that the range that this factor is usually effective within a specific range (Arbitrarily ~20). That said, DOEs can surprise you when prior knowledge is based on OFAT approaches which is why I have opted for a wider range.

 

My DOE ranges this parameter from 0-100 currently. Looking over the DOE, there are three levels it has generated (0,50,100). My concern is that because the 2 upper levels (50, and 100) are potentially both far above the optimal point, I will not get meaningful data. I am debating adding more runs and forcing the DOE to also additionally evaluate a focused, narrower space around the target 20 mark. 

4 REPLIES 4
Victor_G
Super User

Re: DOE - Augment Design to Focus on More Narrow Range

Hi @MetaLizard62080,

 

Yes, you can change the factor ranges when augmenting a design : Augment Design Window (jmp.com)

By default, JMP will add the range values for your factors based on your datatable and the column property "Coding" of each factor, but you still have the possibility to extend or restrict the factors' ranges. 

There are also other changes you can do, about constraints between factors, terms in the model (add more complexity for example), change the optimality criterion, and you can choose which type of augmentation you would like to do.

 

I hope this (short) answer will help you,

Victor GUILLER
L'Oréal Data & Analytics

"It is not unusual for a well-designed experiment to analyze itself" (Box, Hunter and Hunter)
MetaLizard62080
Level III

Re: DOE - Augment Design to Focus on More Narrow Range

To follow up. I tried to use the augment design feature, but rather than change and restrict the range, (I input 0-30) it just added more runs at 0, I assume because the restriction of 0-30 was within the original 0-100 range.

 

Is there a way to force the model to add 0, 15, 30s to the runs that were 0, 50, 100?

Victor_G
Super User

Re: DOE - Augment Design to Focus on More Narrow Range

If you would like to have new runs between 0 and 30 for one factor, then you'll have to change the factor range, else JMP will augment the design based on the original factor range (0 - 100).

The new runs and the values also depend on the model you have assumed : if you want to have points at 0, 15 and 30, you need to have a factor range of 0-30 and augment the design with centrepoints or augment the design by adding a quadratic effect for this factor in the model, depending on your model needs and objectives.
Also note that depending on the model and the optimality criterion chosen for the augmentation, the repartition of points at the borders and in the centre of the experimental space may differ : Optimality Criteria

You may have to switch to an I-optimal design (besides adding quadratic effect in the model) in the augmentation to facilitate the generation of points at middle level 15 (if the range is 0-30).

Victor GUILLER
L'Oréal Data & Analytics

"It is not unusual for a well-designed experiment to analyze itself" (Box, Hunter and Hunter)
statman
Super User

Re: DOE - Augment Design to Focus on More Narrow Range

I have the following thoughts:

1. There is no singular methodology or experiment that will be useful in all cases.

2. I believe in iteration.  That is, the first experiment will help to plan the next experiment.  To this end, start building your understanding with first order models and augment those in subsequent experiments.  In other words, don't try to get everything in the first experiment.  KISS (Keep it Simple and Sequential).

3. Your approach to be bold and increase the inference space while investigating second order effects seems reasonable.  Run this first, then design a follow-up experiment.  Remember, you are not trying to "pick a winner", but understand causal relationships.  Determining optimum levels will occur in subsequent iterations.

4. I would also suggest including noise in your experiments.

"All models are wrong, some are useful" G.E.P. Box