topic Re: DoE with Blocking and Constraints in Discussions

DoE with Blocking and Constraints

peter_michel — Fri, 17 Nov 2017 13:55:34 GMT

I have to set up a DoE with blocking and Constraints. JMP pops up with "Constraint scripts are not supported for experiments with Blocking or mixture factors".

Does anyone of you have an Idea how I can solve my problem anyhow.

Just some background: I have Discrete Numeric and Categorial Factors in the experiment. Blocking is needed because I have 16 Teststations involved. I have a low end level combination of two Discrete Numeric 3Level Factors, that will not function- "Use disallowed combination Filter" works fine if I do it without blocking.

Many thanks for every help Peter

I prefer a interactive solution, but JSL would work also, but please with comments I'm just starter.

Re: DoE with Blocking and Constraints

Mark_Bailey — Fri, 17 Nov 2017 14:28:03 GMT

Perhaps someone else will have a better idea but I think that for this case you will have to cheat. I would make a larger design with the desired blocking but without constraints. Add enough extra runs to account for their elimination. For example, if you want 20 runs and find 3 runs that should be eliminated, ask for 23 runs first.

This approach might require some iteration...

Re: DoE with Blocking and Constraints

peter_michel — Mon, 27 Nov 2017 07:46:54 GMT

Hi Mark, thanks. I tried. I struggled. I thought. Blocking is nothing else than adding a Factor with just a main effect. I added a Factor 16 levels and added the main effect of this factor to the model. JMP works fine with a Categorial factor on 16 Levels and cope with all the constarints I had. Now i doubt myself, that I did it right. Why is Blocking as software setting not able to combine blocking with constraints and manually adding a Factor doing the same works fine to my imagination. Is there a trap I was falling in? Is blocking in JMP more than a Factor with just a main effect?

Thanks Peter

Re: DoE with Blocking and Constraints

Mark_Bailey — Mon, 27 Nov 2017 14:44:13 GMT

It is a common misconception that a blocking factor (blocks have fixed effects on the response) is just another categorical factor but they are much more than that. They represent a constraint on the experimental unit (part of the experimental run but generally not explicitly part of the design except for blocking). Let me look into it more because you are still struggling.

Re: DoE with Blocking and Constraints

Mark_Bailey — Mon, 27 Nov 2017 14:59:36 GMT

To illustrate my idea (not necessary validate it), I created a simple example with three continuous factors but you could have more factors and of different types. There is a linear constraint involving X1 and X2. The model is simple first order but could be any linear combination. I am going to perform eight runs at each of three test stations for a total of twenty-four runs. The specifications in the Custom Design look like this:

Then I click Make Design and finally Make Table. The resulting design is:

Hopefully you can adapt the approach to the specific details of your experiment.

Re: DoE with Blocking and Constraints

peter_michel — Mon, 27 Nov 2017 16:43:19 GMT

Hi Mark,

I followed your instruction and it worked. But with me, it does not work to run with "Use Disallowed Combination Filter" See underneath.

could you please check wether your system is doing it with "Use Disallowed Combination Filter" .

Do you use PC or Mac?

Thanks

Cheers Peter

Re: DoE with Blocking and Constraints

Mark_Bailey — Mon, 27 Nov 2017 18:06:38 GMT

First of all, Custom Design does not depend on the OS that you are using.

OK, then let's go back to my original approach: design the experiment with all of the necessary features except for the disallowed combinations. Make sure to add extra runs (that is, the number you request before clicking Make Design should be greater than the number that you actually intend to run). After you make the table, find the combinations you can't/won't run and delete these rows from the data table.

That way is the only approach that I can think of to get everything that you want.

Re: DoE with Blocking and Constraints

Mark_Bailey — Mon, 27 Nov 2017 18:46:02 GMT

Here is another example to illustrate the idea I suggested. I created a custom design for 4 categorical factors, each with 3 levels. I add a blocking factor with block size of 4 runs, the maximum number of runs per tester. I do not specify any constraints on the factor combinations. I am only using the first order model for this example. I want 20 runs in the final design but there are 4 combinations that I want to avoid, any or all of which might appear. I specify a total of 20 + 4 = 24 runs. Here is the specification so far:

I click Make Design and Make Table and then examine the results to identify the disallowed combinations. Perhaps X1, X2, and X4 cannot all be at level 3, so I select it before deleting it:

I can now evaluate the results by selecting DOE > Design Diagnostics > Evaluate Design. Assuming that this design is the initial study, the emphasis is on estimating and testing model parameters so the Power Analysis, Estimation Efficiency, and Correlations would be of informative:

Separately, would you please explain the disallowed combination that you defined above as [ (9.3333 <= Spring force <= 16) AND (5.5 <= Brush CS <= 6.5278) ]? Perhaps it could be specified as a linear constraint.

Re: DoE with Blocking and Constraints

Mark_Bailey — Mon, 27 Nov 2017 18:48:23 GMT

Towards my last question, what is the factor range for Spring force and Brush cross section?

Re: DoE with Blocking and Constraints

peter_michel — Tue, 28 Nov 2017 11:53:35 GMT

Hi Mark,

here in this screen dump, I have added senseless constraints, just to show: as soon as I add a constraint I will get the failure message shown.

Did you get a failure message as well after changing from "specify linear constraints" to "Use dissallowed combination Filter" and click make table?

Peter

Re: DoE with Blocking and Constraints

Jeff_Perkinson — Tue, 28 Nov 2017 14:18:34 GMT

@peter_michel, you mentioned a screen capture but there's no image in your post above. Can you edit your post and include it?

Re: DoE with Blocking and Constraints

Mark_Bailey — Tue, 28 Nov 2017 15:44:21 GMT

I don't see a picture of your screen.

I did not use the disallowed combinations feature to define a constraint because you proved that this feature won't work with blocking factors. I hoped that your constraint might be defined as a linear constraint instead, but you haven't answered my questions yet.

Re: DoE with Blocking and Constraints

peter_michel — Tue, 28 Nov 2017 15:52:41 GMT

Hi Jeff,

this post was refering to a post above. Sorry for creating confusion. Underneath it is now complete. Mark was asking for the meaning of the constraints.

Hi Mark,

here in this screen dump, I have added senseless constraints, just to show: as soon as I add a constraint I will get the failure message shown underneath.

Did you get a failure message as well after changing from "specify linear constraints" to "Use dissallowed combination Filter" and click make table?

Peter

I would like to correct the post as you pleased me, but there is no button to edit. May be you can guide me that we clean up the postings for other readers.

Peter

Re: DoE with Blocking and Constraints

peter_michel — Tue, 28 Nov 2017 16:31:20 GMT

Hi Mark,

sorry I did not understood your hint, to realize my constraints by "Specify linear constraints" tool/function. I will try.

To answer your question now.

It will not work:

Small cross section with agent 1 or 2

High spring force with Agent 1 or 2

I added also the factors in .JMP

Thanks Peter

Re: DoE with Blocking and Constraints

Mark_Bailey — Tue, 28 Nov 2017 16:49:24 GMT

You only have two levels for Agent (Agent1, Agent2). I understand the constraint now to be such that (1) you cannot use a low cross section with either Agent and (2) you cannot use a high spring force with either Agent. So why not limit the individual factor ranges instead of trying to introduce a constraint?

Also, why are you using the discrete numeric factor type for the first four factors instead of the continuos factor type?

Re: DoE with Blocking and Constraints

peter_michel — Wed, 29 Nov 2017 08:49:32 GMT

Sorry I did not explain that in detail.

There are two materials that can be used. And both of the materials can be improved by an Agent. This Agent is varied with 1% or 2 % (Certainly all these numbers are anonymized). Hence I have 0% --> no Agent, or 1% or 2 % Agent.

I use "Discrete Numeric" Factors because the processes of our supplier who is driving this DoE in his facilities, are not able to adjust the Factors to continuous values that are chosen by optimally criterias.

Sorry my impression is we deviating from my question. Let me raise my originally question again. What is the difference between.

Blocking Factor with 16 Blocks or Categorical Factor with 16 Levels. See underneath. Sure, making Amout of levels equally and adding just a main effect for the categorical value.

You answered above: "It is a common misconception that a blocking factor (blocks have fixed effects on the response) is just another categorical factor but they are much more than that. They represent a constraint on the experimental unit (part of the experimental run but generally not explicitly part of the design except for blocking)".

What does it mean? What causes the difference? What is the failure I have to expect if I choose an optimal design with a categorical Value instead of Blocking?

Many thanks

Peter

Re: DoE with Blocking and Constraints

Mark_Bailey — Wed, 29 Nov 2017 16:30:18 GMT

The purpose of a categorical factor is to model its fixed effects (estimate and test). The purpose of a blocking factor is to provide a homogeneous group of runs based on the inclusion of an external variable (e.g., tester) in the design. The homogeneity provides for estimation or comparison without the additional variation from changing the blocks. Full randomization of the runs across blocks would not necessarily provide the best set of treatments in each block. Full randomization occurs with a categorical factor.

You would not include this variable if you only needed one level, but you have 16 levels out of convenience or necessity. You are not interested in the effect of the blocks but you want the groups to be homogeneous and you want to be able to account for their variance.

Re: DoE with Blocking and Constraints

peter_michel — Thu, 30 Nov 2017 15:41:18 GMT

Hi Mark,

that makes it more clear. Thanks

Peter