help with disallowed combinations for Custom DOE

[#1]

Hi, I'm trying to design a split-plot experiment (4 hard-to-change (HTC) and 1 easy-to-change (ETC) variable) that involves existing hardware, so I've set the role of the four HTC variables as "discrete numeric" in the design.  Each of HTC variables (let's call them A, B, C, D) has three values representing the low, middle, and high levels (let's call the values 1, 2, 3 - but they are numeric, not categorical), so a full-factorial of the HTC factor combinations would yield 3^4 = 81 combinations.  However, the hardware supports just 39 of those 81 combinations, so I've entered 42 disallowed combinations in the form of, for example: (A == 1 & B == 1 & C == 2 & D == 3) | (A == 2 & B == 3 & C == 1 & D == 2) | etc

The ETC variable is continuous with no restrictions on what values can be set, and I chose three levels or values of the ETC variable.

I'm running Custom DOE with JMP 15.2 (on a Mac running OS 10.15.7), and I chose an I-Optimal design and set the number of random starts to 150.  The design generation takes approximately 25 minutes per design.  I chose 45 whole plots with three subplots each, for a total of 135 subplots.  However, the resulting design contains five of the disallowed combinations of HTC variables.  Why is it doing this?  I double checked my list of disallowed combinations, and the list is complete and appears to be entered correctly.  Any suggestions would be appreciated. 

[#2]

Not sure why you are getting that result with combinations that you think you have disallowed. Can you share an anonymised version of the problem?

I would say that creating a design with 45 whole plots might be quite computationally challenging and could explain the long design generation times.

Setting disallowed combinations can be difficult so it is easy to make mistakes.

One option when you have disallowed combinations is to create the table with all combinations (the 81-treatment full factorial) and then delete treatments that are not allowed. Then use this table as candidate runs. JMP 16 has improved features for this and it is possible in earlier versions. JMP Chief Data Scientist, Chris Gotwalt, gave a presentation at the most recent Discovery Summit Europe about this.

I hope this helps.

[#3]

Phil,

Thanks for the reply to my question.  Attached is a JMP file with a sanitized version of the design.  For this version, I did not set the number of random starts, so I just let JMP use its default value, and it ran much faster.  I didn't mind the long generation time with the 45-whole plot design and the number of random starts set to 150 - I certainly understand that it requires a lot of computing time.  Again, I'm getting factor combinations in the design that were set as disallowed combinations.

So that I did not have to laboriously enter the disallowed combinations script each time I ran a design, I typed the script into a text editor and just copied and pasted into the disallowed combinations script window.  I've also attached that text file.  Hopefully there are no "end of line" hidden characters from the text file that are causing issues with JMP.  The text file was created in BBEdit v13.5.6 on a Mac.

I did see on the JMP website that v16 does have improved features for setting candidate points for a design, and I did watch the presentation about it.  I've never tried that with JMP before, although I understand the concept.  But I don't know if that feature is available in v15.

Please take a look at my JMP file and let me know if see anything that I've perhaps done wrong that is causing JMP to ignore the disallowed combinations in the design process.  Thanks. 

[#4 · Accepted solution]

There are a couple of things that might be going on here - in the DOE dialog in your table, A had a level at 30 instead of .30, and D had 0 repeated twice.

I've also run into issues where the double representation on discrete numeric doesn't evaluate to equal, particularly with decimals. To get around this I use a "fuzzy equals" with absolute values.

I haven't checked all of your combinations, but hopefully the attached is working as expected. 

The covariates option was also available in JMP 15, but you might need to duplicate the table first (since some rows may be repeated),, but the I-optimality would be factored in on the controllable factor only (easy-to-change), but it could be another option.

[#5]

Ryan,

Thanks for the reply.  The error for factor "A" with a value of 30 instead of 0.30 was only for the "sanitized" version that I re-ran so that I could post something on the website.  The original design had the correct value of 0.30.  Sorry for my mistake there in what I sent.

I will try using the "fuzzy equals" logic that you included in the JSL file.  I'll let you know how it turns out.

[#6]

Ryan,

I utilized the "fuzzy equals" approach for the disallowable combinations script, and that worked fine.  There were no factor combinations in the resulting design that included a disallowed combination.

Thanks again for your help.