How to Optimize Multiple JMP Models Simultaneously

[#1]

I am trying to optimize the processing of a product which is experiencing about 10% scrap rate launching the product into production. I ran a process DOE w/4 factors at 3 levels and collected data on 9 trials. The defect rating response was set to an ordinal variable and fitted a model by logistic regression. I also collected data on the hardness of this product which is also affected by the process variables and have to meet a physical target. The hardness data is a continuous variable and it was fit to a standard least squares model. These two models exist in different JMP windows. I can optimize both models independently, but I don't know how to optimize the two models simultaneously. I figure this is easy to do in JMP13, but I'm kinda new to JMP modeling and unfortunately I'm not seeing an obvious way for simultaneous optimization. Please help.

[#2 · Accepted solution]

For each model, save the Prediction Formula (it is under the red triangle pop-up menu). Once saved, go to the Graph Menu and choose Profiler. Specify the prediction formulas as the dialog box indicates. Both models will now appear on the profiler for simultaneous optimization.

I should add that for a design with that many factors, that many levels, and so few trials, you may have difficulty truly optimizing this process.

[#3]

Hi Kelly,

Is that 9 trials total or per factor level combination?

As Dan stated, saving the prediction formulas to the data table allows you to combine models from various platforms into a prediction profiler from Graph > Profiler.  If you want the confidence bands in this profiler for your continuous response, save the StdErr Pred Formula as well.  In the dialog for Profiler, put both the prediction formula and the std. err.  prediction formula (probably named "PredSE Hardness") into "Y, Prediction Formula." 

Have you used desirability functions before?  If not, that is a great way to do multiple optimization and the Prediction Profiler has a really great implementation of that technique.

One other thing that may help you out is that when you save the prediction (probability) formula for the defect rating model, JMP will create separate columns for the linear component of the model ("Linear"), the cumulative log odds ratios for each level of your rating scale (e.g. "Cum[2]", "Cum[3]", etc.), and the predicted probability for each level of your rating scale (e.g. "Prob[2]", "Prob[3]", etc).  If you plot the probability columns in the profiler, it will show the cumulative log odds ratio columns as the inputs (someone correct me if they know a better way to do this).  They way I have gotten around this is to copy the formula in "Linear" and paste it into reference to the Linear column in the formulas for each Cum[#] column.  Then, I copy the formulas from each Cum[#] column and paste those columns into where ever they are referenced in the Prob[#] column formulas.  The end result is that the Prob[#] columns are only functions of your original inputs, so the Linear and Cum[#] columns are now superfluous.  If you plot all the Prob[#] columns now, the profiler will look as expected with just your experimental factors as inputs in the profiler.

[#4]

If you check the Expand Intermediate Formulas in the profiler dialogue, doesn't that avoid having to copy the formula?

[#5]

---

@stephen_pearson wrote:
If you check the Expand Intermediate Formulas in the profiler dialogue, doesn't that avoid having to copy the formula?

---

I have never noticed that checkbox down there!  Thanks!  I figured there was likely a better way to do that.

---

[#6]

Where is the profiler dialogue? I checked the red triangle next to Prediction Profiler, not there.

[#7]

I’m referring to the standalone Profiler platform in the Graph menu. Graph > Profiler.