Prediction Profiler issue

Question

@malcolm_moore1

I have a question regarding an issue with the Prediction Profiler, and I’d appreciate your help in understanding it. I built a model using the Fit Model function, where three responses were fitted separately, each including multiple effects in Effect Screening mode. After building the model, I identified the most significant effects and checked the residuals for normality. Once the model diagnostics were complete, I used the Prediction Profiler to explore which factor values would yield a specific target response. Next, I applied the Desirability Function, setting the response to a target value. This resulted in a bell-shaped desirability curve. I then used Maximize Desirability, which adjusted the factor values so that the response matched the target. However, I noticed something unexpected: when I manually adjust a factor line in the Prediction Profiler and then click Maximize Desirability again, the resulting factor values sometimes differ significantly from the previous ones, and the desirability function below the factor also changes shape.

Why does this happen? And which set of factor values should I trust when trying to predict the most suitable settings for achieving the target response?

Victor_G · Answer

Hi @Rokusan91,

The situation you describe is quite common: the more responses you add in your modeling and Profiler, the more "trade-offs"/compromise you may need to have to get an "optimal" solution.In most of the cases, responses do not evolve with the same direction and pattern, so there might not be one single solution that may result in a good compromise.

What you can do to better visualize this situation is to use the Simulator to generate data points from your model, covering your factors ranges with enough points and random uniform distributions. The resulting datatable will contain the predicted responses, as well as  an "Obj" column, containing the overall desirability score.
You can then plot the data points from this table, and see where are the optima located. You'll probably see that there are several combinations that can help you reach an optimum (here an example with two responses for three factors), a Pareto front :

The other option consists in doing the "Maximize and Remember" option under "Optimization and Desirability" several times with different starting points. But this is hard work, and may be not completely exhaustive and prompt to errors.

Hope this Pareto front solution may help you,

Ben_BarrIngh · Answer

Hi Rokusan91,

The prediction profiler in JMP works, in the case of continuous factors, using a gradient descent algorithm to try and find the optimum settings of your input factors to meet the desirability criteria that has been set for each response. You may see different predicted values as there are more than one possible optimums for your system, the profiler is showing you one of those possibilities, when you adjust your inputs, the profiler shows a different but equally desirable optimum.

You will need to look into the maximisation options (Optimisation and Desirability > Maximisation Options) to alter this to be more provide a more constrained result - I found that increasing the number of trips is a good place to start, work iteratively to change.

I’ve answered similarly in this post that has some useful tips: https://community.jmp.com/t5/Discussions/Different-result-from-Prediction-profiler-everytime/td-p/746903

Thanks!
Ben