Hello everyone
This is a rather basic but interesting question.
Take JMP's Popcorn Experiment data as an example. As you know, the factor of this experiment is one categorical and two continuous.
The time is set to 3-5 and the power is set to 5-10.
Check the profiler that has been analyzed and optimized with the experimental results after designing as shown below.
Regardless of the experiment condition, if you expand the time axis, you can get the profiler as below.
And where Hotspot's Use Profile at Boundary to see the values below.
Now the question is.
In Profiler, you can check the level of factors that are beyond the experiment range. If so, are the estimates of the response at values above the experimental level predicted by JMP? For example, you can move a red line to bring Time to eight.
The second question is if the above is true, in the extreme case, if you drag the mouse to 20, you can trust the response and the change in the response and the other factors. Of course, we know that additional experimentation is required to estimate cases outside the experimental range. But the key to this question is how reliable you can be if you don't have that time and money right away.
Finally, is it normal for this case, but is it normal to make a combination that goes beyond the given experimental level when optimized? For example, in the above example, Time is 5 best. Is it possible to get 5.1 when performing optimization? For real customer cases, 0 to 150 is the experimental level, but when optimized, the optimal point for that factor is set to 154.61317. The best case out of the experimental range is not understood.
Thank you.
