Hi @gyoungwonpark,
The prediction profiler in JMP works, in the case of continuous factors a gradient descent algorithm is used 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 the settings.
The other option of course is to apply the Simulation tool, if you know the noise and expected deviation of your inputs you can predict your response distribution and expected 'optimum'.
Reference:
@Mark_Bailey gives a good answer in this post.
I would also recommend reading more here and here
Hope that helps!
Ben
“All models are wrong, but some are useful”