Hi Peter,
to both wuestions this might be of help: http://northhavengroup.com/pdfs/MultipleOptimization.pdf
However, in slides 8-12 they do not specifically talk about the algorithm but the concept.
Martin
Peter,
I ran across this article that was shared by Mark Bailey.
https://www.linkedin.com/pulse/those-new-design-experiments-desirability-functions-bob-launsby?trk=hp-feed-article-title-like
susan.walsh1 shared the following in What is the specific method that JMP uses to maximize desirability in the Profiler:
To maximize the desirability for continuous factors, JMP uses a constrained Newton's method approach with step shortening for the desirability functionality. A reference for this is:
Numerical Methods for Unconstrained Optimization and Nonlinear Equations by Dennis and Schnabel.
For categorical factors, JMP uses a greedy, one variable at a time optimization algorithm. When there are both categorical and continuous factors, JMP goes back and forth between the two strategies. To reduce the odds of finding a local optima, JMP uses multiple random starts. The optimizer chooses and alternates between a variety of optimization steps depending on the problem, and how it is behaving.
And Michael Crotty has shared the documentation around Desirability from the upcoming JMP 13 docs.
