Bayesian Optimization - What about Numeric Discrete factor?

Question

Hello,

I would like to know how to deal with discrete numerical factors in Bayesian optimization. 
How do I specify the steps between two discrete numerical values in JMP?

While it is possible to force a level (in column property) for categorical variables when it does not appear in existing trials, this is not possible for discrete values (and complicated to define in this way because if you only want integer values between 0 and 1000, you have many levels to define, and in any case, we would lose the numerical concept in this case).

Imagine that you only have two trials with a discrete factor that can take even values between 0 and 8. for the exercise, I created a DOE (just to be sure that the column properties are correctly defined)

and imagine you just have two runs.
In this case, it opens the candidate matrix and I have an error message (no bayesian optimization platform)

If I have additional runs, I can access to the platform, but no suggestion for missing levels

Has anyone here ever dealt with discrete digital factors?

Thank you for your feedbacks,

Florence

Phil_Kay · Answer

Yes, hopefully discrete numeric will be supported in Bayes opt in JMP Pro 20. Maybe there is potential for a Marketplace extension for generating candidates sets that could also help with factor constraint situations.

flo_kussener · Answer

Hi Phl,

Thank you for this workaround. It requests some preparation steps but it works.

Phil_Kay · Answer

Have you tried using a candidate set? You should be able to generate a large space-filling design including discrete numeric factors and then load this as the candidate set for Bayes opt.