Bayesian Optimization with disallowed combinations?

Question

I've got a customer that is interested in using Bayesian Optimization for his process. He's got several factors and certain combinations of factors need to be avoided. Since there's no disallowed combinations filter in Bayesian Optimizations, is there another way to deal with this?

Could he add the runs that JMP suggests to the table, and then type in undesirable response values (far from the desired response values) for the disallowed combinations, so that subsequent runs are not recommended in those disallowed spaces?

Is there a way to make Bayesian Optimization work for a problem like this?

MarilynWheatley · Answer

Thanks for the suggestions, everyone. The linear constraints are a bit cumbersome to type out if there are several factors. One other suggestion that Chris made (and I hope he'll post here), is that using an existing data table we could use the option to augment the design where there is an option to add a disallowed filter combination. Chris said the Bayesian Optimization platform will honor the disallowed combinations in the profiler and in the run selection.

MikeD_Anderson · Answer

+1 for Mark's suggestion - I was talking with Chris in passing about this topic at some point and he mentioned that BayesOpt is looking at the profiler for a lot of what it does.

Phil_Kay · Answer

You can also create a space-filling design as a candidate set for Bayes opt. Filter out and delete the disallowed combinations. Then load in the remaining candidates using Load Candidate Set from Data Table. Bayes opt will then only be choosing from your allowed candidates.

Bayesian Optimization Batch Customizer Report Options